WO2024093249A1 - 一种气象预测方法、装置及相关设备 - Google Patents

一种气象预测方法、装置及相关设备 Download PDF

Info

Publication number
WO2024093249A1
WO2024093249A1 PCT/CN2023/100686 CN2023100686W WO2024093249A1 WO 2024093249 A1 WO2024093249 A1 WO 2024093249A1 CN 2023100686 W CN2023100686 W CN 2023100686W WO 2024093249 A1 WO2024093249 A1 WO 2024093249A1
Authority
WO
WIPO (PCT)
Prior art keywords
models
weather
reasoning
meteorological data
model
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2023/100686
Other languages
English (en)
French (fr)
Inventor
毕恺峰
谢凌曦
田奇
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Huawei Cloud Computing Technologies Co Ltd
Original Assignee
Huawei Cloud Computing Technologies Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Huawei Cloud Computing Technologies Co Ltd filed Critical Huawei Cloud Computing Technologies Co Ltd
Priority to EP23884188.6A priority Critical patent/EP4603879A4/en
Publication of WO2024093249A1 publication Critical patent/WO2024093249A1/zh
Priority to US19/192,997 priority patent/US20250258318A1/en
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01W—METEOROLOGY
    • G01W1/00—Meteorology
    • G01W1/10—Devices for predicting weather conditions
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/20—Ensemble learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G06N3/0455—Auto-encoder networks; Encoder-decoder networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/096—Transfer learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/01—Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/04—Inference or reasoning models
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation

