JPH0553605A - Method and device for process control - Google Patents

Abstract

PURPOSE:To attain the method and device for process control in order to secure the quick, stable and accurate follow-up of the output to the required load despite a large response delay of a controlled variable and the nonlinear responsiveness. CONSTITUTION:A process controller consists of a basic program circuit 830 which produces a basic target command Frdo of a process manipulated variable by a load command Ld, a correction circuit 820 which produces a correction signal from the deviation ET secured between a detected process controlled variable and a set controlled variable, and an acceleration circuit 410 which corrects the command Fedo. The circuit 410 inputs the command Ld and then the deviation ET and produces an acceleration signal AFrd to minimize the deviation ET. The circuit 410 is formed in a neural network or with use of a fuzzy inference.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a process control method and a control device, and more particularly to a large delay in response of a process control amount to a change in process operation amount, such as main steam temperature control of a thermal power plant. The present invention relates to a process control method and apparatus particularly suitable for load follow-up control such as a plant whose response characteristic largely or non-linearly changes depending on a load level or the like.

[0002]

2. Description of the Related Art Conventionally, in a process control device employed in a thermal power plant or the like, in order to control the plant output in conformity with a required load, a small load change is required.
In addition to the basic target command value of the process, which is the operation amount of water supply and fuel in the static balance state of the process that is suitable for the required load, the deviations between the main steam pressure and main steam temperature and their set values are calculated. The correction signals for water supply and fuel are calculated by proportional-plus-integral calculation based on the deviation, these are added to the basic target command value, and by this corrected basic target command,
The method of operating water supply or fuel is widely adopted as a basic control method. However, for a large change in the load demand, a response characteristic that is not always satisfactory can be obtained only by the proportional-plus-integral feedback control as described above. Therefore, the load change is further quantified by a differential operation circuit and the like. Based on the result of (3), a method of adding to the corrected basic target command value as a preceding signal and controlling the water supply and the fuel in advance by this signal to improve the load responsiveness is also used.

[0003]

However, the above-mentioned prior art has the following problems. That is, (1) The preceding signal is generated by an acceleration circuit including a function generator and a rate-of-change limiter with the load change as a variable. In this acceleration circuit, the configuration and circuit parameters of the circuit are empirically created and set each time in consideration of the response characteristics in order to suppress transient fluctuations of the main steam pressure and the main steam temperature, which are process control quantities. There is a need to. At the same time, the performance obtained by the circuit is likely to be influenced by individual differences of designers who set the circuit, and thus it has been difficult to realize optimum controllability in all situations. Also,

(2) When the plant characteristics change depending on the required load level, the preceding signal also needs to be changed correspondingly, but since the characteristic change involves nonlinearity, all load levels are required. It has been difficult to set the optimum acceleration circuit configuration and circuit parameters for each.

Therefore, in view of the above-mentioned problems in the prior art, the object of the present invention is to provide a process control amount with a large response delay, and even if the response characteristic changes non-linearly due to a load level or the like. An object of the present invention is to provide a process control method and a control device therefor capable of accurately causing an output to follow a required load while keeping a deviation between a quantity and its set value to a minimum.

[0006]

According to the present invention, in order to achieve the above object, a process control method and a control device therefor are configured as follows. (1) With the process load command, a deviation between the process control amount and the process set value is input, and an acceleration signal that minimizes the deviation is output.

(2) A neural network is employed to learn an acceleration signal that minimizes the deviation between the process control amount and the set value, and based on the result, an acceleration signal corresponding to the load command input can be created. ..

(3) Using the acceleration signal determination rule as a knowledge base, the deviation between the process control amount and the set value is applied to the membership function to calculate the goodness of fit, and the calculated goodness of fit and the rule are used. It is configured to generate the acceleration signal by so-called fuzzy inference.

(4) In order to correct the acceleration signal, a circuit for calculating the correction signal of the acceleration signal is further provided based on the response characteristic of the process control amount in the result of being added to the process. The neural network is configured to learn the optimum acceleration signal by using the signal corrected by the correction signal as the teaching signal.

(5) The circuit for calculating the correction signal of the acceleration signal uses the rule for determining the correction signal as a knowledge base, and applies the deviation between the process control amount and the set value to the membership function to determine the goodness of fit. Is calculated, and the correction signal is calculated by fuzzy inference using the goodness of fit and the rule.

