JPH1063307A - Feedforward controller in thermal power plant main control system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a feedforward signal in a thermal power plant main control system suitable for obtaining a feedforward signal (manipulation amount) relatively easily from a forward model of an NN and improving control accuracy and stability of a process amount. It is to provide a control device. SOLUTION: Turbine control system 10, water supply control system 11,
In a thermal power plant equipped with a fuel control system 12, a means 1 for obtaining a static operation amount corresponding to a required power generation amount, feedback control means 2 and 3 for providing a control characteristic at the time of output change, and improved control characteristics And feed-forward control means 5 and 6 for constructing a neural network that simulates the plant based on the data of the increase and decrease of the load of the plant, and calculates the Jacobian (the sensitivity of the output to the input) from the neural network. , A feed-forward signal that minimizes an evaluation function consisting of an operation amount is determined using Jacobian.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a control device for a thermal power plant, and more particularly, to a feedforward control device in a thermal power plant main control system, which has a remarkable effect on controlling the main steam temperature and pressure of a boiler.

[0002]

2. Description of the Related Art Thermal power plants include boilers, turbines,
The main control system consists of a generator, a feed water supply system, etc. The main control system is a turbine control system that controls turbine (generator) output, a feed water control system that mainly controls main steam pressure, and a fuel control that mainly controls main steam temperature. System. Each control system controls a turbine (generator) output, a main steam pressure, and a main steam temperature based on a static operation amount corresponding to a power generation request amount. By the way, thermal power plants contain many very complicated nonlinear elements as control characteristics, and the response of process quantities such as main steam pressure and main steam temperature to manipulated variables such as water supply and fuel takes a lot of time. Control is very difficult. Therefore, the feedwater control system and the fuel control system are provided with a feedback control (PI calculation, etc.) function for providing control characteristics at the time of output change and a feedforward control function for improving control characteristics. ing. In other words, when the output of the turbine (generator) is excessively increased or decreased, the process amount such as the main steam pressure and the main steam temperature changes. In order to prevent the change in the process amount, the error between the process amount and the set value is set within a specified range. Feedback control to keep However, when the feedback control has a slow response and it is not possible to suppress the variation within the specified range, the feedforward control is performed by adding a feedforward signal created based on the required power generation to reduce the change in the process amount. Control. In particular, in a main control system of a thermal power plant, feedforward control has a remarkable effect on control for minimizing a change in a process amount such as a main steam temperature and a main steam pressure of a boiler.

Conventionally, the generation of this kind of feedforward signal has been left to a veteran engineer, requiring a great deal of labor. Against this background, there has been an eager desire to automate the creation of feedforward signals. Therefore, various methods have been proposed as algorithms for automatically creating a feedforward signal for minimizing a change in a process amount such as a main steam temperature and a main steam pressure of a boiler. First, there is an application of a neural network (hereinafter referred to as NN), and there are many documents concerning control using this NN. However, as a problem of the NN according to these documents, due to the specificity of the thermal power plant, it is not possible to carry out the number of times necessary to make the NN learn the load change (rise, fall). Therefore, online learning methods such as “feedback learning method” cannot be used. On the other hand, it is conceivable that the NN learns the characteristics of the plant offline from several patterns of load change data. The “generalized learning method” can be applied offline, but is a so-called inverse model learning method in which the output of the plant is used as an input of the NN and the NN is learned using the operation amount as a teacher signal.
In this case, a learning error is inevitably generated in the offline learning, and an error in the operation amount occurs during the online execution. If the control is performed with the manipulated variable, errors accumulate and good control cannot be performed. Further, there is a “specialized learning method” for learning NN based on a deviation between a target value and a plant output. In this case, the sensitivity of the output with respect to the input of the plant (the Jacobian of the plant) is required. It is an unknown plant and its value cannot generally be known. There is also a "forward model learning method" for obtaining a forward model of a plant by the NN.
In this forward model, the input is an operation amount and the output is a control amount. To obtain the operation amount, an inverse function of the non-linear function constituting the NN is obtained, but it is generally difficult. There is a literature that uses the simplex method to search for the manipulated variable, but it is very complicated and difficult to deal with thermal power plants. From the above, it is generally difficult to reversely calculate the operation amount from the forward model. As a patent, for example, Japanese Patent Application Laid-Open No. 6-195104, “Process Control Method”, and as a document, “Basic Measures for Automatic Adjustment of Thermal Power Plant Controller” (Thermal Nuclear Power, No. 456, Vol.
1994), but since the algorithm is complicated and learning is performed based on the PI value while repeatedly increasing and decreasing the load, a large number of load changes are required. In addition, as an application of fuzzy technology, for example, there is “Adjustment technology in thermal power plant control system” (71th Annual Meeting of the Japan Society of Mechanical Engineers, 1993). However, in the fuzzy theory, it is necessary to create a fuzzy rule suitable for a target, and a target load characteristic differs for each plant. Therefore, it is necessary to correct the rule for each plant. Therefore, modifying the rules requires a great deal of effort. In addition, since a feedforward signal can generate only a fixed waveform (for example, trapezoid) or a function (for example, first-order lag) and cannot output an arbitrary waveform, the accuracy of the required process amount deteriorates. It is expected that a case will occur. Also, as an application of adaptive control, MRAC (Model Reference) which is one of adaptive controls
(Adaptive Control Model Reference Adaptive Control). in this way,
(1) Since thermal power plants have gradual process fluctuations,
PE to identify plant parameters (Persi
Since it does not satisfy the condition of “sently exciting”, it is difficult to identify parameters by the sequential least squares method. (2) Noise (small fluctuation of MWD) needs to be added because of initial setting, and there is a fluctuation of the fuel signal at that time. (3) There are problems such as poor stability, and the application is difficult.