Definitions

  • the present application relates to the field of artificial intelligence technology, and in particular to a weather forecasting method, device and related equipment.
  • Weather forecast refers to the technology of using modern scientific methods to predict the future weather conditions of the earth's atmosphere (or a certain location under the earth's atmosphere). It plays an important role in arranging work reasonably and warning of severe weather.
  • NWP numerical weather prediction
  • the embodiments of the present application provide a weather forecasting method to reduce the computing power required for weather forecasting and reduce the forecasting delay.
  • the present application also provides a corresponding apparatus, a computing device, a computing device cluster, a computer-readable storage medium, and a computer program product.
  • an embodiment of the present application provides a weather forecasting method, which can be executed by a corresponding weather forecasting device.
  • the weather forecasting device obtains weather data and a target time.
  • the obtained weather data can be, for example, weather data of a certain region at the current moment, or global weather data at the current moment, etc.
  • the target time can be a future moment or a future time period, etc.; then, the weather forecasting device determines multiple first AI models from a model library according to the target time, and different first AI models among the multiple first AI models are used to predict weather at different time intervals, wherein the multiple AI models in the model library can be pre-trained based on sample data, and different AI models in the model library are used to predict weather at different time intervals such as 1 hour, 2 hours, 3 hours, 1 day, and 3 days, respectively.
  • Different AI models for predicting weather at the same time interval have differences in the training algorithms, model structures, or iterative reasoning algorithms used, so that the weather forecasting device uses multiple first AI models for reasoning based on the acquired meteorological data to obtain a first weather forecast result for the target time, wherein each first AI model performs at least one iterative reasoning operation.
  • the weather forecasting device uses multiple AI models to perform weather forecasting without the need to perform complex equation solving processes. This can not only effectively reduce the computing power required for weather forecasting, but also the AI model can quickly output the weather forecast results for the target time, thereby significantly reducing the delay in weather forecasting.
  • the weather forecasting device does not use a single AI model to perform a large number of iterative reasoning, but instead uses multiple AI models that predict the weather at different time intervals to perform a small number of reasonings. This can not only effectively reduce the iterative error, but also reduce the resource consumption required for iterative reasoning.
  • the weather forecasting device when it determines multiple first AI models, it can specifically determine multiple first AI models and the number of iterative reasoning operations that each of the multiple first AI models needs to perform from the model library according to the target time, such as iterative reasoning for some first AI models for 3 times and iterative reasoning for other first AI models for 10 times.
  • the weather forecast device uses multiple first AI models for reasoning, it can specifically use the multiple first AI models to perform iterative reasoning operations in sequence according to the meteorological data and the number of iterative reasoning operations required to be performed by each first AI model. In this way, the weather forecast device uses multiple first AI models to perform at least one iterative reasoning operation respectively to obtain the weather forecast result for the target time, thereby reducing the iteration error and reducing the resource consumption required for iterative reasoning.
  • the meteorological data acquired by the meteorological forecasting device may include three-dimensional meteorological data.
  • the acquired meteorological data may also include two-dimensional meteorological data, that is, the meteorological forecasting device may perform meteorological forecasting based on three-dimensional meteorological data and two-dimensional meteorological data.
  • the accuracy of the meteorological forecasting results obtained by reasoning based on the complete three-dimensional meteorological data can reach a high level, such as exceeding the accuracy of the meteorological forecasting results obtained by traditional numerical weather forecasting technology.
  • the meteorological data obtained by the meteorological forecasting device may also be two-dimensional meteorological data.
  • the meteorological data acquired by the meteorological forecasting device are three-dimensional meteorological data and two-dimensional meteorological data.
  • the meteorological forecasting device can specifically couple the three-dimensional meteorological data and the two-dimensional meteorological data to obtain fused meteorological data, and then use multiple first AI models to perform reasoning based on the fused meteorological data to obtain the first meteorological forecast result for the target time.
  • it when coupling the two-dimensional meteorological data and the three-dimensional meteorological data, it can specifically project the two-dimensional meteorological data and the three-dimensional meteorological data into a high-dimensional vector space respectively, and perform vector splicing in the high-dimensional vector space to obtain the fused meteorological data. In this way, the accuracy of the meteorological forecast results obtained by reasoning based on the complete three-dimensional meteorological data can reach a high level.
  • a bias is added to the calculation results of the intermediate variables or attention mechanisms in the multiple first AI models, and the bias is determined according to the altitude information or latitude information in the weather data.
  • the intermediate calculation results of the AI model can be compensated by using the bias pair, which can overcome the influence of the uneven distribution of weather data on the accuracy of weather forecast results, thereby improving the accuracy of reasoning using the AI model.
  • the weather forecasting device can also determine multiple second AI models from the model library according to the target time, and different second AI models in the multiple second AI models are used to predict the weather at different time intervals; wherein the multiple second AI models and the multiple first AI models mentioned above can have different data inputs, model structures, and adopt different iterative reasoning algorithms, or the second AI model and the first AI model are used to predict the weather at different time intervals, etc.; then, the weather forecasting device uses multiple second AI models to perform reasoning based on the meteorological data to obtain the second weather forecast result of the target time, and each second AI model performs at least one iterative reasoning operation.
  • the weather forecasting device can use another set of AI models for reasoning to obtain another weather forecast result, so that the weather forecasting device can present a variety of possible weather forecasts by presenting the second weather forecast result to improve the user experience; or, the weather forecasting device can comprehensively determine the final output weather forecast result based on multiple weather forecast results such as the first weather forecast result and the second weather forecast result to further improve the accuracy of weather forecasting.
  • the weather forecast device can determine the target weather forecast result based on the first weather forecast result and the second weather forecast result, such as by voting or calculating the average value, and output the target weather forecast result. In this way, the weather forecast device can comprehensively determine and output a unique weather forecast result by integrating the weather inference results obtained by inference of multiple groups of AI models, thereby improving the accuracy of weather forecasts.
  • the weather forecasting device can also obtain location information, which is used to indicate a certain region, such as city A, or to indicate a global range, and the first weather forecast result inferred by using multiple first AI models is used to indicate the weather corresponding to the location information. Then, when determining multiple first AI models, the weather forecasting device can specifically determine multiple first AI models from the model library for inferring the weather of the region indicated by the location information based on the location information and time information. In this way, independent forecasting of the weather in a certain region can be achieved.
  • the weather forecast device when obtaining the target time, may first output an interactive interface, such as presenting the interactive interface through an external client, etc., so as to obtain the target time input by the user in response to the user's operation on the interactive interface.
  • the user can specify the future time or time period for which the weather forecast is to be predicted, so that the inferred weather forecast result can meet the user's needs and improve the user experience.
  • an embodiment of the present application also provides a weather forecasting device, including: a data acquisition module, used to acquire weather data and a target time; a model determination module, used to determine multiple first artificial intelligence AI models from a model library according to the target time, and different first AI models among the multiple first AI models are used to predict the weather at different time intervals; an inference module, used to perform inference using the multiple first AI models based on the weather data to obtain a first weather forecast result for the target time, and each first AI model performs at least one iterative inference operation.
  • the model determination module is used to determine, from a model library, multiple first artificial intelligence AI models and the number of iterative reasoning operations that each of the multiple first AI models needs to perform according to the target time; the reasoning module is used to perform iterative reasoning operations in sequence using the multiple first AI models according to the meteorological data and the number of iterative reasoning operations that each of the first AI models needs to perform.
  • the meteorological data includes three-dimensional meteorological data.
  • the meteorological data includes the three-dimensional meteorological data and the two-dimensional meteorological data; the reasoning module is used to: couple the three-dimensional meteorological data and the two-dimensional meteorological data to obtain fused meteorological data; and perform reasoning using the multiple first AI models based on the fused meteorological data.
  • the reasoning module is used to add a bias to the calculation results of the intermediate variables or attention mechanisms in the multiple first AI models during the process of reasoning the meteorological data using the multiple first AI models, and the bias is determined based on the altitude information or latitude information in the meteorological data.
  • the model determination module is further used to determine multiple second AI models from a model library according to the target time, and different second AI models among the multiple second AI models are used to predict the weather at different time intervals; the reasoning module is also used to perform reasoning using the multiple second AI models based on the meteorological data to obtain a second meteorological forecast result for the target time, and each second AI model performs at least one iterative reasoning operation.
  • the reasoning module is further used to: determine a target meteorological forecast result based on the first meteorological forecast result and the second meteorological forecast result; and output the target meteorological forecast result.
  • the data acquisition module is also used to acquire location information, and the first weather forecast result is used to indicate the weather corresponding to the location information; the model determination module is used to determine the multiple first artificial intelligence AI models from the model library based on the location information and the target time.
  • the data acquisition module is used to: output an interactive interface; and acquire the target time input by the user in response to an operation of the user on the interactive interface.
  • the weather forecasting device provided in the second aspect corresponds to the weather forecasting method provided in the first aspect. Therefore, the technical effects of the second aspect and any implementation method thereof can refer to the technical effects of the first aspect or the corresponding implementation methods of the first aspect.
  • the present application provides a computing device, the computing device comprising a processor and a memory; the memory is used to store instructions, and the processor executes the instructions stored in the memory so that the computing device performs the meteorological forecasting method in the above-mentioned first aspect or any possible implementation of the first aspect.
  • the memory can be integrated into the processor or can be independent of the processor.
  • the computing device may also include a bus.
  • the processor is connected to the memory via the bus.
  • the memory may include a readable memory and a random access memory.
  • the present application provides a computing device cluster, the computing device includes at least one computing device, the at least one computing device includes at least one processor and at least one memory; the at least one memory is used to store instructions, and the at least one processor executes the instructions stored in the at least one memory, so that the computing device cluster executes the meteorological forecasting method in the above-mentioned first aspect or any possible implementation of the first aspect.
  • the memory can be integrated into the processor or can be independent of the processor.
  • the at least one computing device may also include a bus.
  • the processor is connected to the memory via a bus.
  • the memory may include a readable memory and a random access memory.
  • the present application provides a computer-readable storage medium, wherein instructions are stored in the computer-readable storage medium, and when the computer-readable storage medium is executed on at least one computing device, the at least one computing device executes the method described in the first aspect or any one of the implementations of the first aspect.
  • the present application provides a computer program product comprising instructions, which, when executed on at least one computing device, enables the at least one computing device to execute the method described in the first aspect or any one of the implementations of the first aspect.
  • FIG1 is a schematic diagram of an exemplary application scenario provided by the present application.
  • FIG2 is a schematic diagram of another exemplary application scenario provided by the present application.
  • FIG3 is a schematic diagram of a flow chart of a weather forecasting method provided by the present application.
  • FIG4 is a schematic diagram of a serial execution reasoning process of multiple AI models provided by the present application.
  • FIG5 is a schematic diagram of weather forecasting based on two-dimensional weather data and three-dimensional weather data provided by the present application
  • FIG6 is a schematic diagram of the structure of a computing device provided by the present application.
  • FIG. 7 is a schematic diagram of the structure of a computing device cluster provided in the present application.
  • NWP technology When making weather forecasts for a certain region or the world, NWP technology is usually used to predict the weather in the region or the world at a certain point in the future based on data reflecting the actual atmospheric situation, mathematical and physical process modeling, and solving complex fluid mechanics and thermodynamics equations.
  • solving complex equations requires high computing power, and the prediction delay is usually high.
  • the present application provides a weather forecasting method, which can be executed by a corresponding weather forecasting device to reduce the computing power required for weather forecasting and shorten the prediction delay.
  • the weather forecasting device obtains weather data and target time, and determines multiple artificial intelligence (AI) models from a model library according to the target time.
  • AI artificial intelligence
  • different AI models are used to predict weather at different time intervals, such as determining AI model a, AI model b and AI model c, and AI model a is used to predict weather for 1 hour, AI model b is used to predict weather for 12 hours, and AI model c is used to predict weather for 3 days, etc.
  • the weather forecasting device uses multiple AI models for reasoning based on the weather data to obtain the weather forecast result for the target time, wherein each AI model performs at least one iterative reasoning operation.
  • the weather forecasting device uses multiple AI models to perform weather forecasts, there is no need to perform complex equation solving processes. This can not only effectively reduce the computing power required for weather forecasts, but the AI model can also quickly output weather forecast results for the target time, thereby significantly reducing the delay in weather forecasts.
  • the weather forecasting device does not use a single AI model to perform a large number of iterative reasoning, but instead uses multiple AI models that predict the weather at different time intervals to perform a smaller number of reasonings. This can not only effectively reduce the iterative error, but also reduce the resource consumption required for iterative reasoning.
  • the weather forecasting device needs to use the AI model a for iterative reasoning 86 times (i.e. 3*24+14) to infer the weather forecast result 3 days and 14 hours later.
  • AI model a, AI model b and AI model c are used for reasoning at the same time, since AI model c can predict the weather 3 days later, AI model b can predict the weather 12 hours later, and AI model a can predict the weather 1 hour later, the weather forecasting device can use AI model c for reasoning once, AI model b for reasoning once, and finally AI model c for reasoning twice to obtain the weather forecast result 3 days and 14 hours later (i.e. 3 days + 12 hours + 1 hour * 2).
  • the above-mentioned weather forecast device can be deployed in the cloud to provide users with cloud services for weather forecasts.
  • the weather forecast device 100 can be deployed in the cloud, for example, it can be implemented by a computing device or a computing device cluster in the cloud.
  • the weather forecast device 100 can provide a client 200 to the outside to interact with the user 300, such as receiving the above-mentioned target time input by the user 300, or feeding back the target weather forecast result to the user 300.
  • the client 200 can be, for example, an application running on a user-side device, or it can be a web browser provided by the weather forecast device 100 to the outside.
  • the weather forecast device 100 can include a data acquisition module 101, a model determination module 102, and an inference module 103.
  • the data acquisition module 101 is used to obtain the target time input by the user 300 through the client 200, provide the target time to the model determination module 102, and obtain meteorological data from the cloud, and provide the meteorological data to the reasoning module 103;
  • the model determination module 102 is used to determine multiple AI models from the model library according to the target time (the model library is deployed in the cloud with the meteorological forecasting device 100), and provide the multiple AI models to the reasoning module 103;
  • the reasoning module 103 is used to use multiple AI models to perform reasoning based on the meteorological data to obtain the meteorological forecast result of the target time.
  • the reasoning module 103 can also send the meteorological forecast result of the target time to the client 200, so that the client 200 presents the meteorological forecast result to the user 300.
  • the above-mentioned weather forecast device can be deployed locally, so as to provide local weather forecast services for users.