(6) Based on the basic target command of the process corrected by the acceleration signal, the response characteristic of the process control amount is estimated using a prediction model, and the acceleration signal is estimated based on the estimated response characteristic. Configured to fix.

[0012]

That is, according to the process control method and the control apparatus thereof of the present invention having the above-mentioned configuration, (1) the deviation between the process control amount and the set value is input together with the load command to respond to the change of the load command. Process that outputs the acceleration signal that minimizes the deviation between the process control amount and the set value, so that the plant output accurately follows the load command even for a load command that changes drastically and rapidly. It becomes possible to output the manipulated variable and stably control the process control variable, and to make the plant output accurately follow the load command.

(2) An acceleration signal in which the neural network inputs the load command, the basic target command of the process, and the deviation between the process control amount and the set value, minimizes the deviation, and causes the plant output to accurately follow the load command. , So that the plant output accurately follows the load command while optimally maintaining the process control amount at the desired value even for variously changing load commands. It is possible to output the amount quickly and accurately.

(3) The acceleration signal determination rule is used as a knowledge base, and the acceleration signal is determined by fuzzy inference, so that the know-how of an expert can be held as a knowledge base, and a change in the load command can be performed. However, it is possible to improve the performance with a device that is highly flexible and is relatively simple, and it is possible to eliminate the need for human adjustment work.

(4) The acceleration signal, which is an output signal, is added to the basic target command and added to the plant. Based on the response characteristic of the process control amount obtained as a result, a correction signal for the acceleration signal is obtained and added. Even if the response characteristic of the process changes from the original one by retraining the neural network using the corrected signal as the teaching signal,
A constant response characteristic can always be realized, and desired controllability can always be automatically maintained even if the process state changes.

(5) Knowing the expert's know-how by constructing the acceleration signal correction circuit in the acceleration circuit of the above (4) so that the correction signal is created by fuzzy inference using a rule for determining the correction signal as a knowledge base. Can be incorporated as a modification rule, and further, the acceleration signal can be automatically modified and re-learned by the neural network, so plant control characteristics can be improved in a self-growing manner. Become.

(6) Based on the basic target command of the process corrected by the acceleration signal, a response characteristic of the process control amount is estimated by using a prediction model, and the correction is performed based on the estimated response characteristic. By determining the amount and correcting the acceleration signal with this correction amount, it is possible to prevent disturbance to the plant due to the incompatibility of the acceleration signal in the response characteristic.

[0018]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described in detail below with reference to the accompanying drawings. First, FIG. 2 shows the overall configuration of a thermal power plant to which the process control device of the present invention is preferably applied. This figure shows the equipment configuration of the most typical thermal power plant, and internal fluid such as water discharged from the water supply pump 100 is a high-pressure feed water heater 110 and a economizer 120 provided in the boiler 105. The furnace water cooling wall 130, the primary superheater 140 raises and superheats the temperature, the desuperheater 150 reduces the temperature, and the secondary superheater 160 again.
And is supplied to the high-pressure turbine 180. In addition,
The desuperheater 150 reduces the temperature of the secondary superheater inlet steam of the internal fluid that has been heated / heated to a specified value. Subsequently, the steam that has worked in the high-pressure turbine 180 is heated again in the primary reheater 190 and the secondary reheater 200, and is supplied to the medium / low pressure turbine 210. In addition,
The steam that worked in the low-pressure turbine 210 is returned to the condenser 230.
The water is reconstituted in and is supplied to the water supply pump 100 again. A low-pressure feed water heater, a desalting device, a deaerator, etc. are generally provided between the condenser 230 and the water supply pump 100. However, these devices are not directly related to the present invention. Therefore, it is omitted from the description. As mentioned above,
The energy given from the steam to the high-pressure turbine 180 and the medium / low-pressure turbine 210 becomes rotational energy to rotate the generator 220, and thus is converted into electric energy.