[0004]

As described above, conventionally,
There are various algorithms for automatically creating a feedforward signal that minimizes the change in the process amount.However, none of these methods is deterministic, and there are problems with the control accuracy and stability of the process amount with respect to the operation amount. is there.

An object of the present invention is to relatively easily obtain a feedforward signal (manipulation amount) from a forward model of an NN, and to provide a feedforward signal in a thermal power plant main control system suitable for improving control accuracy and stability of a process amount. It is to provide a control device.

[0006]

The object of the present invention is to simulate a plant using an NN, calculate the Jacobian of the NN (the sensitivity of an output to an input), and use the Jacobian to minimize the feedforward function. The problem is solved by obtaining a signal (manipulated variable).

[0007] Since the NN Jacobian corresponds to the coefficient of the predictive expression of the control amount when the plant is linearly approximated, it is relatively easy to obtain a feedforward signal (operating amount) by using this Jacobian. By incorporating this into a control system of a thermal power plant, control accuracy and stability can be improved.

[0008]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 shows a conceptual diagram of plant control of a typical large-capacity commercial thermal power plant. MWD (Mega Watt Demand) based on the output commanded from the central power supply command station or the output set at the power plant
In the creating unit 1, the required output MWD4 is calculated. The MWD 4 is controlled by the turbine control device 10
3 to control the turbine inflow steam supplied from the boiler and control the generator output. At this time, the water supply, fuel, and air must be changed in accordance with the change in the amount of steam flowing into the turbine. However, even if these parameters are changed, the process amount (main steam pressure, temperature, etc.) of the plant cannot be maintained at the specified value. Various control functions are added to keep the plant process amount at a specified value even when the generator output increases or decreases. First,
The main steam pressure is statically expressed as BID (Bo
The main steam pressure is controlled by controlling the feed water pump (BFP) with the feed water control device 11, but the main steam pressure changes when the output of the generator increases or decreases. Therefore, the set value and measurement of the main steam pressure are performed. By taking the deviation of the values and performing PI calculation 2, the main steam pressure is kept constant. However, since the response is slow only by the feedback control, the feedforward signal 5 created based on the MWD signal 4 is added. The BID signal 8 is obtained by adding the value of the PI operation 2, the static balance signal (equal to MWD) 4, and the feedforward signal 5. Also, regarding the main steam temperature, a deviation between the set value and the measured value of the main steam temperature is obtained, and based on the value of the PI calculation 3, the function 7 defining the static balance relationship between the water supply amount and the fuel amount, and the MWD signal 4, To which the feedforward signal 6 created by
(Fire Rate Demand) 9 and the fuel control device 12 controls the fuel (heavy oil, gas, coal, etc.).

Here, in the thermal power plant control system, the control of the main steam temperature is particularly difficult because the dead time is long and nonlinear. FIG. 2 shows a conceptual diagram of the temperature control system. In FIG. 2, the required output (required power generation) M
The following three control signals are used in a transient state in which WD changes in a ramp shape and the load changes. (1) Static fuel signal corresponding to MWD (w (t)) (f) (2) Main steam temperature set value and feedback of measured value (y)
(T)) PI control (3) Feedforward signal (BIR) covering delay of feedback (1) is a fuel amount (f) corresponding to the generator output in a certain settling state. The temperature changes in the transient state when the temperature rises and falls. To prevent that temperature change,
(2) feeds back the measured value (y (t)), and P
Perform I (or P) control. However, the feedback control has a slow response to supply the extra energy (u (t)) required during a transition, and a sufficient control function cannot be obtained. On the other hand, in (3), the temperature change can be reduced by adding a feedforward signal (BIR) created based on MWD (w (t)).