  • the above-mentioned weather forecast device can be specifically a local terminal 400, so that the user 300 can input the target time and weather data into the terminal 400, and the terminal 400 uses the target time to determine multiple AI models from the model library deployed locally, and uses the multiple AI models to perform reasoning based on the weather data to obtain the weather forecast result of the target time, and present the weather forecast result to the user 300.
  • the above-mentioned weather forecasting device can be implemented by software or hardware.
  • a weather forecasting device may include code running on a computing instance.
  • the computing instance may include at least one of a host, a virtual machine, and a container.
  • the computing instance may be one or more.
  • the weather forecasting device may include code running on multiple hosts/virtual machines/containers.
  • the multiple hosts/virtual machines/containers used to run the code may be distributed in the same region or in different regions.
  • the multiple hosts/virtual machines/containers used to run the code may be distributed in the same availability zone (AZ) or in different AZs, each AZ including one data center or multiple data centers with close geographical locations.
  • a region may include multiple AZs.
  • VPC virtual private cloud
  • multiple hosts/virtual machines/containers used to run the code can be distributed in the same virtual private cloud (VPC) or in multiple VPCs.
  • VPC virtual private cloud
  • a VPC is set up in a region.
  • a communication gateway needs to be set up in each VPC to achieve interconnection between VPCs through the communication gateway.
  • the weather forecast device may include at least one computing device, such as a server, etc.
  • the weather forecast device may also be a device implemented using an application-specific integrated circuit (ASIC) or a programmable logic device (PLD).
  • the PLD may be a complex programmable logical device (CPLD), a field-programmable gate array (FPGA), a generic array logic (GAL), a data processing unit (DPU) or any combination thereof.
  • the multiple computing devices included in the weather forecasting device can be distributed in the same region or in different regions.
  • the multiple computing devices included in the weather forecasting device can be distributed in the same AZ or in different AZs.
  • the multiple computing devices included in the weather forecasting device can be distributed in the same VPC or in multiple VPCs.
  • the multiple computing devices can be any combination of computing devices such as servers, ASICs, PLDs, CPLDs, FPGAs, and GALs.
  • FIG 3 is a flow chart of a weather forecasting method in an embodiment of the present application.
  • the method can be applied to the application scenario shown in Figure 1 or Figure 2 above, or it can also be applied to other applicable application scenarios.
  • the following is an example of application to the application scenario shown in Figure 1.
  • the functions of the data acquisition module 101, the model determination module 102, and the reasoning module 103 in the weather forecasting device 100 are specifically described in the following embodiments.
  • the weather forecasting method shown in FIG3 may specifically include:
  • the data acquisition module 101 acquires target time and meteorological data.
  • the target time when making weather forecasts, you can first determine the time corresponding to the weather to be predicted, such as predicting 24 The weather conditions after 12 hours are referred to as the target time.
  • the weather forecasting device 100 can provide a client 200 to the outside, and the client 200 can present an interactive interface to the user 300, so that the user 300 can input the target time to be predicted on the interactive interface. Accordingly, the weather forecasting device 100 obtains the target time input by the user in response to the operation of the user 300 on the interactive interface.
  • the target time input by the user 300 can be a moment, such as the current moment can be 8:00:00, then the target time can be 18:00:00 in the future 10 hours away from the current moment, so that the weather forecasting device 100 can subsequently feedback the weather forecast result corresponding to the moment 18:00:00 to the user 300.
  • the target time input by the user 300 may also be a time period, such as a time period between 8:00:00 and 18:00:00, so that the weather forecast device 100 may subsequently feed back to the user 300 weather forecast results corresponding to multiple moments in the time period, such as the weather forecast result corresponding to 9:00:00, the weather forecast result corresponding to 10:00:00, the weather forecast result corresponding to 13:00:00, and the weather forecast result corresponding to 18:00:00, etc.
  • the weather forecast device 100 may perform periodic weather forecasts by default, and the moment corresponding to the period may be the target time, etc.
  • the data acquisition module 101 also acquires meteorological data, which may be two-dimensional meteorological data, or three-dimensional meteorological data, or both two-dimensional meteorological data and three-dimensional meteorological data, etc.
  • the two-dimensional meteorological data may include, for example, information such as temperature, humidity, wind speed, total rainfall, and total light intensity;
  • the three-dimensional meteorological data may include, for example, information such as isobaric surface, altitude, temperature, and humidity.
  • the user 300 can import meteorological data to the meteorological forecasting device 100 through the client 200, so that the data acquisition module 101 can receive the meteorological data sent by the client 200; or, the meteorological forecasting device 100 can obtain meteorological data from the network, such as remotely accessing meteorological detection equipment to obtain current meteorological data in real time.
  • the data acquisition module 101 can also acquire the target time and meteorological data in other ways, and this embodiment does not limit this.
  • the data acquisition module 101 may provide the target time to the model determination module 102 and provide the meteorological data to the reasoning module 103 .
  • the model determination module 102 determines a plurality of AI models from a model library according to a target time, and different AI models among the plurality of AI models are used to predict weather at different time intervals.
  • a model library may be deployed in the weather forecasting device 100, and the model library includes multiple different AI models, and different AI models are used to predict the weather at different time intervals.
  • the model library may include 10 AI models, and the 10 AI models are used to predict the weather after 1 hour, 2 hours, 3 hours, 5 hours, 7 hours, 12 hours, 24 hours (i.e., 1 day), 72 hours (i.e., 3 days), 168 hours (i.e., 7 days) and 360 hours (i.e., 15 days).
  • the model library may also include multiple AI models for predicting the weather at the same time interval, and the multiple AI models differ in terms of model input, model structure, iterative reasoning algorithm used for model reasoning, model training algorithm, etc.
  • the model library may also include AI model m, AI model n and AI model w, and these three AI models are all used to predict the weather for 1 hour, wherein AI model m and AI model n are trained using different training algorithms, and AI model m and AI model w have different neural network architectures, etc.
  • the AI model in the model library can be constructed based on a neural network with a transformer structure, such as a vision self-attention network (vision transformer), or can be based on other types of neural networks.
  • a transformer structure such as a vision self-attention network (vision transformer)
  • the network is constructed, but this embodiment does not limit this.
  • the weather forecasting device 100 can select multiple AI models from the model library to predict the weather at the target time.
  • the model determination module 102 can first select multiple AI models for weather forecasting from the model library according to the target time, and the total duration of the weather time intervals predicted by the multiple AI models matches the target time.
  • the model determination module 102 can filter out from the model library, according to the target time, an AI model x for predicting the weather 3 hours later, an AI model y for predicting the weather 12 hours later, and an AI model z for predicting the weather 72 hours later.
  • the interval time that these three AI models can predict the weather is 87 hours (i.e. 3 hours + 12 hours + 72 hours), which is 3 days and 15 hours.
  • the model determination module 102 can provide the multiple AI models to the reasoning module 103.
  • the reasoning module 103 uses multiple AI models to perform reasoning based on the acquired meteorological data to obtain the meteorological forecast result for the target time, wherein each AI model performs at least one iterative reasoning operation.
  • the reasoning module 103 can input meteorological data into one of the AI models, and the AI model can perform reasoning based on the meteorological data.
  • the obtained reasoning result 1 can be used as the input of the second AI model, and the second AI model can perform reasoning based on the reasoning result 1, and the obtained reasoning result 2 can be input into the next AI model. That is, after the reasoning module 103 inputs the meteorological data at time T1 into the first AI model, the input of the remaining AI models is the reasoning result of the output of the previous AI model, as shown in Figure 4.
  • the reasoning module 103 uses the multiple AI models to perform reasoning operations in sequence, and can obtain the meteorological forecast result output by the last AI model, that is, the meteorological forecast result at time T2 shown in Figure 4.
  • the order in which multiple AI models are inferred in sequence can be determined by the inference module 103 according to preset rules.
  • the order in which multiple AI models perform inference operations in sequence can be the order of the time intervals of the weather that can be predicted by the multiple AI models from small to large.
  • the first AI model to perform the inference operation can be the AI model x used to predict the weather after 3 hours.
  • the inference result 1 obtained by the AI model x based on the meteorological data can be used as the input of the AI model y used to predict the weather after 12 hours, and the AI model y outputs the inference result 2 after performing the inference operation, and the inference result 2 is used as the input of the AI model z used to predict the weather after 72 hours, so that the AI model z outputs the final inference result after performing the inference operation, that is, the weather forecast result at the target time.
  • the order in which multiple AI models are inferred in sequence can be randomly determined by the inference module 103, such as the inference module 103 can determine the inference order of AI model x, AI model y and AI model z based on a random algorithm as AI model z—>AI model x—>AI model y, etc.
  • the order in which multiple AI models perform inference operations is not limited.
  • the reasoning module 103 can use each determined AI model to perform a reasoning process respectively, and obtain the weather forecast result for the target time. For example, assuming that the target time is 3 days and 15 hours, the reasoning module 103 can use AI model x, AI model y and AI model z to perform a reasoning process respectively, and obtain a weather forecast result for 3 days and 15 hours.
  • the model determination module 102 can determine multiple AI models and the number of operations required to execute each AI model from the model library according to the target time.
  • the number of iterative reasoning operations, and the determined multiple AI models and the number of iterative reasoning operations required to be performed by each AI model are provided to the reasoning module 103.
  • the reasoning module 103 can iteratively perform the reasoning process on some or all AI models multiple times according to the number of iterative reasoning operations required to be performed by each AI model, so as to obtain the weather forecast result of the target time.
  • the model determination module 102 can determine from the model library the AI model o for predicting the weather after 1 hour, the AI model p for predicting the weather after 12 hours, and the AI model q for predicting the weather after 72 hours according to the target time.
  • the number of iterative reasoning operations required to be performed by AI model o is 3 times
  • the number of iterative reasoning operations required to be performed by AI model p is 1 time
  • the number of iterative reasoning operations required to be performed by AI model q is 1 time; accordingly, the total duration of the time interval of the weather that can be predicted by multiple AI models is 87 hours (i.e., 1 hour * 3 + 12 hours * 1 + 72 hours * 1).
  • the reasoning module 103 can perform three reasoning operations using AI model o according to the number of iterations corresponding to each AI model, and provide the results obtained after three iterative reasoning operations to AI model p; then, the reasoning module 103 performs one reasoning operation using AI model p, and provides the result obtained after performing one reasoning operation to AI model q, so that after AI model q performs one reasoning operation, the weather forecast result for the target time can be output.
  • T refers to the target time
  • f1 to fn respectively refer to the n AI models determined by the model determination module 102
  • N1 to Nn respectively refer to the number of reasoning operations required to be performed by AI models f1 to fn , such as N1 refers to the number of reasoning operations required to be performed by AI model f1
  • Nn refers to the number of reasoning operations required to be performed by AI model fn .
  • multiple AI models in the model library can be trained in advance based on meteorological data in the past time period.
  • the meteorological data at 5:00:00 and the meteorological data at 6:00:00 in the past time period can be used to complete the training.
  • the meteorological data at 5:00:00 can be used as the input of the AI model, and the AI model is inferred based on the input to obtain the meteorological data at 6:00:00 predicted by the AI model.
  • the parameters in the AI model are gradient adjusted, so that the AI model is iteratively trained using multiple sets of meteorological data until the training termination condition of the AI model is met, such as the value of the loss function of the AI model is less than a preset value.
  • multiple AI models for predicting weather at different time intervals can be trained.
  • multiple AI models in the model library can also be trained based on other methods, and this embodiment does not limit this.
  • a model library can be created based on the multiple AI models, and the model library can be deployed in the weather forecasting device 100.
  • the multiple AI models in the model library can be trained by the weather forecasting device 100, or they can be sent to the weather forecasting device 100 by other devices after the AI models are trained, etc. This embodiment does not limit this.
  • the meteorological data input into one of the AI models by the inference module 103 may be two-dimensional meteorological data, or three-dimensional meteorological data, or may include both two-dimensional meteorological data and three-dimensional meteorological data.
  • the inference module 103 may couple the two-dimensional meteorological data and the three-dimensional meteorological data in advance before inputting the meteorological data into the AI model to obtain fused meteorological data, so that the inference module 103 performs inference using multiple AI models based on the fused meteorological data.
  • the inference module 103 can input the two-dimensional meteorological data into the neural network 1, so that the neural network 1 projects the two-dimensional meteorological data into a high-dimensional (greater than or equal to three-dimensional) space to obtain a high-dimensional feature vector.
  • the inference module 103 also inputs the three-dimensional meteorological data into the neural network 2, so that the neural network 2 The three-dimensional meteorological data is projected into a high-dimensional (greater than or equal to three-dimensional) space to obtain a high-dimensional feature vector, as shown in FIG5.
  • the reasoning module 103 can splice the high-dimensional feature vector output by the neural network 1 with the high-dimensional feature vector output by the neural network 2, wherein the high-dimensional feature vectors output by the two neural networks have differences in the feature vectors corresponding to the dimension of height in the three-dimensional meteorological data, and the feature vectors of the remaining dimensions match, that is, the values of the feature vectors of the remaining dimensions can be summed in the process of splicing the high-dimensional feature vectors, and the calculated new high-dimensional feature vector is the above-mentioned fused meteorological data.
  • the neural network 1 outputs a 5-dimensional feature vector
  • the neural network 2 outputs a 6-dimensional feature vector, wherein the feature vectors of the 5 dimensions in the 6-dimensional feature vector match the feature vectors of the 5 dimensions output by the neural network 1, and the feature vector of the 6th dimension in the 6-dimensional feature vector is the feature vector corresponding to the dimension of height in the three-dimensional meteorological data.
  • the value of the feature vector of the 6th dimension in the 6-dimensional feature vector remains unchanged, and the values of the feature vectors of the remaining 5 dimensions in the 6-dimensional feature vector can be summed corresponding to the values of the 5-dimensional feature vector.
  • modeling and weather forecasting based on complete three-dimensional meteorological data can not only integrate more dimensional information, but also retain the relationship characteristics between more dimensional data, compared with the method of weather forecasting based only on two-dimensional meteorological data, thereby effectively improving the accuracy of weather forecasting.
  • the neural network 1 and the neural network 2 used to couple the two-dimensional meteorological data can be, for example, any one or more of a fully connected neural network (FCNN), a convolutional neural network (CNN) with a convolution kernel of 1, or other applicable neural networks, which is not limited in this embodiment.
  • FCNN fully connected neural network
  • CNN convolutional neural network
  • CNN convolution kernel of 1, or other applicable neural networks, which is not limited in this embodiment.
  • the reasoning module 103 can also decouple the meteorological forecast results to obtain two-dimensional meteorological forecast results and three-dimensional meteorological forecast results, as shown in Figure 5.
  • the process of data decoupling can be the inverse operation of the above-mentioned data coupling process.
  • the meteorological forecast result output by the last AI model among the multiple AI models can be a high-dimensional feature vector, so that the reasoning module 103 can split the high-dimensional feature vector to obtain two different high-dimensional feature vectors, and the process of splitting the high-dimensional feature vector is opposite to the process of splicing the high-dimensional feature vector.
  • the reasoning module 103 can input the two high-dimensional feature vectors obtained by the split into the neural network 3 and the neural network 4 respectively, so that the two-dimensional meteorological forecast results and the three-dimensional meteorological forecast results are obtained by the output of the two neural networks, as shown in Figure 5.
  • the meteorological data includes both two-dimensional meteorological data and three-dimensional meteorological data
  • multiple AI models in the model library can be trained based on the two-dimensional meteorological data and three-dimensional meteorological data in the past time period.
  • the AI model can be trained by first coupling the two-dimensional meteorological data and the three-dimensional meteorological data to obtain the fused meteorological data.
  • the process of training the AI model using the fused meteorological data can be referred to the aforementioned description of the relevant parts of training the AI model, which will not be repeated here.
  • the distribution of meteorological data is usually uneven and irregular.
  • the distribution of meteorological data near the equator ie, lower latitudes
  • the distribution of meteorological data near the South Pole or the North Pole ie, higher latitudes