The control of the thermal power plant as described above is generally performed as follows. That is, the electrical output is
The main turbine control valve 170 is operated and controlled by the deviation signal between the load demand and the output detector 560 signal. Also,
Main steam pressure / temperature, reheat steam temperature, gas O ₂ , and 2
The inlet superheater steam temperature is the main steam pressure detector 510, the main steam temperature detector 500, the reheat steam temperature detector 550, the gas O ₂ detector 570, and the steam temperature detectors before and after the desuperheater. Signals from 520 and 530 are used to adjust the regulating valve 250 of the turbine 240 for driving the water supply pump,
A fuel amount control valve 260, a recirculation gas amount control damper 280,
Air amount adjusting damper 270, and spray amount adjusting valve 290
It is controlled by operating.

FIG. 3 shows a control system for realizing the control of the thermal power plant described above. First, the basic part will be described. The main turbine control valve controller 710 controls the main turbine control valve 170 by the load command value Ld from the load command setting device 700 and the actual output La from the output detector 560 to control the electrical output. Control. In addition, the main steam pressure controller 7
40 is a load command setting via the adder 750 based on the deviation between the output from the main steam pressure detector 510 and the output from the main steam pressure setter 720 (that is, the output of the subtractor 730). Pressure correction is added to the load command value Ld from the device 700 to obtain the boiler input demand Bid. This boiler input demand Bid is converted into a water supply demand Fwd by the water supply basic program 755, and the water supply flow controller 76 is supplied.
0 indicates the value Fwd of the water supply demand and the water supply detector 57.
Water supply pump drive turbine control valve 2 according to the output from 1
Control 50.

On the other hand, the boiler input demand Bid output from the adder 750 is a static fuel demand signal Frd which is a basic target command by the fuel basic program 830.
is converted into o, and the deviation between the output of the main steam temperature detector 500 and the output of the main steam temperature setter 800 (subtractor 81
The fuel quantity demand Frd is obtained by being corrected by the output from the main steam temperature controller 820 corresponding to the (zero output). That is, the boiler input demand Bid is converted into the static fuel demand signal Frdo by the fuel basic program 830, and then passes through the adder 440 and then the adder 84.
Lead to zero. Then, the adder 840 adds the output from the main steam temperature controller 820 corresponding to the deviation between the output of the main steam temperature detector 500 and the output of the main steam temperature setting device 800 (output of the subtractor 810). Therefore, the fuel amount demand Frd is output. Then, according to the value of the fuel amount demand Frd and the output from the fuel amount detector 572,
The fuel amount controller 850 controls the fuel amount control valve 260.
Further, the output of the gas O ₂ controller 880 corresponding to the deviation (output of the subtractor 870) between the output from the gas O ₂ detector 570 and the output from the gas O ₂ setting device 860 is also output via the adder 900. The air amount demand Fad is obtained by adding the fuel amount demand Frd as a correction signal to the value of the static air amount demand signal Fado obtained by converting the fuel amount demand Frd by the air basic program 890. And this air demand F
The air amount controller 910 controls the air amount adjusting damper 270 according to the value of ad and the output from the air amount detector 573.

Further, the reheat steam temperature controller 940 produces an output Fgrd corresponding to the deviation (output of the subtractor 930) between the output from the reheat steam temperature detector 550 and the output from the reheat steam temperature setter 920. The recirculation gas amount controller 950 controls the opening degree of the recirculation gas amount adjustment damper 280 according to the value of the output Fgrd and the output from the recirculation gas amount detector 574. Similarly, the secondary superheater inlet steam temperature controller corresponding to the deviation (output of the subtractor 970) between the output from the secondary superheater inlet steam temperature detector 520 and the output from the temperature setter 960. The output of 980 determines the primary superheater outlet steam temperature T1SH, d, and according to this T1SH, d value and the output of the primary superheater outlet steam temperature detector 530, the primary superheater outlet steam temperature controller 990 will control the superheater spray amount control valve 290.

The above is the description of the operation of the basic part of the control system of the thermal power plant. In addition to the basic part of the control system described above, the electric output is faithfully adjusted to the required value in response to a large change in the load command. At the same time, in order to stably control the main steam pressure and various steam temperatures to set values while following them, correction signals are created, and these correction signals are respectively supplied to the water supply demand signal Fwd and the fuel demand signal Frd described above. A circuit for outputting an acceleration signal to the air amount demand signal Fad is added. That is, in FIG. 3, the first acceleration circuit 400 outputs the acceleration signal AFwd and corrects the output from the water supply basic program 755 via the adder 430. In addition, the second acceleration circuit 410 outputs the acceleration signal AFrd.
To correct the static fuel demand signal Frdo output from the fuel basic program 830 via the adder 440, and the third acceleration circuit 420 outputs the acceleration signal A
Fad is output, and thus the air amount demand Fad output from the adder 900 is corrected via the adder 450.