As described above, the points at which the main steam pressure and the like and the main steam pressure and the like are kept at specified values when the output of the generator is increased or decreased are the feedforward signals 5 and 6, and the method of determining these is important. .

The present invention is characterized in that a plant is simulated using NNs, and a feedforward signal (operation amount) that minimizes an evaluation function is obtained by using the NN Jacobian. The details will be described below.

(1) Modeling of Plant As an example of a model configuration of a plant, the response of temperature to fuel is a first-order lag + dead time, the response to MWD is governor and turbine, and the temperature control system of FIG. And linear display approximately. Here, the dead time is small and approximately expressed by a first-order delay, as follows. This model configuration can be expressed in any manner according to the simplicity of the model, and the present embodiment is only an example. When this is represented by a block diagram, it is shown as a simplified model as shown in FIG. FIG. 3 is converted to the simplified model of FIG. From FIG.

(Equation 1) When the equation (1) is z-transformed and represented by a difference equation, the equation (2) is obtained.

(Equation 2) w (t): disturbance / MWD u (t): manipulated variable / fuel (PI output + BIR) y (t): controlled variable / temperature (set temperature-measured temperature) d: dead time Is introduced and the equation (2) is converted, the control amount becomes the equation (3).

(Equation 3) The order is S: d-1, R: 2, BS: d, CS: 2d-
It becomes 1, and R, BS, and CS are represented as follows.

(Equation 4)

(Equation 5)

(Equation 6)

(2) Construction of neural network If a linear approximation of a simplified model of a plant is performed, (3) to
Expression (6) is used. On the other hand, if a non-linear function is applied to apply a neural network, expression (7) is used.

(Equation 7) The nonlinear function of equation (7) is learned by NN. y (t),
y (t-1), y (t-2), u (t), u (t-
1),... U (t−d), w (t + d), w (t +
d-1),... NN for learning y (t + d) is constructed based on the sequential data w (t−d + 1). Here, y (t), y (t-1), y (t-
2) are past MWD values, u (t), u (t−1),.
... U (t−d) is the current value of MWD, w (t +
d), w (t + d-1),..., w (t + 1) are future MWD values. For this NN, load increase and decrease with different load change rates are performed, time series data of y, u, and w are stored, and the time series data is learned off-line. As shown in FIG. 5 (learning data configuration), the learning method uses the input (y (t), y (t−1),...)
And a teacher signal y (t + d) are randomly generated, input one after another, and learned by the back propagation method.

FIG. 6 shows the structure of the NN. FIG. 6 shows a three-layer hierarchical network (three layers in this embodiment,
Any number of layers is acceptable. ), And y (t), y (t-1), y (t-2), u
(T), u (t-1), ... u (td), w (t
+ D), w (t + d−1),..., W (t−d + 1)
And the teacher signal y (t + d) are successively input, and the NN is made to learn. The learning algorithm will be described below.

From FIG. 6, the forward calculation

(Equation 8) Backward calculation

(Equation 9) _{However, O 1k, O 2i, O} 3 ......... network node output w _ik, w _i .................. weight x _k between network nodes, NN .................. network input, output p _k , Q _k , r, s... Linear transformation of network input / output f... Sigmoid function η...…... Learning coefficient DEV... NN output

(3) Feed forward signal (operating amount)
Next, a method of calculating the feedforward signal (manipulation amount) will be described. The manipulated variable u (t) is given by y (t +
The goal is to minimize d). On the other hand, consider minimizing the evaluation function in consideration of the following operation amount u (t).

(Equation 10)

[Equation 11] If equation (11) is set to 0,

(Equation 12) If dy (t + d) / du (t) is obtained, the optimal u
(T) can be obtained.

(4) Calculation of Jacobian Next, the coefficients R _i (t), BS expressed by the equations (4) to (6)
_i (t) and CS _i (t) are obtained by performing linearization every moment.
It can be considered as a Jacobian calculated from the NN in the preceding section. The Jacobian of NN is obtained as follows from FIG.