  • the heights of isobaric surfaces (a type of meteorological data) in different regions are usually not the same, such as the atmospheric pressure at an altitude of 1 km in area A is the same as the atmospheric pressure at an altitude of 2 km in area B. Therefore, in the process of using multiple AI models for meteorological forecasting, the reasoning module 103 can compensate for the intermediate calculation results of the AI model to overcome the impact of the uneven distribution of meteorological data on the accuracy of meteorological forecast results.
  • the reasoning module 103 can A bias (bais) is added to the calculation result of the intermediate variable of each calculation unit, and the bias is determined according to the altitude information or latitude information in the meteorological data. That is, for the calculation result of the intermediate variable calculated by the AI model according to the meteorological data at different locations, if the latitude or altitude is different, the size of the bias added to the calculation result of the intermediate variable will also be different. When the latitude (and altitude) are the same, the size of the bias added to the calculation result of the intermediate variable is the same.
  • each AI model uses an attention mechanism for reasoning. Then, in the process of reasoning of each AI model, the reasoning module 103 can add a bias to the calculation result of the attention mechanism in each AI model.
  • the bias is determined according to the altitude information or latitude information in the meteorological data, that is, the calculation result of the attention mechanism calculated by the AI model according to the meteorological data at different locations. If the latitude or altitude is different, the size of the bias added to the calculation result of the attention mechanism will also be different. When the latitude (and altitude) are the same, the size of the bias added to the calculation result of the attention mechanism is the same.
  • the bias added to the calculation result of the attention mechanism can also satisfy translational symmetry in longitude, that is, when calculating the calculation result of the attention mechanism according to meteorological data at different locations, because different geographical locations rotate around the earth's axis to satisfy symmetry, therefore, for the calculation result of the attention mechanism corresponding to two geographical locations with the same longitude difference, the same size of bias can be added (while satisfying the consistency of altitude and latitude information).
  • the same size of bias can be added (while satisfying the consistency of altitude and latitude information).
  • the weather forecasting device 100 can also predict the weather at the target time based on the weather data based on multiple groups of AI models, so as to provide multiple prediction possibilities of the weather at the target time.
  • the following is an example of determining two groups of AI models for weather forecasting.
  • the AI model used for weather forecasting is referred to as the first AI model (i.e., the first group of AI models), and the weather forecast result at the target time is referred to as the second forecast result.
  • the model determination module 102 can not only determine the above-mentioned multiple first AI models, but also determine multiple second AI models (i.e., the second group of AI models), and different second AI models in the multiple second AI models are used to predict the weather at different time intervals, and the total duration of the time intervals of the weather predicted by the multiple second AI models matches the target time, and each second AI model performs at least one iterative reasoning operation.
  • the determined second AI model may have differences in the combination of the predicted weather time intervals.
  • the multiple first AI models determined by the model determination module 102 are respectively AI model x for predicting the weather after 3 hours, AI model y for predicting the weather after 12 hours, and AI model z for predicting the weather after 72 hours, and the multiple second AI models determined are respectively AI model u for predicting the weather after 1 hour, AI model v for predicting the weather after 6 hours, and AI model w for predicting the weather after 24 hours, or other possible combinations.
  • the reasoning module 103 can use multiple second AI models to reason according to the meteorological data to obtain the second meteorological forecast result for the target time.
  • it can be iterative reasoning using AI model u 3 times, iterative reasoning using AI model v 2 times, and iterative reasoning using AI model w 3 times.
  • the total time intervals corresponding to these three AI models are 3 days and 15 hours (i.e., 1 hour * 3 + 6 hours * 2 + 24 hours * 3).
  • the specific implementation process of the reasoning module 103 using multiple second AI models to reason and obtain the second meteorological forecast result for the target time can be found in the description of the relevant parts of using multiple first AI models to reason and obtain the first meteorological forecast result for the target time in the aforementioned embodiment, which will not be repeated here.
  • the multiple second AI models determined by the reasoning module 103 differ from the first AI model in terms of model input, model structure, iterative reasoning algorithm used for model reasoning, model training algorithm, etc., and this embodiment does not further investigate this. Line limitation.
  • the reasoning module 103 After the reasoning module 103 obtains multiple different weather forecast results based on multiple groups of AI model reasoning, the multiple different weather forecast results can be presented to the user 300 through the client 200, so that the user 300 can evaluate the weather at the future target time based on the multiple weather forecast results.
  • the reasoning module 103 may also determine a unique meteorological forecast result based on a plurality of different meteorological forecast results. In specific implementation, the reasoning module 103 may compare the difference between the first meteorological forecast result and the second meteorological forecast result, and when the difference between the first meteorological forecast result and the second meteorological forecast result is within a preset error range, the reasoning module 103 may use the first meteorological forecast result or the second meteorological forecast result as the final target meteorological forecast result, and output the target meteorological forecast result.
  • the reasoning module 103 may re-predict the weather at the target time based on the meteorological data, or the reasoning module 103 may use a third group of AI models to determine the third meteorological forecast result based on the meteorological data, and according to the voting method of the minority obeys the majority, determine the meteorological forecast result with the largest number of identical or similar meteorological forecast results among the multiple meteorological forecast results as the final unique output target meteorological forecast result, or the reasoning module 103 may determine the unique target meteorological forecast result by weighted calculation, etc., and this embodiment does not limit this. In this way, the meteorological forecast device 100 can predict multiple meteorological results based on multiple groups of AI models, and thereby realize a large number of integrated meteorological forecasts.
  • the reasoning module 103 can also output the target weather forecast result to the user 300. For example, the reasoning module 103 can send the target weather forecast result to the client 200, and the client 200 can present the target weather forecast result to the user 300.
  • the weather forecast device 100 defaults to forecasting the weather in a certain area or the global weather by default. In other embodiments, the weather forecast device 100 can also forecast the weather for the region specified by the user.
  • the data acquisition module 101 can also obtain location information, which is used to indicate the region where weather forecasting is required, such as the data acquisition module 101 can obtain the location information input by the user through the output client 200, so that the model determination module 102 can determine from the model library multiple AI models (such as the first AI model or the second AI model) for predicting the weather in the region indicated by the location information according to the location information and the target time, so as to use the determined multiple AI models to predict the weather in the region indicated by the location information.
  • the user can specify the region for which the weather forecasting is to be performed, and realize independent forecasting for regional weather, thereby improving the user's selectivity in weather forecasting and improving the user experience.
  • the weather forecasting device 100 uses multiple AI models to perform weather forecasting without executing a complex equation solving process, which can not only effectively reduce the computing power required for weather forecasting, but also the AI model can quickly output the weather forecast results of the target time, thereby significantly reducing the delay of weather forecasting.
  • the weather forecasting device 100 does not use a single AI model to perform a large number of iterative reasoning, but uses multiple AI models that predict the weather at different time intervals to perform a small number of reasoning, which can not only effectively reduce errors, but also reduce the resource consumption required for iterative reasoning.
  • meteorological data includes three-dimensional meteorological data
  • modeling and meteorological forecasting based on the complete three-dimensional meteorological data can not only integrate more dimensional information, but also retain the relationship characteristics between more dimensional data, compared with the method of forecasting based on two-dimensional meteorological data only, thereby effectively improving the accuracy of meteorological forecasting.
  • the influence of the irregularity of meteorological data on the meteorological forecast results can be overcome, so as to further Improve the accuracy of weather forecast results.
  • the accuracy of weather forecasts based on the above method can exceed the accuracy of weather forecasts based on NWP technology.
  • the weather forecasting device 100 (including the above data acquisition module 101, model determination module 102 and reasoning module 103) involved in the weather forecasting process can be software configured on a computing device or a computing device cluster, and by running the software on the computing device or the computing device cluster, the computing device or the computing device cluster can realize the functions of the above weather forecasting device 100.
  • the weather forecasting device 100 involved in the weather forecasting process is introduced in detail.
  • FIG 6 shows a structural diagram of a computing device, on which the above-mentioned weather forecasting device 100 can be deployed.
  • the computing device can be a computing device in a cloud environment (such as a server), or a computing device in an edge environment, or a terminal device, etc., which can be specifically used to implement the functions of the data acquisition module 101, the model determination module 102 and the reasoning module 103 in the embodiment shown in Figure 3 above.
  • the computing device 600 includes a processor 610, a memory 620, a communication interface 630, and a bus 640.
  • the processor 610, the memory 620, and the communication interface 630 communicate with each other through the bus 640.
  • the bus 640 may be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus.
  • the bus may be divided into an address bus, a data bus, a control bus, and the like.
  • FIG6 is represented by only one thick line, but it does not mean that there is only one bus or one type of bus.
  • the communication interface 630 is used to communicate with the outside, such as obtaining weather data, target time and location information, and outputting weather forecast results.
  • the processor 610 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), a graphics processing unit (GPU) or one or more integrated circuits.
  • the processor 610 may also be an integrated circuit chip having signal processing capabilities.
  • the functions of each module in the weather forecasting device 100 may be completed by hardware integrated logic circuits or software instructions in the processor 610.
  • the processor 610 may also be a general processor, a digital signal processor (DSP), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, and may implement or execute the methods, steps and logic diagrams disclosed in the embodiments of the present application.
  • the general processor may be a microprocessor or the processor may also be any conventional processor, etc.
  • the method disclosed in the embodiments of the present application may be directly embodied as being executed by a hardware decoding processor, or executed by a combination of hardware and software modules in a decoding processor.
  • the software module may be located in a storage medium mature in the art, such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc.
  • the storage medium is located in the memory 620, and the processor 610 reads the information in the memory 620 and completes part or all of the functions in the weather forecasting device 100 in combination with its hardware.
  • the memory 620 may include a volatile memory, such as a random access memory (RAM).
  • the memory 620 may also include a non-volatile memory, such as a read-only memory (ROM), a flash memory, a HDD, or a SSD.
  • the memory 620 stores executable codes, and the processor 610 executes the executable codes to execute the method executed by the aforementioned weather forecasting device 100 .
  • the data acquisition module 101, the model determination module 102, and the reasoning module 103 described in the embodiment shown in FIG. 3 are implemented by software
  • the data acquisition module 101, the model determination module 102, and the reasoning module 103 in FIG. 3 are executed.
  • the software or program code required for the functions of module 101, model determination module 102 and reasoning module 103 are stored in memory 620.
  • the interaction between data acquisition module 101 and other devices is realized through communication interface 630.
  • the processor is used to execute instructions in memory 620 to implement the method executed by meteorological forecasting device 100.
  • FIG7 shows a schematic diagram of the structure of a computing device cluster.
  • the computing device cluster 70 shown in FIG7 includes multiple computing devices, and the above-mentioned weather forecasting device 100 can be distributedly deployed on multiple computing devices in the computing device cluster 70.
  • the computing device cluster 70 includes multiple computing devices 700, each computing device 700 includes a memory 720, a processor 710, a communication interface 730 and a bus 740, wherein the memory 720, the processor 710, and the communication interface 730 are connected to each other through the bus 740.
  • the processor 710 may be a CPU, a GPU, an ASIC, or one or more integrated circuits.
  • the processor 710 may also be an integrated circuit chip with signal processing capabilities. In the implementation process, some functions of the weather forecasting device 100 may be completed by the hardware integrated logic circuit or software instructions in the processor 710.
  • the processor 710 may also be a DSP, an FPGA, a general processor, other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, and may implement or execute some methods, steps, and logic block diagrams disclosed in the embodiments of the present application.
  • the general processor may be a microprocessor or the processor may also be any conventional processor, etc., and the steps of the method disclosed in the embodiments of the present application may be directly embodied as a hardware decoding processor to be executed, or a combination of hardware and software modules in the decoding processor to be executed.
  • the software module may be located in a mature storage medium in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc.
  • the storage medium is located in the memory 720.
  • the processor 710 reads the information in the memory 720, and in combination with its hardware, some functions of the weather forecasting device 100 may be completed.
  • the memory 720 may include ROM, RAM, static storage device, dynamic storage device, hard disk (such as SSD, HDD), etc.
  • the memory 720 may store program codes, for example, part or all of the program codes for implementing the data acquisition module 101, part or all of the program codes for implementing the model determination module 102, part or all of the program codes for implementing the reasoning module 103, etc.
  • the processor 710 executes part of the method executed by the weather forecasting apparatus 100 based on the communication interface 730, such as part of the computing device 700 may be used to execute the method executed by the above-mentioned data acquisition module 101, part of the computing device 700 may be used to execute the method executed by the above-mentioned model determination module 102, and part of the computing device 700 may be used to execute the method executed by the above-mentioned reasoning module 103.
  • the memory 720 may also store data, for example: intermediate data or result data generated by the processor 710 during the execution process, for example, the above-mentioned first weather forecast result, the second weather forecast result, the target weather forecast result, etc.
  • the communication interface 703 in each computing device 700 is used for external communication, such as interacting with other computing devices 700 .
  • the bus 740 may be a peripheral component interconnect standard bus or an extended industry standard architecture bus, etc.
  • the bus 740 in each computing device 700 in FIG. 7 is represented by only one thick line, but does not mean that there is only one bus or one type of bus.
  • the plurality of computing devices 700 establish communication paths through a communication network to implement the functions of the weather forecasting apparatus 100.
  • Any computing device may be a computing device in a cloud environment (eg, a server), or a computing device in an edge environment, or a terminal device.
  • an embodiment of the present application also provides a computer-readable storage medium, which stores instructions.
  • the computer-readable storage medium When the computer-readable storage medium is run on one or more computing devices, the one or more computing devices execute the methods executed by the various modules of the meteorological forecasting device 100 of the above embodiment.
  • the embodiment of the present application further provides a computer program product, and when the computer program product is executed by one or more computing devices, the one or more computing devices execute any of the aforementioned weather forecasting methods.
  • the computer program product may be a software installation package, and when any of the aforementioned weather forecasting methods is required, the computer program product may be downloaded and executed on a computer.
  • the device embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed over multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment.
  • the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines.
  • the technical solution of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a readable storage medium, such as a computer floppy disk, a U disk, a mobile hard disk, a ROM, a RAM, a disk or an optical disk, etc., including a number of instructions to enable a computer device (which can be a personal computer, a training device, or a network device, etc.) to execute the methods described in each embodiment of the present application.
  • a computer device which can be a personal computer, a training device, or a network device, etc.
  • all or part of the embodiments may be implemented by software, hardware, firmware or any combination thereof.
  • all or part of the embodiments may be implemented in the form of a computer program product.
  • the computer program product includes one or more computer instructions.
  • the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.
  • the computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium.
  • the computer instructions may be transmitted from a website site, a computer, a training device, or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, training device, or data center.
  • the computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device, a data center, etc. that includes one or more available media integrations.
  • the available medium may be a magnetic medium, (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)), etc.
  • a magnetic medium e.g., a floppy disk, a hard disk, a tape
  • an optical medium e.g., a DVD
  • a semiconductor medium e.g., a solid-state drive (SSD)