FIG. 1 shows only the construction of the fuel control section in which the process control method according to the present invention is adopted from the overall construction of the thermal power plant described above.
In the structure of this fuel control unit, the acceleration circuit 410 is particularly characteristic in comparison with the related art, and this acceleration circuit 410 includes the load command setter 70.
In addition to the load command value Ld from 0, it is the deviation between the output signal from the main steam temperature detector 500 which is a so-called process control amount and the output signal from the main steam temperature setter 800 which is a process set value. Enter the so-called control deviation ΕT,
The point is that an acceleration signal is output based on these. That is, the acceleration circuit 410 outputs an acceleration signal AFrd that minimizes the control deviation ΕT of the main steam temperature based on the load command value Ld and the control deviation ΕT of the main steam temperature. Through the adder 440, the fuel basic target command F from the fuel basic program 830.
Added to rdo. Furthermore, this fuel basic target command Frdo
The output formed by the main steam temperature controller 820 is added by the adder 840 based on the control deviation ΕT of the main steam temperature, and the result is output as the fuel amount demand Frd. Then, the fuel amount controller 850 controls the fuel amount adjusting valve 260 based on the fuel amount demand Frd and the output from the fuel amount detector 572.

The output waveform diagram of FIG. 4 illustrates the operation of the acceleration circuit 410 in the present invention. As is clear from the figure, a static fuel for statically controlling the main steam temperature to a set value against a sudden change in the load command value Ld, which is a load request (see (a) of FIG. 4). Quantity demand signal Frdo
In the case of only (see the broken line in (b) of FIG. 4), the deviation ET of the main steam temperature (see (d) of FIG. 4) shows the heat transfer delay in the transient state as shown by the broken line ETo. A large temperature deviation is generated due to a flow delay, and it takes time to settle the temperature. Therefore, by adding the fuel amount acceleration signal AFrd from the acceleration circuit 410 as a correction signal to the static fuel demand Frdo to obtain the fuel amount demand signal Frd (see the solid line in FIG. 4B), the main steam temperature The main steam temperature deviation ΕT can be made small as shown by the solid line ET in FIG. Incidentally, in (c) and (d) of the same figure, if the value of the fuel amount acceleration signal AFrd at time ti is AFrd (ti),
This effect is due to the main steam temperature deviation ΕT after the response delay time L due to heat transfer delay or flow delay, that is, at time ti + L.
It is represented as (ti + L). Therefore, the main steam temperature deviation ΕT at the time ti + L after the delay time L
In order to reduce (ti + L), it is important to properly select the fuel amount acceleration signal AFrd (ti) at the time ti.

FIG. 5 shows an embodiment in which the acceleration circuit is constructed by a so-called neural network in order to appropriately select the fuel amount acceleration signal AFrd. Hereinafter, in this figure, the neural network is composed of three layers of an input layer, an intermediate layer, and an output layer. In this embodiment, as the input signal, the slope θ of the load command Ld and the level H ((a in FIG. 4) are used. )), The main steam temperature deviation values during the time T after the time t ₀ + L, that is, ΕT (t ₀ + L), ΕT (t ₁ +
L) ... ET (t _i + L) ... ET (t _N + L) are used as output signals, and acceleration signals AFrd (t ₀ ), AFrd (t ₁ ) ...
AFrd (t _i ) ... AFrd (t _N ) is output (see (c) and (d) of FIG. 4). Here, N represents N = T / Δ (Δ: sampling period, T: time required for the main steam temperature to settle).