(Equation 13)

[Equation 14]

(Equation 15) (15) Jacobian J _k (t) (k = 1, 2,...)
.., 3d + 4) are coefficients Ri (i =
0, 1, 2), BSi (i = 0, 1,..., D), CS
i (i = 0, 1,..., 2d-1). here,
Dy (t + d) / du (t) in equation (12) corresponds to BS ₀ . Therefore, the optimal operation amount u shown in Expression (12)
(T) is BS ₀ obtained by calculating the NN Jacobian.
It can be obtained from the value of (t) and y (t + d) obtained every moment. Here, y (t + d) is obtained from equation (7). At the current time t, w (t + d) and w (t + d) are obtained.
Since -1),..., W (t + 1) are future MWD values, these predicted values are predicted and obtained under the condition that the current load change rate will continue in the future.

As described above, in this embodiment, the plant is simulated using the NN, and the Jacobian of the NN is used.
An operation amount for minimizing the evaluation function, that is, a feedforward signal for minimizing a temperature change is automatically and relatively easily generated. The control accuracy and stability of the amount can be improved. The PI control is used for fine adjustment as shown in FIG.

FIG. 7 is a flowchart showing the above procedure. First, a plant model is created, and equation (1) is established based on this model configuration. Next, the equation (1) is z-transformed and represented by a difference equation (2). Then, Dioph
The formula (2) is converted by introducing the antine equation, and the predictive formulas (3) to (6) for the control amount (in the case of temperature) are obtained.
On the other hand, an NN is constructed, weights are calculated by the equations (7) to (9), and the Jacobian is calculated from the NN by the equations (13) to (1).
5) Calculate by equation. From the value obtained by the calculation of the NN Jacobian and the value of equation (7) obtained from time to time, (1
2) Calculate the feedforward signal (manipulation amount) using the equation.

FIGS. 8 and 9 show the results of simulations performed using a simulator that simulates an actual typical constant-pressure once-through plant in detail. 8 and 9,
The horizontal axis is time, and the vertical axis is MWD1, temperature 2, BIR3, PI
Output value 4 is shown. FIG. 8 shows an example of control using only appropriately adjusted PI control, in which a load change rate represented by MWD1 is 2% / min. As the MWD 1 rises and falls, the PI output value 4 changes, but the temperature 2 changes considerably. Here, BIR
3 is 0 because it is not functioning. FIG. 9 shows the control by the feedforward of the present embodiment.
A load change rate of 2% / min, represented by D1,
It is an example of a descent. As MWD1 rises and falls, BI
It can be seen that R3 changes following, the change in temperature 2 is greatly improved, and the temperature change is small.

As described above, the main steam temperature control of the thermal power plant has been described as the embodiment of the present invention.
It is all applicable to feedforward control of non-linear, dead-time plants, let alone other controls in thermal plants.

[0022]

As described above, according to the present invention,
Configure a neural network that simulates a plant,
The Jacobian (the sensitivity of the output to the input) is calculated from this neural network, and by using this Jacobian as a coefficient parameter when the plant is linearly approximated, a feedforward signal (operation amount) that minimizes the evaluation function is obtained. By incorporating this into a control system of a thermal power plant, control accuracy and stability can be improved. Further, according to the present invention, it is possible to automatically and relatively easily create a feedforward signal for minimizing a change in the process amount. Control accuracy and stability can be improved.

[Brief description of the drawings]

FIG. 1 is a conceptual diagram of plant control of a large-scale commercial thermal power plant.

FIG. 2 is a conceptual diagram of a temperature control system.

FIG. 3 is a simplified model

FIG. 4 is a simplified model

FIG. 5 is a diagram showing a configuration of learning data.

FIG. 6 shows the configuration of the NN.

FIG. 7 is a flowchart showing a procedure of the present invention.

FIG. 8 is a simulation example (only PI control).

FIG. 9 is a simulation example (the present invention).

[Explanation of symbols]

Reference Signs List 1 MWD creation unit 2 Main steam pressure control PI calculation 3 Main steam temperature control PI calculation 4 MWD signal 5 Feedwater feedforward signal 6 Fuel feedforward signal 7 Water / fuel ratio function 8 BID (Boiler Input Deman)
d) 9 FRD (Fire Rate Demand) 10 Turbine control device 11 Water supply control device 12 Fuel control device 13 Steam control valve

Claims (1)

[Claims] Claims: 1. A thermal power plant having a turbine control system for controlling the output of a turbine (generator), a feedwater control system for mainly controlling the main steam pressure, and a fuel control system for mainly controlling the main steam temperature. Means for obtaining a static manipulated variable corresponding to the above, feedback control (PI calculation) means for providing control characteristics at the time of output change, and feedforward control means for improving control characteristics, and simulating a plant. A feedforward control device in a thermal power plant main control system, comprising forming a neural network, calculating a Jacobian from the neural network, and obtaining a feedforward signal that minimizes an evaluation function consisting of an operation amount using the Jacobian. .