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Health & Medical Sciences (AREA)
  • Environmental & Geological Engineering (AREA)
  • Ecology (AREA)
  • Biodiversity & Conservation Biology (AREA)
  • Atmospheric Sciences (AREA)
  • Environmental Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本申请提供了一种气象预测方法，包括：获取气象数据以及目标时间，并根据该目标时间，从模型库中确定多个第一AI模型，该多个第一AI模型中不同第一AI模型用于预测不同时间间隔的气象，从而根据获取的气象数据，利用多个第一AI模型进行推理，得到目标时间的第一气象预测结果，其中，每个第一AI模型执行至少一次迭代推理操作。如此，利用多个AI模型进行气象预测，无需执行复杂的方程求解过程，这不仅可以有效减小气象预测所需消耗的算力，而且能够显著降低气象预测的时延。另外，利用多个预测不同时间间隔的气象的AI模型进行较少次数的推理，可以有效减小迭代误差、降低迭代推理所需的资源消耗。此外，本申请还提供了对应的装置及相关设备。

Description

一种气象预测方法、装置及相关设备
本申请要求于2022年10月31日提交中国国家知识产权局、申请号为202211351491.5、申请名称为“一种气象预测方法、装置及相关设备”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及人工智能技术领域，尤其涉及一种气象预测方法、装置及相关设备。
背景技术
气象预测(weather forecast，或者可以称为天气预报)，是指使用现代科学手段对未来地球大气层(或者地球大气层下的某一地点)的天气状态进行预测的技术，对于合理安排工作、预警恶劣天气等方面具有重要作用。
目前，主要是基于数值天气预报(numerical weather prediction,NWP)技术，根据大气实际情况，通过对天气状况进行数学物理过程建模，并通过求解复杂的流体力学和热力学的方程的方式，预测未来一定时段的大气运动状态和天气现象。但是，复杂的方程求解过程，不仅需要消耗较高的算力，而且预测时延也较高。
发明内容
有鉴于此，本申请实施例提供了一种气象预测方法，以减少气象预测所需的算力，降低预测时延。本申请还提供了对应的装置、计算设备、计算设备集群、计算机可读存储介质以及计算机程序产品。
第一方面，本申请实施例提供了一种气象预测方法，该方法可以由相应的气象预测装置执行，具体地，气象预测装置获取气象数据以及目标时间，所获取的气象数据例如可以是某个地域在当前时刻的气象数据，或者可以是当前时刻的全球气象数据等，该目标时间可以是未来的一个时刻或者未来的时间段等；然后，气象预测装置根据该目标时间，从模型库中确定多个第一AI模型，该多个第一AI模型中不同第一AI模型用于预测不同时间间隔的气象，其中，模型库中的多个AI模型可以预先基于样本数据完成训练，并且，模型库中的不同AI模型分别用于预测1小时、2小时、3小时、1天、3天等不同时间间隔的气象，对于预测同一时间间隔的气象的不同AI模型在采用的训练算法、模型结构或者迭代推理算法等方面存在差异，从而气象预测装置根据获取的气象数据，利用多个第一AI模型进行推理，得到目标时间的第一气象预测结果，其中，每个第一AI模型执行至少一次迭代推理操作。
如此，气象预测装置是利用多个AI模型进行气象预测，无需执行复杂的方程求解过程，这不仅可以有效减小气象预测所需消耗的算力，而且AI模型能够快速输出目标时间的气象预测结果，以此能够显著降低气象预测的时延。另外，在气象预测过程中，气象预测装置并非利用单个AI模型执行较多次数的迭代推理，而是利用多个预测不同时间间隔的气象的AI模型进行较少次数的推理，这不仅可以有效减小迭代误差，而且也能降低迭代推理所需的资源消耗。
在一种可能的实施方式中，气象预测装置在确定多个第一AI模型时，具体可以是根据目标时间，从模型库中确定多个第一AI模型以及该多个第一AI模型中各个第一AI模型所需执行的迭代推理操作的次数，如部分第一AI模型迭代推理3次，其它第一AI模型迭代推理10 次等，从而气象预测装置在利用多个第一AI模型进行推理时，具体可以是根据气象数据以及各个第一AI模型所需执行的迭代推理操作的次数，利用该多个第一AI模型依次执行迭代推理操作。如此，气象预测装置利用多个第一AI模型分别执行至少一次的迭代推理操作，即可得到目标时间的气象预测结果，以此实现减小迭代误差、降低迭代推理所需的资源消耗。
在一种可能的实施方式中，气象预测装置所获取的气象数据，可以包括三维气象数据。进一步地，所获取的气象数据还可以包括二维气象数据，即气象预测装置可以基于三维气象数据以及二维气象数据进行气象预测。如此，基于完整的三维气象数据进行推理所得到的气象预测结果的精度可以达到较高水平，如可以超过传统的基于数值天气预报技术所得到的气象预测结果的精度等。
实际应用时，气象预测装置所获取的气象数据，也可以是二维气象数据。
在一种可能的实施方式中，气象预测装置所获取的气象数据为三维气象数据以及二维气象数据，则气象预测装置在进行气象预测时，具体可以是将三维气象数据以及二维气象数据进行耦合，得到融合气象数据，从而根据该融合气象数据，利用多个第一AI模型进行推理，以此得到目标时间的第一气象预测结果。其中，在耦合二维气象数据以及三维气象数据时，具体可以是分别将二维气象数据以及三维气象数据投影到高维的向量空间，并在高维的向量空间进行向量拼接，以此得到融合气象数据。如此，基于完整的三维气象数据进行推理所得到的气象预测结果的精度可以达到较高水平。
在一种可能的实施方式中，气象预测装置在进行气象预测时，具体可以是在利用多个第一AI模型对气象数据进行推理的过程中，对多个第一AI模型中的中间变量或者注意力机制的计算结果增加偏置，该偏置根据气象数据中的高度信息或者纬度信息进行确定。如此，利用偏置对可以对AI模型的中间计算结果进行补偿，能够克服气象数据的不均匀分布对气象预测结果的精度的影响，以此可以提高利用AI模型进行推理的准确性。
在一种可能的实施方式中，气象预测装置还可以根据目标时间，从模型库中确定多个第二AI模型，该多个第二AI模型中不同第二AI模型用于预测不同时间间隔的气象；其中，该多个第二AI模型与上述多个第一AI模型之间可以具有不同的数据输入、模型结构、采用不同的迭代推理算法，或者，第二AI模型与第一AI模型用于预测不同时间间隔的气象等；然后，气象预测装置根据该气象数据，利用多个第二AI模型进行推理，得到目标时间的第二气象预测结果，每个第二AI模型执行至少一次迭代推理操作。如此，气象预测装置可以利用另一组AI模型进行推理，能够得到另一种气象预测结果，从而气象预测装置可以通过呈现该第二气象预测结果来呈现预测出的多种可能的气象，以提高用户体验；或者，气象预测装置可以根据第一气象预测结果、第二气象预测结果等多种气象预测结果来综合确定最终输出的气象预测结果，以进一步提高气象预测的准确性。
在一种可能的实施方式中，气象预测装置可以根据第一气象预测结果以及第二气象预测结果确定目标气象预测结果，如通过投票或者计算平均值等方式确定目标气象预测结果等，并输出目标气象预测结果。如此，气象预测装置可以通过综合多组AI模型分别推理得到的气象推理结果，综合确定出唯一的气象预测结果并进行输出，可以提高气象预测的准确性。
在一种可能的实施方式中，气象预测装置还可以获取位置信息，该位置信息用于指示某个地域，如A城市等，或者指示全球范围等，并且，上述利用多个第一AI模型所推理出的第一气象预测结果用于指示该位置信息对应的气象，则，气象预测装置在确定多个第一AI模型时，具体可以是根据位置信息以及时间信息，从模型库中确定用于对该位置信息所指示的地域的气象进行推理的多个第一AI模型。如此，可以实现针对某个地域的气象进行独立预测。
在一种可能的实施方式中，气象预测装置在获取目标时间时，具体可以是先输出交互界面，如通过对外呈现的客户端呈现该交互界面等，从而响应于用户在该交互界面的操作，获取用户输入的目标时间。如此，可以由用户指定对未来哪个时刻或者哪个时间段内的气象进行预测，以便推理出的气象预测结果能够满足用户需求，提高用户体验。
第二方面，本申请实施例还提供了一种气象预测装置，包括：数据获取模块，用于获取气象数据以及目标时间；模型确定模块，用于根据所述目标时间，从模型库中确定多个第一人工智能AI模型，所述多个第一AI模型中不同第一AI模型用于预测不同时间间隔的气象；推理模块，用于根据所述气象数据，利用所述多个第一AI模型进行推理，得到所述目标时间的第一气象预测结果，每个第一AI模型执行至少一次迭代推理操作。
在一种可能的实施方式中，所述模型确定模块，用于根据所述目标时间，从模型库中确定多个第一人工智能AI模型以及所述多个第一AI模型中各个第一AI模型所需执行的迭代推理操作的次数；所述推理模块，用于根据所述气象数据、所述各个第一AI模型所需执行的迭代推理操作的次数，利用所述多个第一AI模型依次执行迭代推理操作。
在一种可能的实施方式中，所述气象数据包括三维气象数据。
在一种可能的实施方式中，所述气象数据包括所述三维气象数据以及二维气象数据；所述推理模块，用于：将所述三维气象数据以及所述二维气象数据进行耦合，得到融合气象数据；根据所述融合气象数据，利用所述多个第一AI模型进行推理。
在一种可能的实施方式中，所述推理模块，用于在利用所述多个第一AI模型对所述气象数据进行推理的过程中，对所述多个第一AI模型中的中间变量或者注意力机制的计算结果增加偏置，所述偏置根据所述气象数据中的高度信息或者纬度信息进行确定。
在一种可能的实施方式中，所述模型确定模块，还用于根据所述目标时间，从模型库中确定多个第二AI模型，所述多个第二AI模型中不同第二AI模型用于预测不同时间间隔的气象；所述推理模块，还用于根据所述气象数据，利用所述多个第二AI模型进行推理，得到所述目标时间的第二气象预测结果，每个第二AI模型执行至少一次迭代推理操作。
在一种可能的实施方式中，所述推理模块，还用于：根据所述第一气象预测结果以及所述第二气象预测结果，确定目标气象预测结果；输出所述目标气象预测结果。
在一种可能的实施方式中，所述数据获取模块，还用于获取位置信息，所述第一气象预测结果用于指示所述位置信息对应的气象；所述模型确定模块，用于根据所述位置信息以及所述目标时间，从所述模型库中确定所述多个第一人工智能AI模型。
在一种可能的实施方式中，所述数据获取模块，用于：输出交互界面；响应于用户在所述交互界面的操作，获取所述用户输入的所述目标时间。
值得注意的是，第二方面提供的气象预测装置，对应于第一方面提供的气象预测方法， 故第二方面以及第二方面中任一实施方式所具有的技术效果，可参见第一方面或者第一方面的相应实施方式所具有的技术效果。
第三方面，本申请提供一种计算设备，所述计算设备包括处理器和存储器；所述存储器用于存储指令，所述处理器执行所述存储器存储的该指令，以使所述计算设备执行上述第一方面或第一方面任一种可能实现方式中的气象预测方法。需要说明的是，该存储器可以集成于处理器中，也可以是独立于处理器之外。所述计算设备还可以包括总线。其中，处理器通过总线连接存储器。其中，存储器可以包括可读存储器以及随机存取存储器。
第四方面，本申请提供一种计算设备集群，所述计算设备包括至少一个计算设备，所述至少一个计算设备包括至少一个处理器和至少一个存储器；所述至少一个存储器用于存储指令，所述至少一个处理器执行所述至少一个存储器存储的该指令，以使所述计算设备集群执行上述第一方面或第一方面任一种可能实现方式中的气象预测方法。需要说明的是，该存储器可以集成于处理器中，也可以是独立于处理器之外。所述至少一个计算设备还可以包括总线。其中，处理器通过总线连接存储器。其中，存储器可以包括可读存储器以及随机存取存储器。
第五方面，本申请提供一种计算机可读存储介质，所述计算机可读存储介质中存储有指令，当其在至少一个计算设备上运行时，使得所述至少一个计算设备执行上述第一方面或第一方面的任一种实现方式所述的方法。
第六方面，本申请提供了一种包含指令的计算机程序产品，当其在至少一个计算设备上运行时，使得所述至少一个计算设备执行上述第一方面或第一方面的任一种实现方式所述的方法。
本申请在上述各方面提供的实现方式的基础上，还可以进行进一步组合以提供更多实现方式。
附图说明
为了更清楚地说明本申请实施例中的技术方案，下面将对实施例描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请中记载的一些实施例，对于本领域普通技术人员来讲，还可以根据这些附图获得其他的附图。
图1为本申请提供的一示例性应用场景的示意图；
图2为本申请提供的另一示例性应用场景的示意图；
图3为本申请提供的一种气象预测方法的流程示意图；
图4为本申请提供的一种多个AI模型串行执行推理过程的示意图；
图5为本申请提供的基于二维气象数据以及三维气象数据进行气象预测的示意图；
图6为本申请提供的一种计算设备的结构示意图；
图7为本申请提供的一种计算设备集群的结构示意图。
具体实施方式
下面将结合本申请中的附图，对本申请提供的实施例中的方案进行描述。
本申请的说明书和权利要求书及上述附图中的术语“第一”、“第二”等是用于区别类似的对象，而不必用于描述特定的顺序或先后次序。应该理解这样使用的术语在适当情况下可以互换，这仅仅是描述本申请的实施例中对相同属性的对象在描述时所采用的区分方式。
在对某个地区或者全球进行气象预测时，通常会基于NWP技术，根据反映大气实际情况的数据，通过数学物理过程建模以及求解复杂的流体力学和热力学的方程的方式，预测出该地区或者全球在未来某个时刻的气象。但是，求解复杂的方程需要消耗较高的算力，而且预测时延通常也较高。
基于此，本申请提供了一种气象预测方法，该气象预测方法可以由相应的气象预测装置执行，用于降低气象预测所需的算力、缩短预测时延。具体实现时，气象预测装置获取气象数据以及目标时间，并根据该目标时间，从模型库中确定多个人工智能(artificial intelligence，AI)模型。其中，不同AI模型用于预测不同时间间隔的气象，如确定AI模型a、AI模型b以及AI模型c，并且，AI模型a用于预测1小时的气象，AI模型b用于预测12小时的气象，AI模型c用于预测3天的气象等。这样，气象预测装置根据该气象数据，利用多个AI模型进行推理，能够得到该目标时间的气象预测结果，其中，每个AI模型执行至少一次迭代推理操作。
由于气象预测装置是利用多个AI模型进行气象预测，无需执行复杂的方程求解过程，这不仅可以有效减小气象预测所需消耗的算力，而且AI模型能够快速输出目标时间的气象预测结果，以此能够显著降低气象预测的时延。
另外，在气象预测过程中，气象预测装置并非利用单个AI模型执行较多次数的迭代推理，而是利用多个预测不同时间间隔的气象的AI模型进行较少次数的推理，这不仅可以有效减小迭代误差，而且也能降低迭代推理所需的资源消耗。
举例来说，假设当前需要预测3天14小时后的全球天气，此时，如果采用上述单个AI模型a进行迭代推理，则气象预测装置需要利用该AI模型a迭代推理86次(即3*24+14)才能推理得到3天14小时后的气象预测结果。而当同时采用AI模型a、AI模型b以及AI模型c进行推理时，由于AI模型c可以预测3天后的气象、AI模型b可以预测12小时后的气象、AI模型a可以预测1小时后的气象，因此，气象预测装置可以利用AI模型c推理1次、再利用AI模型b推理1次、最后再利用AI模型c推理2次，即可得到3天14小时后的气象预测结果(即3天+12小时+1小时*2)。
作为一种示例，上述气象预测装置可以被部署于云端，用于为用户提供气象预测的云服务。例如，在图1所示的应用场景中，气象预测装置100可以部署于云端，例如可以是由云端的计算设备或者计算设备集群实现。并且，气象预测装置100可以对外提供客户端200，用于实现与用户300的交互，如接收用户300输入的上述目标时间，或者向用户300反馈目标的气象预测结果等。实际应用时，客户端200例如可以是运行在用户侧设备上的应用程序，或者可以是气象预测装置100对外提供的网络浏览器等。气象预测装置100可以包括数据获取模块101、模型确定模块102、推理模块103。其中，数据获取模块101，用于获取用户300通过客户端200输入的目标时间，将目标时间提供给模型确定模块102，并从云端获取气象数据，将气象数据提供给推理模块103；模型确定模块102，用于根据目标时间，从模型库中确定多个AI模型(该模型库随气象预测装置100被部署于云端)，并将该多个AI模型提供给推理模块103；推理模块103，用于根据气象数据，利用多个AI模型进行推理，得到目标时间的气象预测结果。进一步地，推理模块103还可以将目标时间的气象预测结果发送给客户端200，以便客户端200将气象预测结果呈现给用户300。