On the other hand, in the neural network NN,
First, the switch circuits SW ₁ , SW ₂ , SW _{3O to} SW _3N , S
W _{4O to} SW _4N are switched to the b side, and various teaching patterns are input from the teaching pattern input circuit TCHi. For example,
From the switch circuits SW ₁ and SW ₂ , the inclination ^ θ and the level ^ H of the load demand Ld of various patterns are obtained, and from the switch circuits SW _{3O to} SW _3N , the desired main steam temperature deviation ^ ET (t ₀ + L). ), ^ ET (t ₁ + L) ... ^ ET (t _i +
L) ... ^ ET (t _N + L) is input respectively. At the same time, the outputs AFrd ′ (t ₀ ), AFrd ′ of the output layer at this time
(T ₁ ) ... AFrd '(t _i ) ... AFrd' (t _N ) and desired acceleration signal output ^ AFrd (t ₀ ), ^ A corresponding to the teaching pattern input from the teaching pattern output circuit TCho.
Frd (t ₁ ) ... ^ AFrd (t _i ) ... ^ AFrd (t _N ) is compared by the comparison circuits CM, CM ... Of the learning unit LC, and the first learning control is performed so that the outputs of both of them may match. Circuit L
The CCK1 learns the synapse weighting of the input circuit of the neuron in the output layer, while the second learning control circuit LCCK2 adjusts the synapse weighting of the input circuit of the neuron in the middle layer to adjust various load patterns. Train the students to obtain appropriate output. As a result, in the control mode, the above switch circuit S
By switching W ₁ , SW ₂ , SW _{3o to} SW _3N , and SW _{40 to} SW _4N from the b side to the a side, the neural network can be applied to the input that is the same as or close to the above learning pattern. The NN outputs an acceleration signal that is the same as or similar to the pattern at the time of learning.

Next, the structure of the neural network NN and the learning method in the learning unit LC will be described. FIG. 6 shows a unit model of one neuron that constitutes the neural network NN. Here, the input signals X1, X2, ... Xn to the unit 11 take the range (0, 1),
Synapse weights W1, W2, ... Wn are in the range (-∞, + ∞)
Shall be taken. Here, the input Ui transmitted from the i-th input Xi to the unit is

[Equation 1] Then the total input U to the unit is

[Equation 2] Becomes The unit output Y is

[Equation 3] Define in. However, in the equation 3, U0 is a bias. In this embodiment, the unit models 11 described above are arranged in layers as shown in FIG.
The neural network NN has a configuration in which the output signal from 1 is used as an input signal to each unit 11 in the next layer. Regarding the structures of the unit model 11 and the neural network NN, for example, see
Tea Press, Neurocomputing Foundations of Research, 1988, pages 318-362 (The MIT Press, Neurocomputing Found
ations of Research, 1988, pp 318-362). Also, in this paper, as shown in FIG. 7, when a certain input signal pattern 12 is given to the input layer, the output signal pattern 13 becomes a desired signal pattern, that is, a teacher signal pattern 14, depending on the error between the two. A learning algorithm (referred to as backpropagation) for correcting the connection strength of the input unit to each unit of the intermediate layer, that is, the synapse weighting is shown.

Also, in the learning section LC of this embodiment,
The learning algorithm itself uses the backpropagation shown in the above paper.

FIG. 8 specifically explains the backpropagation algorithm. In this figure, in order to make the algorithm easy to understand, the kth output signal Y3k of the output layer is focused and this is used as a teacher signal. 9 shows a procedure for modifying synaptic weights to match Ytk. That is, specifically describing the algorithm shown in FIG.
First, the error e between the kth output signal Y3k and the teacher signal Ytk
k is

[Equation 4] It is represented by. If the influence degree of the error at the unit operation level U3k is d3k, the influence degree of this error d3k is

[Equation 5] It is represented by. However, f ′ (U) is as in the following Expression 6.

[Equation 6]

Therefore, the correction amount ΔW3k, 2j (N + 1) of the synapse weight W3k, 2j (N + 1) at the jth input section in the kth unit of the output layer is expressed by the following equation 7. ..

[Equation 7] Here, N is a symbol representing the previous time, and η is called a learning constant. Further, y2j indicates the j-th output signal of the intermediate layer. However, in order to realize stable convergence, the correction amount obtained in the equation 7 is not used as it is, but is corrected by the method represented by the following equation 8.
A new synaptic weight W3k, 2j (N + 1) is obtained by the following equation.

[Equation 8] Here, α is called a stabilization constant. The method of correcting the synaptic weight of the input part of the output layer has been described above.

Next, a method of correcting the synapse weight of the input section of the intermediate layer will be described. In FIG. 8 above, the synaptic weight W2j, 1i at the i-th input of the j-th unit in the hidden layer is
Paying attention to, and showing the correction method. The influence degree d2j of the error in the unit operation level U2j in this case should be determined in consideration of the error of all unit outputs of the output layer, and is represented by the following Expression 9.