作为另一种示例，上述气象预测装置可以被部署于本地，从而可以为用户提供本地的气象预测服务。例如，在图2所示的应用场景中，上述气象预测装置具体可以是本地的终端400，从而用户300可以向终端400输入目标时间以及气象数据，终端400利用该目标时间，从部署于本地的模型库中确定多个AI模型，并根据该气象数据，利用多个AI模型进行推理，得到目标时间的气象预测结果，并向用户300呈现该气象预测结果。
实际应用时，上述气象预测装置可以通过软件实现，或者可以通过硬件实现。
气象预测装置作为软件功能单元的一种举例，可以包括运行在计算实例上的代码。其中，计算实例可以包括主机、虚拟机、容器中的至少一种。进一步地，上述计算实例可以是一台或者多台。例如，气象预测装置可以包括运行在多个主机/虚拟机/容器上的代码。需要说明的是，用于运行该代码的多个主机/虚拟机/容器可以分布在相同的区域(region)中，也可以分布在不同的region中。进一步地，用于运行该代码的多个主机/虚拟机/容器可以分布在相同的可用区(availability zone，AZ)中，也可以分布在不同的AZ中，每个AZ包括一个数据中心或多个地理位置相近的数据中心。其中，通常一个region可以包括多个AZ。
同样，用于运行该代码的多个主机/虚拟机/容器可以分布在同一个虚拟私有云(virtual private cloud，VPC)中，也可以分布在多个VPC中。其中，通常一个VPC设置在一个region内，同一region内两个VPC之间，以及不同region的VPC之间跨区通信需在每个VPC内设置通信网关，经通信网关实现VPC之间的互连。
气象预测装置作为硬件功能单元的一种举例，气象预测装置可以包括至少一个计算设备，如服务器等。或者，气象预测装置也可以是利用专用集成电路(application-specific integrated circuit，ASIC)实现、或可编程逻辑器件(programmable logic device，PLD)实现的设备等。其中，上述PLD可以是复杂程序逻辑器件(complex programmable logical device，CPLD)、现场可编程门阵列(field-programmable gate array，FPGA)、通用阵列逻辑(generic array logic，GAL)、数据处理单元(Data processing unit，DPU)或其任意组合实现。
气象预测装置包括的多个计算设备可以分布在相同的region中，也可以分布在不同的region中。气象预测装置包括的多个计算设备可以分布在相同的AZ中，也可以分布在不同的AZ中。同样，气象预测装置包括的多个计算设备可以分布在同一个VPC中，也可以分布在多个VPC中。其中，所述多个计算设备可以是服务器、ASIC、PLD、CPLD、FPGA和GAL等计算设备的任意组合。
接下来，对气象预测过程的各种非限定性的具体实施方式进行详细描述。
参阅图3，为本申请实施例中一种气象预测方法的流程示意图。该方法可以应用于上述图1或者图2所示的应用场景中，或者也可以是应用于其它可适用的应用场景中。下面以应用于图1所示的应用场景为例进行说明。在图1所示的应用场景中，气象预测装置100中的数据获取模块101、模型确定模块102以及推理模块103的功能，具体参见下述实施例的相关描述。
图3所示的气象预测方法具体可以包括：
S301：数据获取模块101获取目标时间以及气象数据。
通常情况下，在进行气象预测时，可以先确定所要预测的天气对应的时间，如预测24 小时后的气象等，以下称之为目标时间。
作为一种获取目标时间的示例，气象预测装置100可以对外提供客户端200，并且，该客户端200可以向用户300呈现交互界面，从而用户300可以在该交互界面上输入所要预测的目标时间。相应地，气象预测装置100响应于用户300在该交互界面上的操作，获取用户输入的目标时间。其中，用户300输入的目标时间，可以是一个时刻，如当前时刻可以为8:00:00，则目标时间可以是与当前时刻距离10小时的未来时刻18:00:00，从而气象预测装置100后续可以向用户300反馈18:00:00这一时刻所对应的气象预测结果。或者，用户300输入的目标时间，也可以是一个时间段，如8:00:00～18:00:00之间的时间段，从而气象预测装置100后续可以向用户300反馈该时间段内的多个时刻分别对应的气象预测结果，如反馈9:00:00对应的气象预测结果、10:00:00对应的气象预测结果、13:00:00对应的气象预测结果以及18:00:00对应的气象预测结果等。或者，气象预测装置100可以默认执行周期性的气象预测，该周期对应的时刻可以是目标时间等。
另外，数据获取模块101还会获取气象数据，该气象数据可以是二维气象数据，或者可以是三维气象数据，又或者同时包括二维气象数据以及三维气象数据等。其中，二维气象数据，例如可以是温度、湿度、风速、总降雨量、总光照量等信息；三维气象数据，例如可以是等压面、高度、温度、湿度等信息。
作为一些获取气象数据的示例，用户300可以通过客户端200向气象预测装置100导入气象数据，从而数据获取模块101可以接收客户端200发送的气象数据；或者，气象预测装置100可以从网络中获取气象数据，如可以通过远程访问气象探测设备的方式实时获取当前的气象数据等。
值得注意的是，上述数据获取模块101获取目标时间以及气象数据的实现方式，仅作为一些示例性说明，实际应用时，数据获取模块101也可以通过其它方式获取目标时间以及气象数据，本实施例对此并不进行限定。
数据获取模块101在获取到目标时间以及气象数据后，可以将目标时间提供给模型确定模块102，将气象数据提供给推理模块103。
S302：模型确定模块102根据目标时间，从模型库中确定多个AI模型，该多个AI模型中的不同AI模型用于预测不同时间间隔的气象。
其中，气象预测装置100中可以部署有模型库，该模型库中包括多个不同的AI模型，并且，不同AI模型用于预测不同时间间隔的气象。比如，模型库中可以包括10个AI模型，该10个AI模型分别用于预测1小时、2小时、3小时、5小时、7小时、12小时、24小时(即1天)、72小时(即3天)、168小时(即7天)以及360小时(即15天)后的气象。进一步的，模型库中还可以包括用于预测同一时间间隔的气象的多个AI模型，该多个AI模型在模型输入、模型结构、模型推理所采用的迭代推理算法、模型训练算法等方面存在差异。比如，模型库中还可以包括AI模型m、AI模型n以及AI模型w，并且，这三个AI模型均用于预测1小时的气象，其中，AI模型m与AI模型n采用不同训练算法完成训练，AI模型m与AI模型w具有不同的神经网络架构等。
示例性地，模型库中的AI模型，可以基于transformer结构的神经网络进行构建，如基于视觉自注意力网络(vision transformer)进行构建等，或者可以是基于其它类型的神经网 络进行构建，本实施例对此并不进行限定。
本实施例中，气象预测装置100可以从模型库中筛选出多个AI模型，实现对目标时间的气象进行预测。具体实现时，模型确定模块102可以先根据目标时间，从模型库中筛选出此次进行气象预测的多个AI模型，该多个AI模型所预测的气象的时间间隔的总时长与该目标时间相匹配。
比如，假设目标时间为3天15小时，如预测距离当前时刻3天15小时后的气象，此时，模型确定模块102可以根据该目标时间，从模型库中筛选出用于预测3小时后气象的AI模型x、用于预测12小时后气象的AI模型y以及用于预测72小时后气象的AI模型z，这三个AI模型所能预测的气象的间隔时长可以为87小时(即3小时+12小时+72小时)，也即3天15小时。
模型确定模块102在确定出多个AI模型后，可以将该多个AI模型提供给推理模块103。
S303：推理模块103根据获取的气象数据，利用多个AI模型进行推理，得到目标时间的气象预测结果，其中，每个AI模型执行至少一次迭代推理操作。
具体实现时，推理模块103可以将气象数据，输入至其中一个AI模型，并由该AI模型根据该气象数据进行推理，所得到的推理结果1可以作为第二个AI模型输入，并由第二个AI模型根据该推理结果1进行推理，并将得到的推理结果2继续输入至下一个AI模型中。即推理模块103在将T1时刻的气象数据输入至第一个AI模型后，其余AI模型的输入为上一个AI模型的输出的推理结果，如图4所示。如此，推理模块103利用该多个AI模型依次执行推理操作，可以得到最后一个AI模型所输出的气象预测结果，也即图4中所示的T2时刻的气象预测结果。
其中，多个AI模型依次推理的顺序，可以由推理模块103根据预设规则进行确定。比如，多个AI模型依次执行推理操作的顺序，可以为该多个AI模型所能预测的气象的时间间隔从小到大的顺序。仍以目标时间为3天15小时为例，第一个执行推理操作的AI模型可以为用于预测3小时后气象的AI模型x，该AI模型x根据气象数据进行推理所得到的推理结果1，可以作为用于预测12小时后气象的AI模型y的输入，并由AI模型y执行推理操作后输出推理结果2，并将该推理结果2作为用于预测72小时后气象的AI模型z的输入，从而由该AI模型z执行推理操作后输出最终的推理结果，也即目标时间的气象预测结果。或者，多个AI模型依次推理的顺序，可以由推理模块103随机确定，如推理模块103可以基于随机算法确定AI模型x、AI模型y以及AI模型z的推理顺序为AI模型z—>AI模型x—>AI模型y等。本实施例中，对于多个AI模型执行推理操作的顺序并不进行限定。
实际应用时，当模型库中用于预测不同时间间隔的气象的AI模型较多时，如模型库中可以包括分别用于预测1小时、2小时、3小时、5小时、7小时、12小时、24小时、72小时、168小时以及360小时的气象的10个AI模型等，对于任意的目标时间，推理模块103可以利用所确定出的各个AI模型分别执行一次推理过程，即可得到目标时间的气象预测结果。举例来说，假设目标时间为3天15小时，推理模块103可以通过依次利用AI模型x、AI模型y以及AI模型z分别执行一次推理过程，即可得到为3天15小时的气象预测结果。
而当模型库中用于预测不同时间间隔的气象的AI模型较少时，如模型库中包括分别用于预测1小时、12小时、72小时以及168小时后的气象的4个AI模型等，对于任意的目标时间，模型确定模块102可以根据目标时间，从模型库中确定多个AI模型以及每个AI模型所需执行 的迭代推理操作的次数，并将确定出的多个AI模型以及每个AI模型所需执行的迭代推理操作的次数提供给推理模块103。推理模块103可以根据各个AI模型所需执行的迭代推理操作的次数，对部分或者全部AI模型迭代多次执行推理过程，以此得到目标时间的气象预测结果。仍以目标时间为3天15小时为例，模型确定模块102可以根据该目标时间可以从模型库中确定用于预测1小时后气象的AI模型o、用于预测12小时后气象的AI模型p、用于预测72小时后气象的AI模型q。其中，AI模型o所需执行的迭代推理操作的次数为3次，AI模型p所需执行的迭代推理操作的次数为1次，AI模型q所需执行的迭代推理操作的次数为1次；相应地，多个AI模型所能预测的气象的时间间隔的总时长为87小时(即1小时*3+12小时*1+72小时*1)。如此，推理模块103可以根据各个AI模型对应的迭代次数，利用AI模型o迭代执行3次推理操作，并将3次迭代推理后所得到的结果提供给AI模型p；然后，推理模块103再利用AI模型p执行1次推理操作，并将执行1次推理操作后所得到的结果提供给AI模型q，从而由AI模型q执行1次推理操作后，可以输出得到目标时间的气象预测结果。示例性地，模型确定模块102所确定出的多个AI模型以及各个AI模型所需执行的推理操作的次数，可以满足下述公式(1)。
T＝f1*N1+f2*N2+…+fn*Nn     (1)
其中，T是指目标时间；f1至fn分别是指模型确定模块102所确定出的n个AI模型；N1至Nn分别是指AI模型f1至fn分别所需执行的推理操作的次数，如N1是指AI模型f1所需执行的推理操作的次数、Nn是指AI模型fn所需执行的推理操作的次数。
相应地，模型库中的多个AI模型，可以预先根据过去时间段内的气象数据完成训练。比如，对于模型库中的用于预测1小时时间间隔的AI模型，可以利用过去时间段内的5:00:00的气象数据以及6:00:00的气象数据完成训练。具体地，5:00:00的气象数据可以作为该AI模型的输入，并由AI模型根据该输入进行推理，得到AI模型预测的6:00:00的气象数据。然后，通过比较该AI模型预测的6:00:00的气象数据与实际的6:00:00的气象数据之间的差异，对AI模型中的参数进行梯度调整，从而利用多组气象数据对该AI模型进行迭代训练，直至满足AI模型的训练终止条件，如该AI模型的损失函数的值小于预设值等。参照类似方式，可以训练得到多个用于预测不同时间间隔的气象的AI模型。实际应用时，模型库中的多个AI模型也可以基于其它方式训练得到，本实施例对此并不进行限定。在训练得到多个AI模型后，可以基于该多个AI模型创建模型库，并将该模型库部署于气象预测装置100中。其中，模型库中的多个AI模型，可以由气象预测装置100完成训练，也可以是由其它设备在完成AI模型的训练后，由其它设备发送给气象预测装置100等，本实施例对此并不进行限定。
推理模块103在预测目标时间的气象的过程中，输入至其中一个AI模型的气象数据，可以是二维气象数据，或者可以是三维气象数据，或者同时包括二维气象数据以及三维气象数据。其中，当气象数据同时包括二维气象数据以及三维气象数据时，推理模块103在将该气象数据输入至AI模型之前，可以预先将二维气象数据以及三维气象数据进行耦合，得到融合气象数据，从而推理模块103根据该融合气象数据，利用多个AI模型进行推理。
在一种可能的实施方式中，推理模块103可以将二维气象数据输入至神经网络1中，从而由该神经网络1将该二维气象数据投影到高维(大于或者等于三维)空间，得到高维的特征向量。并且，推理模块103还将三维气象数据输入至神经网络2中，从而由该神经网络2 将该三维气象数据投影到高维(大于或者等于三维)空间，得到高维的特征向量，如图5所示。然后，推理模块103可以将神经网络1输出的高维特征向量与神经网络2输出的高维特征向量进行拼接，其中，这两个神经网络所输出的高维特征向量之间，在三维气象数据中的高度这一维度对应的特征向量存在差异之外，其余维度的特征向量相匹配，即其余维度的特征向量的值可以在拼接高维特征向量的过程中进行求和运算，所计算出的新的高维特征向量即为上述融合气象数据。例如，神经网络1输出5维的特征向量，神经网络2输出6维的特征向量，其中，6维特征向量中的5个维度的特征向量与神经网络1输出的5个维度的特征向量相匹配，6维特征向量中的第6个维度的特征向量为三维气象数据中的高度这一维度对应的特征向量。则，在拼接神经网络1输出的5维特征向量与神经网络2输出的6维特征向量时，6维特征向量中的第6维度的特征向量的值保持不变，而6维特征向量中的其余5个维度的特征向量的值可以与5维特征向量的值对应求和。如此，基于完整的三维气象数据进行建模以及气象预测，相比于仅基于二维气象数据进行气象预测的方式，不仅可以集成更多维度的信息，而且也保留了更多维度的数据之间的关系特征，从而可以有效提高气象预测的精度。
示例性地，上述用于耦合二维气象数据以及三维气象数据的神经网络1以及神经网络2，例如可以是全连接神经网络(full connect neural network，FCNN)、卷积核为1的卷积神经网络(convolutional neural networks,CNN)中的任意一种或者多种，也可以是其它可适用的神经网络，本实施例对此并不进行限定。
在利用多个AI模型根据融合气象数据进行推理所得到气象预测结果后，推理模块103还可以对该气象预测结果进行解耦，得到二维的气象预测结果与三维的气象预测结果，如图5所示。其中，数据解耦的过程可以为上述数据耦合的过程的逆运算。具体地，多个AI模型中的最后一个AI模型所输出的气象预测结果可以是高维特征向量，从而推理模块103可以将该高维特征向量进行拆分，得到两个不同的高维特征向量，拆分高维特征向量的过程与前述拼接高维特征向量的过程相反。然后，推理模块103可以将拆分得到的两个高维特征向量分别输入至神经网络3以及神经网络4中，从而由这两个神经网络输出得到二维的气象预测结果与三维的气象预测结果，如图5所示。
相应地，当气象数据同时包括二维气象数据以及三维气象数据时，模型库中的多个AI模型可以基于过去时间段内的二维气象数据以及三维气象数据进行训练，具体可以是先对该二维气象数据以及三维气象数据进行耦合所得到的融合气象数据训练AI模型，其利用融合气象数据训练AI模型的过程可以参见前述关于训练AI模型的相关之处描述，在此不做赘述。