[Equation 9]

Therefore, the correction amount ΔW2j, 1i of the synapse weight at the i-th input portion in the j-th unit of the intermediate layer is obtained.
(N + 1) is expressed by the following equation.

[Equation 10] Here, N is a symbol indicating the previous time, and η is called a learning constant. Further, y1i represents the i-th output signal of the input layer. However, as in the case of the output layer, to achieve stable convergence,
Instead of using the correction amount obtained by the above equation 10 as it is, it is corrected by the method represented by the following equation, and a new synaptic weight W2j, 1i (N +
1) is obtained as the following formula 11.

[Equation 11] Here, α is called a stabilization constant.

In this way, the error ek can be minimized by repeating the arithmetic processing represented by the above equations 4 to 11. That is, the output signal pattern from the output layer can be matched with the teacher signal pattern, and as a result, the input signal pattern is stored (learned) as a synapse weighted distribution (that is, in-circuit connection strength distribution) in the neural network. It will be. If another input signal pattern is presented to the input layer and the teacher signal pattern is also presented correspondingly to this pattern, the above algorithm operates and the new synapse weight distribution is stored.

By using such an algorithm, it is possible to store a plurality of learning samples in the same neural network. By using the neural network after learning is completed, when the same pattern as the already learned pattern is input, the same output signal pattern used at the time of learning is output from the output layer. Further, even if an unlearned pattern is input, an output signal pattern similar to the learned pattern can be obtained.

Next, FIG. 9 shows an embodiment in which fuzzy inference is applied to the above-described acceleration circuit 410 to calculate an acceleration signal. In this example,
In the load command characteristic quantification unit 1100, the load command Ld
Is quantified as a slope θ and a level H. In addition, the control characteristic quantification unit 1200 displays the past acceleration signal AFrd ^P
(To) ... AFrd ^P (tN) and main steam temperature deviation signal Et
^P (to + L) ... evaluates control characteristics as a function of Et ^P (tN + L), a quantitative as the value of the evaluation index I _E. On the other hand, the storage unit 1300 accumulates and stores a required amount of historical data of past acceleration signals and main steam temperature deviation signals. Then, the fuzzy inference unit 1400 applies the feature amount of the load command Ld and the value of the evaluation index I _{E to} the membership function shown in FIG. 10 to determine the degree of suitability for each. Here, the symbols in the figure are as follows: NB: Negative Big (negative large) NM: Negative Medium (negative) ZO: Zero (zero) PS: Positive Small (PM) Positive Medium (Positive middle) PB: Positive Big (large positive).

Next, the value of the fitness of the membership function described above is applied to the acceleration signal generation rule stored in the knowledge base 1500 to the antecedent part (IF) of the fuzzy inference unit 1400, and fuzzy inference is performed by fuzzy logic operation. The conclusions are drawn in part 1400, the companion part (THEN). That is, in the case of this embodiment, the front companion part of the rule 1 First, theta, H, (wherein, X.theta, xH, xI and rear) value of I _E is applied to the fitness ω1 by the following equation (12) Ask.

[Equation 12] In the present embodiment, ω1 = PS (xI), and from this, acceleration signals AFrd (ti), AFrd (ti + 1) ...
AFrd (ti + M) are the shaded areas Ai, Ai + 1, Ai + 2, respectively.
… (Consequent part of the fuzzy inference unit 1400 in FIG. 9 (THE
N))). Here, M indicates a period during which the acceleration signal is output.

Next, with respect to rule 2, the same inference operation is performed,

[Equation 13] In the present embodiment, ω2 = PS (xI), and from this, the acceleration signals according to rule 2 are respectively shaded areas Bi,
Bi + 1, Bi + 2 ... (See the consequent part of the fuzzy inference unit 1400 in FIG. 9).

In this embodiment, only Rule 1 and Rule 2 described above are applied, and the results obtained by integrating the calculation results of both rules are respectively AFrd (t
i), AFrd (ti + 1) ... AFrd (ti + M), the respective values are the shaded areas Ai and Bi, Ai + 1 and Bi + 1 ... Ai + M and Bi.
+ M is calculated as the x-coordinate of the center of gravity of the shaded area, which is the maximum value (the consequent part of the fuzzy inference unit 1400 (T
(See the bottom diagram of HEN section).