实际应用场景中，气象数据的分布通常并不均匀以及具有不规则性。比如，靠近赤道(即纬度较低)的气象数据的分布通常较为稀疏，而靠近南极或者北极(即纬度较高)的气象数据的通常分布较为密集。又比如，不同地区的等压面的高度(气象数据中的一种)通常并不相同，如A地区的1km高度处的大气压与B地区的2km高度处的大气压相同。因此，推理模块103在利用多个AI模型进行气象预测的过程中，可以对AI模型的中间计算结果进行补偿，以克服气象数据的不均匀分布对气象预测结果的精度的影响。
作为一种实现示例，在各个AI模型进行推理的过程中，推理模块103可以对各个AI模型 中各个计算单元的中间变量的计算结果增加偏置(bais)，该偏置根据气象数据中的高度信息或者纬度信息进行确定，即对于AI模型根据不同位置的气象数据所计算出的中间变量的计算结果，如果纬度或者高度不同，则为该中间变量的计算结果增加的偏置的大小也会存在差异，而当纬度(以及高度)相同时，为该中间变量的计算结果增加的偏置的大小相同。
作为另一种实现示例，各个AI模型采用注意力机制(attention)进行推理，则，在各个AI模型进行推理的过程中，推理模块103可以对各个AI模型中的注意力机制的计算结果增加偏置，该偏置根据气象数据中的高度信息或者纬度信息进行确定，即AI模型根据不同位置的气象数据所计算出的注意力机制的计算结果，如果纬度或者高度不同，则对于该注意力机制的计算结果所增加的偏置的大小也会存在差异，而当纬度(以及高度)相同时，对于该注意力机制的计算结果所增加的偏置的大小相同。进一步地，对于注意力机制的计算结果所增加的偏置，还可以在经度上满足平移对称性，即在根据不同位置的气象数据计算注意力机制的计算结果时，由于不同的地理位置绕地轴旋转满足对称性，因此，对于经度差相同的两个地理位置对应的注意力机制的计算结果，可以增加相同大小的偏置(同时满足高度以及纬度信息一致)。如此，通过对AI模型中的中间变量或者注意力机制的计算结果增加偏置，可以克服气象数据的不规则性对气象预测结果所产生的影响，以此可以进一步提高气象预测结果的精度。
在进一步可能的实施方式中，气象预测装置100还可以基于多组AI模型分别根据气象数据预测目标时间的气象，以此提供目标时间的气象的多种预测可能，下面以确定出2组AI模型进行气象预测为例进行示例性说明。为便于区分以及描述，以下将上述进行气象预测所使用的AI模型称之为第一AI模型(即第一组AI模型)，所得到的目标时间的气象预测结果称之为第二预测结果。
模型确定模块102不仅可以确定出上述多个第一AI模型，还可以确定多个第二AI模型(即第二组AI模型)，该多个第二AI模型中的不同第二AI模型用于预测不同时间间隔的气象，并且，该多个第二AI模型所预测的气象的时间间隔的总时长与该目标时间相匹配，每个第二AI模型执行至少一次迭代推理操作。其中，相对于第一AI模型，所确定出的第二AI模型可以在所预测的气象的时间间隔的组合存在差异。比如，假设目标时间为3天15小时，则模型确定模块102所确定出的多个第一AI模型分别为用于预测3小时后气象的AI模型x、用于预测12小时后气象的AI模型y、用于预测72小时后气象的AI模型z，而确定出的多个第二AI模型分别为用于预测1小时后气象的AI模型u、用于预测6小时后气象的AI模型v、用于预测24小时后气象的AI模型w，又或者可以是其它可能的组合。然后，推理模块103可以利用多个第二AI模型根据气象数据进行推理，得到目标时间的第二气象预测结果，具体可以是利用AI模型u迭代推理3次、AI模型v迭代推理2次、利用AI模型w迭代推理3次，如此，这三个AI模型对应的时间间隔的总时长为3天15小时(即1小时*3+6小时*2+24小时*3)。其中，推理模块103利用多个第二AI模型推理得到目标时间的第二气象预测结果的具体实现过程，可参见前述实施例中利用多个第一AI模型推理得到目标时间的第一气象预测结果的相关之处描述，在此不做赘述。
或者，推理模块103所确定出的多个第二AI模型，与第一AI模型在模型输入、模型结构、模型推理所采用的迭代推理算法、模型训练算法等方面存在差异等，本实施例对此并不进 行限定。
推理模块103在基于多组AI模型推理得到多个不同的气象预测结果后，可以将该多个不同的气象预测结果，通过客户端200呈现给用户300，以便由用户300根据多个气象预测结果评估未来目标时间的气象。
或者，推理模块103还可以根据多个不同的气象预测结果，确定唯一的气象预测结果。具体实现时，推理模块103可以比较第一气象预测结果与第二气象预测结果之间差异，并且，当第一气象预测结果与第二气象预测结果之间的差异处于预设误差范围之内，则推理模块103可以将该第一气象预测结果或者第二气象预测结果作为最终确定的目标气象预测结果，并输出该目标气象预测结果。而当第一气象预测结果与第二气象预测结果之间的差异超出该预设误差范围，则推理模块103可以重新根据气象数据预测目标时间的气象，或者，推理模块103可以利用第三组AI模型根据气象数据确定第三气象预测结果，并按照少数服从多数的投票方式，将多个气象预测结果中相同或者相似数量最多的气象预测结果，确定为最终唯一输出的目标气象预测结果，或者推理模块103可以通过加权计算等方式确定唯一的目标气象预测结果等，本实施例对此并不进行限定。如此，气象预测装置100可以基于多组AI模型，预测多种气象结果，并以此实现较大数目的集成气象预测。
推理模块103在确定出目标气象预测结果后，还可以将该目标气象预测结果输出给用户300，例如推理模块103可以将该目标气象预测结果发送给客户端200，并由客户端200将该目标气象预测结果呈现给用户300等。
值得注意的是，本实施例中是以气象预测装置100默认对某个地区的气象或者默认对全球的气象的进行预测为例进行说明。在其它实施例中，气象预测装置100还可以对用户指定的地域进行气象预测。具体实现时，数据获取模块101还可以获取位置信息，该位置信息用于指示需要进行气象预测的地域，如数据获取模块101可以通过输出的客户端200获取用户输入的位置信息等，从而模型确定模块102可以根据该位置信息以及目标时间，从模型库中确定用于预测该位置信息所指示的地域的气象的多个AI模型(如第一AI模型或第二AI模型)，以便利用所确定的多个AI模型预测该位置信息所指示的地域的气象。如此，可以由用户指定所要预测气象的地域，实现针对地域气象进行独立预测，以此提高用户进行气象预测的可选择性，提高用户体验。
本实施例中，气象预测装置100利用多个AI模型进行气象预测，无需执行复杂的方程求解过程，这不仅可以有效减小气象预测所需消耗的算力，而且AI模型能够快速输出目标时间的气象预测结果，以此能够显著降低气象预测的时延。在气象预测过程中，气象预测装置100并非利用单个AI模型执行较多次数的迭代推理，而是利用多个预测不同时间间隔的气象的AI模型进行较少次数的推理，这不仅可以有效降低减小误差，而且也能降低迭代推理所需的资源消耗。
另外，当气象数据包括三维气象数据时，基于完整的三维气象数据进行建模并进行气象预测，相比于仅基于二维气象数据进行气象预测的方式，不仅可以集成更多维度的信息，而且也保留了更多维度的数据之间的关系特征，从而可以有效提高气象预测的精度。并且，在利用AI模型进行推理的过程中，通过对中间变量的计算结果或者注意力机制的计算结果增加偏置，可以克服气象数据的不规则性对气象预测结果所产生的影响，以此可以进一步 提高气象预测结果的精度。实际应用场景中，基于上述方式进行气象预测的精度，能够超过基于NWP技术进行气象预测的精度。
上述图3所示实施例中，针对气象预测过程中所涉及到的气象预测装置100(包括上述数据获取模块101、模型确定模块102以及推理模块103)可以是配置于计算设备或者计算设备集群上的软件，并且，通过在计算设备或者计算设备集群上运行该软件，可以使得计算设备或者计算设备集群实现上述气象预测装置100所具有的功能。下面，基于硬件设备实现的角度，对气象预测的过程中所涉及的气象预测装置100进行详细介绍。
图6示出了一种计算设备的结构示意图，上述气象预测装置100可以部署在该计算设备上，该计算设备可以是云环境中的计算设备(如服务器)，或边缘环境中的计算设备，或终端设备等具体可以用于实现上述图3所示实施例中数据获取模块101、模型确定模块102以及推理模块103的功能。
如图6所示，计算设备600包括处理器610、存储器620、通信接口630和总线640。处理器610、存储器620和通信接口630之间通过总线640通信。总线640可以是外设部件互连标准(peripheral component interconnect，PCI)总线或扩展工业标准结构(extended industry standard architecture，EISA)总线等。总线可以分为地址总线、数据总线、控制总线等。为便于表示，图6中仅用一条粗线表示，但并不表示仅有一根总线或一种类型的总线。通信接口630用于与外部通信，例如获取气象数据、目标时间以及位置信息，以及输出气象预测结果等。
其中，处理器610可以为中央处理器(central processing unit，CPU)、专用集成电路(application specific integrated circuit,ASIC)、图形处理器(graphics processing unit,GPU)或者一个或多个集成电路。处理器610还可以是一种集成电路芯片，具有信号的处理能力。在实现过程中，气象预测装置100中各个模块的功能可以通过处理器610中的硬件的集成逻辑电路或者软件形式的指令完成。处理器610还可以是通用处理器、数据信号处理器(digital signal process,DSP)、现场可编程逻辑门阵列(field programmable gate array,FPGA)或者其他可编程逻辑器件，分立门或者晶体管逻辑器件，分立硬件组件，可以实现或者执行本申请实施例中公开的方法、步骤及逻辑框图。其中，通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等，结合本申请实施例所公开的方法可以直接体现为硬件译码处理器执行完成，或者用译码处理器中的硬件及软件模块组合执行完成。软件模块可以位于随机存储器、闪存、只读存储器，可编程只读存储器或者电可擦写可编程存储器、寄存器等本领域成熟的存储介质中。该存储介质位于存储器620，处理器610读取存储器620中的信息，结合其硬件完成气象预测装置100中的部分或全部功能。
存储器620可以包括易失性存储器(volatile memory)，例如随机存取存储器(random access memory，RAM)。存储器620还可以包括非易失性存储器(non-volatile memory)，例如只读存储器(read-only memory，ROM)，快闪存储器，HDD或SSD。
存储器620中存储有可执行代码，处理器610执行该可执行代码以执行前述气象预测装置100所执行的方法。
具体地，在实现图3所示实施例的情况下，且图3所示实施例中所描述的数据获取模块101、模型确定模块102以及推理模块103为通过软件实现的情况下，执行图3中的数据获取 模块101、模型确定模块102以及推理模块103的功能所需的软件或程序代码存储在存储器620中，数据获取模块101与其它设备的交互通过通信接口630实现，处理器用于执行存储器620中的指令，实现气象预测装置100所执行的方法。
图7示出的一种计算设备集群的结构示意图。其中，图7所示的计算设备集群70包括多个计算设备，上述气象预测装置100可以分布式地部署在该计算设备集群70中的多个计算设备上。如图7所示，计算设备集群70包括多个计算设备700，每个计算设备700包括存储器720、处理器710、通信接口730以及总线740，其中，存储器720、处理器710、通信接口730通过总线740实现彼此之间的通信连接。
处理器710可以采用CPU、GPU、ASIC或者一个或多个集成电路。处理器710还可以是一种集成电路芯片，具有信号的处理能力。在实现过程中，气象预测装置100的部分功能可用通过处理器710中的硬件的集成逻辑电路或者软件形式的指令完成。处理器710还可以是DSP、FPGA、通用处理器、其他可编程逻辑器件，分立门或者晶体管逻辑器件、分立硬件组件，可以实现或者执行本申请实施例中公开的部分方法、步骤及逻辑框图。其中，通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等，结合本申请实施例所公开的方法的步骤可以直接体现为硬件译码处理器执行完成，或者用译码处理器中的硬件及软件模块组合执行完成。软件模块可以位于随机存储器、闪存、只读存储器，可编程只读存储器或者电可擦写可编程存储器、寄存器等本领域成熟的存储介质中。该存储介质位于存储器720，在每个计算设备700中，处理器710读取存储器720中的信息，结合其硬件可以完成气象预测装置100的部分功能。
存储器720可以包括ROM、RAM、静态存储设备、动态存储设备、硬盘(例如SSD、HDD)等。存储器720可以存储程序代码，例如，用于实现数据获取模块101的部分或者全部程序代码、用于实现模型确定模块102的部分或者全部程序代码、用于实现推理模块103的部分或者全部程序代码等。针对每个计算设备700，当存储器720中存储的程序代码被处理器710执行时，处理器710基于通信接口730执行气象预测装置100所执行的部分方法，如其中一部分计算设备700可以用于执行上述数据获取模块101所执行的方法，一部分计算设备700可以用于执行上述模型确定模块102所执行的方法、一部分计算设备700用于执行上述推理模块103所执行的方法。存储器720还可以存储数据，例如：处理器710在执行过程中产生的中间数据或结果数据，例如，上述第一气象预测结果、第二气象预测结果、目标气象预测结果等。
每个计算设备700中的通信接口703用于与外部通信，例如与其它计算设备700进行交互等。
总线740可以是外设部件互连标准总线或扩展工业标准结构总线等。为便于表示，图7中每个计算设备700内的总线740仅用一条粗线表示，但并不表示仅有一根总线或一种类型的总线。
上述多个计算设备700之间通过通信网络建立通信通路，以实现气象预测装置100的功能。任一计算设备可以是云环境中的计算设备(例如，服务器)，或边缘环境中的计算设备，或终端设备。
此外，本申请实施例还提供了一种计算机可读存储介质，该计算机可读存储介质中存储有指令，当其在一个或者多个计算设备上运行时，使得该一个或者多个计算设备执行上述实施例气象预测装置100的各个模块所执行的方法。
此外，本申请实施例还提供了一种计算机程序产品，所述计算机程序产品被一个或者多个计算设备执行时，所述一个或者多个计算设备执行前述气象预测方法中的任一方法。该计算机程序产品可以为一个软件安装包，在需要使用前述气象预测方法的任一方法的情况下，可以下载该计算机程序产品并在计算机上执行该计算机程序产品。
另外需说明的是，以上所描述的装置实施例仅仅是示意性的，其中所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部模块来实现本实施例方案的目的。另外，本申请提供的装置实施例附图中，模块之间的连接关系表示它们之间具有通信连接，具体可以实现为一条或多条通信总线或信号线。
通过以上的实施方式的描述，所属领域的技术人员可以清楚地了解到本申请可借助软件加必需的通用硬件的方式来实现，当然也可以通过专用硬件包括专用集成电路、专用CPU、专用存储器、专用元器件等来实现。一般情况下，凡由计算机程序完成的功能都可以很容易地用相应的硬件来实现，而且，用来实现同一功能的具体硬件结构也可以是多种多样的，例如模拟电路、数字电路或专用电路等。但是，对本申请而言更多情况下软件程序实现是更佳的实施方式。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品存储在可读取的存储介质中，如计算机的软盘、U盘、移动硬盘、ROM、RAM、磁碟或者光盘等，包括若干指令用以使得一台计算机设备(可以是个人计算机，训练设备，或者网络设备等)执行本申请各个实施例所述的方法。
在上述实施例中，可以全部或部分地通过软件、硬件、固件或者其任意组合来实现。当使用软件实现时，可以全部或部分地以计算机程序产品的形式实现。
所述计算机程序产品包括一个或多个计算机指令。在计算机上加载和执行所述计算机程序指令时，全部或部分地产生按照本申请实施例所述的流程或功能。所述计算机可以是通用计算机、专用计算机、计算机网络、或者其他可编程装置。所述计算机指令可以存储在计算机可读存储介质中，或者从一个计算机可读存储介质向另一计算机可读存储介质传输，例如，所述计算机指令可以从一个网站站点、计算机、训练设备或数据中心通过有线(例如同轴电缆、光纤、数字用户线(DSL))或无线(例如红外、无线、微波等)方式向另一个网站站点、计算机、训练设备或数据中心进行传输。所述计算机可读存储介质可以是计算机能够存储的任何可用介质或者是包含一个或多个可用介质集成的训练设备、数据中心等数据存储设备。所述可用介质可以是磁性介质，(例如，软盘、硬盘、磁带)、光介质(例如，DVD)、或者半导体介质(例如固态硬盘(Solid State Disk，SSD))等。