Next, FIG. 11 shows another embodiment in which, in the embodiment of FIG. 5 described above, further learning is performed by looking at the response result of the process by the acceleration signal AFrd (ti) calculated by the neural network NN. Is shown.
In another embodiment, whether or not the acceleration signal is appropriate is determined based on the response result of the process by the acceleration signal AFrd (ti), and if it is inappropriate, the correction signal ΔAFrd ( ti) to ΔAFrd (ti + N) are created, and the neural network NN is made to re-learn the optimum acceleration signal by using the acceleration signal corrected by this correction signal as a teaching signal. That is, the response characteristic evaluation unit PCE
When the response characteristic of the main steam temperature deviation ET does not satisfy the target performance, the acceleration signal correction signal generating circuit ACC generates the correction signals ΔAFrd (ti) to ΔAFrd (ti + N) of the acceleration signal, and It is added as a correction signal to the output of the neural network NN via the adder AD, and the result is stored in the output teaching pattern OTP as the optimum acceleration signal. Then, the switch circuits SW1, SW2, SW30 to SW3N, SW
40 to SW4N are switched to the learning mode on the side b, and the learning unit LC causes the neuron units of the intermediate layer and the output layer of the neural network NN so that the output of the neural network NN matches the optimum acceleration signal of the output teaching pattern OTP. Adjust individual synaptic weights. When the output of the neural network NN coincides with the output teaching pattern, re-learning is finished, and the switches SW1, SW2, SW
The normal control mode is set by switching 30 to SW3N and SW40 to SW4N to the a side.

Further, FIG. 12 shows the response characteristic evaluation unit PCE and the acceleration signal correction signal generation circuit ACC in FIG.
In the present embodiment, fuzzy inference is applied to the acceleration signal correction signal generation circuit ACC. First, the response characteristic evaluation unit PCE quantifies the temporal change of the control deviation ET with respect to the operation of the acceleration signal AFrd and stores them. The fuzzy inference unit 1600 applies the above-mentioned value of the control deviation signal ET to the membership function shown in FIG. 10 to determine the suitability for each. Next, knowledge base 1
The value of the fitness of the membership function described above is added to the acceleration signal correction amount generation rule stored in 700, and the fuzzy inference unit 16
It is applied to the antecedent part (IF) of 00, and the conclusion is derived by the consequent part (THEN) of the fuzzy inference unit 1600 by fuzzy logic operation.

That is, in the case of this embodiment, first, in the antecedent part of rule 1, the control deviations ET (ti + L) and ET (ti + 1 + L)
Value (here, xi, xi + 1) is applied, and the goodness of fit ω
1 is calculated by the following equation 14.

[Equation 14] In this embodiment, ω1 = PB (xi), and from this, the acceleration signal correction signals ΔAFrd (ti) and ΔAFr according to rule 1 are obtained.
d (ti + 1) is obtained as the shaded portions Ai and Ai + 1 (see the consequent portion of the fuzzy inference unit 1600 in FIG. 12). Next, the same inference operation is performed for rule 2,

[Equation 15] In the present embodiment, ω2 = PS (xi), and from this, the acceleration signal correction signal according to rule 2 is obtained as the shaded portions Bi and Bi + 1 (see the consequent portion of the fuzzy inference unit 1600 in FIG. 12).

In this embodiment, rule 1 and rule 2
It is assumed that only the results are applied, and the results obtained by integrating the calculation results of both rules are ΔAFrd (ti),
If ΔAFrd (ti + 1) is set, each value is obtained as the x-coordinate of the center of gravity of the shaded area formed by the maximum values of the shaded areas Ai and Bi and Ai + 1 and Bi + 1 (fuzzy in FIG. 12). Reasoning unit 1
(See the bottom diagram of the consequent part of 600).

Further, in this embodiment, the control deviation ET is focused only on two points of time ti + L and ti + 1 + L, and the acceleration signal correction signal ΔAFrd also corresponds to this, two points of ti and ti + 1. (Ie, ΔAFrd (t
i), ΔAFrd (ti + 1)), but not limited to this, also in the case of handling data at three or more time points, the acceleration signal correction signal ΔA is processed by the same method as described above.
Frd can be generated.