Claims (21)

  1. 一种气象预测方法，其特征在于，所述方法包括：
    获取气象数据以及目标时间；
    根据所述目标时间，从模型库中确定多个第一人工智能AI模型，所述多个第一AI模型中不同第一AI模型用于预测不同时间间隔的气象；
    根据所述气象数据，利用所述多个第一AI模型进行推理，得到所述目标时间的第一气象预测结果，每个第一AI模型执行至少一次迭代推理操作。
  2. 根据权利要求1所述的方法，其特征在于，所述根据所述目标时间，从模型库中确定多个第一人工智能AI模型，包括：
    根据所述目标时间，从模型库中确定多个第一人工智能AI模型以及所述多个第一AI模型中各个第一AI模型所需执行的迭代推理操作的次数；
    则，所述根据所述气象数据，利用所述多个第一AI模型进行推理，包括：
    根据所述气象数据、所述各个第一AI模型所需执行的迭代推理操作的次数，利用所述多个第一AI模型依次执行迭代推理操作。
  3. 根据权利要求1或2所述的方法，其特征在于，所述气象数据包括三维气象数据。
  4. 根据权利要求3所述的方法，其特征在于，所述气象数据包括所述三维气象数据以及二维气象数据；
    所述根据所述气象数据，利用所述多个第一AI模型进行推理，包括：
    将所述三维气象数据以及所述二维气象数据进行耦合，得到融合气象数据；
    根据所述融合气象数据，利用所述多个第一AI模型进行推理。
  5. 根据权利要求1至4任一项所述的方法，其特征在于，所述根据所述气象数据，利用所述多个第一AI模型进行推理，包括：
    在利用所述多个第一AI模型对所述气象数据进行推理的过程中，对所述多个第一AI模型中的中间变量或者注意力机制的计算结果增加偏置，所述偏置根据所述气象数据中的高度信息或者纬度信息进行确定。
  6. 根据权利要求1至5任一项所述的方法，其特征在于，所述方法还包括：
    根据所述目标时间，从模型库中确定多个第二AI模型，所述多个第二AI模型中不同第二AI模型用于预测不同时间间隔的气象；
    根据所述气象数据，利用所述多个第二AI模型进行推理，得到所述目标时间的第二气象预测结果，每个第二AI模型执行至少一次迭代推理操作。
  7. 根据权利要求6所述的方法，其特征在于，所述方法还包括：
    根据所述第一气象预测结果以及所述第二气象预测结果，确定目标气象预测结果；
    输出所述目标气象预测结果。
  8. 根据权利要求1至7任一项所述的方法，其特征在于，所述方法还包括：
    获取位置信息，所述第一气象预测结果用于指示所述位置信息对应的气象；
    所述根据所述目标时间，从模型库中确定多个第一人工智能AI模型，包括：
    根据所述位置信息以及所述目标时间，从所述模型库中确定所述多个第一人工智能AI模型。
  9. 根据权利要求1至8任一项所述的方法，其特征在于，所述获取目标时间，包括：
    输出交互界面；
    响应于用户在所述交互界面的操作，获取所述用户输入的所述目标时间。
  10. 一种气象预测装置，其特征在于，所述装置包括：
    数据获取模块，用于获取气象数据以及目标时间；
    模型确定模块，用于根据所述目标时间，从模型库中确定多个第一人工智能AI模型，所述多个第一AI模型中不同第一AI模型用于预测不同时间间隔的气象；
    推理模块，用于根据所述气象数据，利用所述多个第一AI模型进行推理，得到所述目标时间的第一气象预测结果，每个第一AI模型执行至少一次迭代推理操作。
  11. 根据权利要求10所述的装置，其特征在于，所述模型确定模块，用于根据所述目标时间，从模型库中确定多个第一人工智能AI模型以及所述多个第一AI模型中各个第一AI模型所需执行的迭代推理操作的次数；
    所述推理模块，用于根据所述气象数据、所述各个第一AI模型所需执行的迭代推理操作的次数，利用所述多个第一AI模型依次执行迭代推理操作。
  12. 根据权利要求10或11所述的装置，其特征在于，所述气象数据包括三维气象数据。
  13. 根据权利要求12所述的装置，其特征在于，所述气象数据包括所述三维气象数据以及二维气象数据；
    所述推理模块，用于：
    将所述三维气象数据以及所述二维气象数据进行耦合，得到融合气象数据；
    根据所述融合气象数据，利用所述多个第一AI模型进行推理。
  14. 根据权利要求10至13任一项所述的装置，其特征在于，所述推理模块，用于在利用所述多个第一AI模型对所述气象数据进行推理的过程中，对所述多个第一AI模型中的中间变量或者注意力机制的计算结果增加偏置，所述偏置根据所述气象数据中的高度信息或者纬度信息进行确定。
  15. 根据权利要求10至14任一项所述的装置，其特征在于，所述模型确定模块，还用于根据所述目标时间，从模型库中确定多个第二AI模型，所述多个第二AI模型中不同第二AI模型用于预测不同时间间隔的气象；
    所述推理模块，还用于根据所述气象数据，利用所述多个第二AI模型进行推理，得到所述目标时间的第二气象预测结果，每个第二AI模型执行至少一次迭代推理操作。
  16. 根据权利要求15所述的装置，其特征在于，所述推理模块，还用于：
    根据所述第一气象预测结果以及所述第二气象预测结果，确定目标气象预测结果；
    输出所述目标气象预测结果。
  17. 根据权利要求10至16任一项所述的装置，其特征在于，
    所述数据获取模块，还用于获取位置信息，所述第一气象预测结果用于指示所述位置信息对应的气象；
    所述模型确定模块，用于根据所述位置信息以及所述目标时间，从所述模型库中确定所述多个第一人工智能AI模型。
  18. 根据权利要求10至17任一项所述的装置，其特征在于，所述数据获取模块，用于：
    输出交互界面；
    响应于用户在所述交互界面的操作，获取所述用户输入的所述目标时间。
  19. 一种计算设备集群，其特征在于，包括至少一个计算设备，每个计算设备包括处理器和存储器；
    所述处理器用于执行所述存储器中存储的指令，以使得所述计算设备集群执行权利要求1至9中任一项所述的方法。
  20. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质中存储有指令，当其在至少一个计算设备上运行时，使得所述至少一个计算设备执行如权利要求1至9任一项所述的方法。
  21. 一种包含指令的计算机程序产品，其特征在于，当其在至少一个计算设备上运行时，使得所述至少一个计算设备执行如权利要求1至9中任一项所述的方法。
PCT/CN2023/100686 2022-10-31 2023-06-16 一种气象预测方法、装置及相关设备 Ceased WO2024093249A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
EP23884188.6A EP4603879A4 (en) 2022-10-31 2023-06-16 METHOD AND APPARATUS FOR WEATHER FORECASTING AND ASSOCIATED DEVICE
US19/192,997 US20250258318A1 (en) 2022-10-31 2025-04-29 Weather forecast method and apparatus, and related device

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202211351491.5 2022-10-31
CN202211351491.5A CN117991410A (zh) 2022-10-31 2022-10-31 一种气象预测方法、装置及相关设备

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US19/192,997 Continuation US20250258318A1 (en) 2022-10-31 2025-04-29 Weather forecast method and apparatus, and related device

Publications (1)

Publication Number Publication Date
WO2024093249A1 true WO2024093249A1 (zh) 2024-05-10

Family

ID=90891599

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/100686 Ceased WO2024093249A1 (zh) 2022-10-31 2023-06-16 一种气象预测方法、装置及相关设备

Country Status (4)

Country Link
US (1) US20250258318A1 (zh)
EP (1) EP4603879A4 (zh)
CN (1) CN117991410A (zh)
WO (1) WO2024093249A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118568472A (zh) * 2024-08-02 2024-08-30 中国气象局公共气象服务中心(国家预警信息发布中心) 自适应Transformer风速预报订正方法、装置及设备
WO2025237117A1 (zh) * 2024-05-14 2025-11-20 维沃移动通信有限公司 感知处理的方法、装置、设备及介质

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2008134145A (ja) * 2006-11-28 2008-06-12 Toshiba Corp 気象予測データ解析装置及び気象予測データ解析方法
CN111856618A (zh) * 2020-06-11 2020-10-30 上海眼控科技股份有限公司 气象要素的预测方法及设备
KR20220095624A (ko) * 2020-12-30 2022-07-07 주식회사 스튜디오엑스코 머신러닝을 기반으로 하는 기상청 제공 기상영상정보를 활용한 강우확률정보 제공 시스템 및 이를 이용한 방법
CN115016040A (zh) * 2022-08-08 2022-09-06 广东省气象公共服务中心(广东气象影视宣传中心) 基于多模型智能选择的海雾预测方法、装置、设备及介质
CN115097548A (zh) * 2022-08-08 2022-09-23 广东省气象公共服务中心(广东气象影视宣传中心) 基于智能预测的海雾分类预警方法、装置、设备及介质
CN115587650A (zh) * 2022-09-19 2023-01-10 中节能天融科技有限公司 中短期分时段大气常规污染物多目标混合预测方法

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US12423574B2 (en) * 2020-02-18 2025-09-23 Royal Bank Of Canada System and method for weather dependent machine learning architecture

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2008134145A (ja) * 2006-11-28 2008-06-12 Toshiba Corp 気象予測データ解析装置及び気象予測データ解析方法
CN111856618A (zh) * 2020-06-11 2020-10-30 上海眼控科技股份有限公司 气象要素的预测方法及设备
KR20220095624A (ko) * 2020-12-30 2022-07-07 주식회사 스튜디오엑스코 머신러닝을 기반으로 하는 기상청 제공 기상영상정보를 활용한 강우확률정보 제공 시스템 및 이를 이용한 방법
CN115016040A (zh) * 2022-08-08 2022-09-06 广东省气象公共服务中心(广东气象影视宣传中心) 基于多模型智能选择的海雾预测方法、装置、设备及介质
CN115097548A (zh) * 2022-08-08 2022-09-23 广东省气象公共服务中心(广东气象影视宣传中心) 基于智能预测的海雾分类预警方法、装置、设备及介质
CN115587650A (zh) * 2022-09-19 2023-01-10 中节能天融科技有限公司 中短期分时段大气常规污染物多目标混合预测方法

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
See also references of EP4603879A4
SUN, JIAN; CAO, ZHUO; LI, HENG; QIAN, SIMENG; WANG, XIN; YAN, LIMIN; XUE, WEI: "Application of Artificial Intelligence and Internet of Things in Atmospheric Science", JOURNAL OF APPLIED METEOROLOGICAL SCIENCE, vol. 32, no. 1, 31 January 2021 (2021-01-31), CN, pages 1 - 11, XP009554582, ISSN: 1001-7313, DOI: 10.11898/10017313.2021010 *
WOON YANG TAN: "State-of-the-Art Development of Two-Waves Artificial Intelligence Modeling Techniques for River Streamflow Forecasting", ARCHIVES OF COMPUTATIONAL METHODS IN ENGINEERING, vol. 29, no. 7, 1 November 2022 (2022-11-01), Dordrecht , pages 5185 - 5211, XP093168955, ISSN: 1134-3060, DOI: 10.1007/s11831-022-09763-2 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2025237117A1 (zh) * 2024-05-14 2025-11-20 维沃移动通信有限公司 感知处理的方法、装置、设备及介质
CN118568472A (zh) * 2024-08-02 2024-08-30 中国气象局公共气象服务中心(国家预警信息发布中心) 自适应Transformer风速预报订正方法、装置及设备

Also Published As

Publication number Publication date
EP4603879A1 (en) 2025-08-20
EP4603879A4 (en) 2026-01-14
US20250258318A1 (en) 2025-08-14
CN117991410A (zh) 2024-05-07

Similar Documents

Publication Publication Date Title
CN111950225B (zh) 一种芯片布局方法、装置、存储介质和电子设备
US11106567B2 (en) Combinatoric set completion through unique test case generation
CN111723933B (zh) 神经网络模型的训练方法和相关产品
CN113449859B (zh) 一种数据处理方法及其装置
KR102859775B1 (ko) 전자 장치 및 이의 제어 방법
JP7567049B2 (ja) 点群分割方法、装置、機器および記憶媒体
US10198693B2 (en) Method of effective driving behavior extraction using deep learning
CN115423190B (zh) 训练模型的方法及系统和预测序列数据的方法及系统
CN113361680A (zh) 一种神经网络架构搜索方法、装置、设备及介质
CN115061386B (zh) 智能驾驶的自动化仿真测试系统及相关设备
US20250258318A1 (en) Weather forecast method and apparatus, and related device
CN114861910A (zh) 神经网络模型的压缩方法及装置、设备和介质
US11182674B2 (en) Model training by discarding relatively less relevant parameters
EP4174710B1 (en) Generating synthetic training data for perception machine learning models using simulated environments
CN114626503A (zh) 模型的训练方法、目标检测方法、装置、电子设备及介质
CN114897664B (zh) 图模型部署方法及装置、gpu和存储介质
US20190385091A1 (en) Reinforcement learning exploration by exploiting past experiences for critical events
US12412108B2 (en) System and method for inference generation via optimization of inference model portions
CN114785693B (zh) 基于分层强化学习的虚拟网络功能迁移方法及装置
CN113761289B (zh) 图学习方法、框架、计算机系统及可读存储介质
CN111582456B (zh) 用于生成网络模型信息的方法、装置、设备和介质
Chen et al. The implementation and performance evaluation for a smart robot with edge computing algorithms
KR102689100B1 (ko) 시간 가변적 예측(anytime prediction)을 위한 얇은 하위 네트워크를 활용하는 방법 및 시스템
CN118896603B (zh) 全局路径规划方法和移动机器人
CN118708769A (zh) 一种图结构数据的节点预测方法、系统、存储介质和设备

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23884188

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 2023884188

Country of ref document: EP

ENP Entry into the national phase

Ref document number: 2023884188

Country of ref document: EP

Effective date: 20250516

NENP Non-entry into the national phase

Ref country code: DE

WWP Wipo information: published in national office

Ref document number: 2023884188

Country of ref document: EP