In FIG. 13, further, before the fuel quantity demand signal Frd is given to the fuel quantity controller 850, the fuel quantity demand signal Frd is given to the plant prediction model as a trial signal to predict the main steam temperature deviation ET, When the response characteristic does not satisfy the target performance, the acceleration signal ΔAFrd is corrected, and when it is confirmed that the target performance is satisfied, after confirming that, the fuel amount controller 850 is operated to confirm the fuel consumption. An embodiment is shown configured to provide a quantity demand signal Frd. Here, the fuel control signal CS is given to the so-called control response prediction model RPM as the fuel amount demand signal Frd before being given to the fuel regulating valve 260. In this control response prediction model RPM, the predicted value of the main steam temperature deviation ^ E
T is obtained and given to the response characteristic evaluation unit PCE. In the response characteristic evaluation unit PCE, the predicted value of this main steam temperature deviation ^
The response characteristic is evaluated by the evaluation function value based on ET, and as a result, if the evaluation standard is satisfied, the switch S
W is switched to the b side and the control signal CS is set to the fuel adjustment valve 260.
On the other hand, if the evaluation criterion is not satisfied, the predicted value ^ ET of the main steam temperature deviation is given to the acceleration signal correction signal generation circuit ACC. Then, the acceleration signal correction signal generation circuit A
The CC creates an acceleration signal correction signal ΔAFrd on the basis of this predicted value ^ ET and adds it to the acceleration signal AFrd via the adder ADD to generate an acceleration signal AFrd ', thereby generating a new ^ Frd. create. Hereinafter, the above procedure is repeated until the evaluation criteria are satisfied by PCE.

The acceleration signal correction signal generation circuit AC
C considers ET in FIG. 12 as ^ ET,
It can also be realized by the same method as in FIG. In addition,
The control response prediction model RPM in FIG. 13 can be realized by the following mathematical expression when the secondary superheater is targeted.

[Equation 16]

[Equation 17]

[Equation 18] However, in the above formula, Ts, Tg, Ks, Kg, Kf
Are respectively expressed by the following mathematical formulas.

[Formula 19]

[Equation 20]

[Equation 21]

[Equation 22]

[Equation 23] Here, each code represents the following quantity. Vs: Pipe flow volume γs: Internal fluid specific weight Fs: Internal fluid flow rate Cpo: Secondary superheater outlet steam constant pressure specific heat Cpi: Secondary superheater inlet steam constant pressure specific heat A: Secondary superheater heat transfer area αms: From pipe Heat transfer rate to steam θso: Secondary superheater outlet steam temperature θsi: Secondary superheater inlet steam temperature θm: Pipe temperature θs: Internal fluid temperature Cm: Pipe specific heat Mm: Pipe weight αgm: Heat transfer from gas to pipe Rate Ff: Fuel flow rate Fa: Air flow rate Fgγf: Recirculated gas flow rate Fgβf: Boiler gas flow rate

In the above-mentioned embodiments, the example in which the present invention is applied to the fuel control of the thermal power plant has been described, but the present invention is not limited to this, and is also applied to, for example, water supply control and air control. However, it is clear that the same effect can be obtained.

[0048]

As is apparent from the above detailed description, the process control method and its control device of the present invention have the following effects. (1) The acceleration circuit responds to changes in the load command and outputs an acceleration signal that always minimizes the deviation between the process control amount and the set value. Even if the output of the plant is made to accurately follow the load command, the process amount can be stably controlled. (2) By configuring the acceleration circuit with a neural network, it becomes possible to quickly and accurately output the process operation amount for keeping the process control amount at a desired value even for various load commands. (3) By configuring the accelerating circuit to determine the accelerating signal by fuzzy inference using the accelerating signal decision rule as the knowledge base, it is possible to have the know-how of the expert as the knowledge base and improve the performance with a simple device. In addition to being able to achieve this, human work for adjusting the circuit becomes unnecessary. (4) By providing a correction function of the output so that the optimum acceleration output can be obtained from the acceleration circuit according to the response characteristics of the process based on the acceleration signal, the desired controllability is automatically maintained even if the process state changes. It becomes possible to maintain. (5) By realizing the correction function by means of a correction rule and fuzzy inference, the know-how of a skilled person can be incorporated as a correction rule, and the acceleration signal is automatically corrected so that the neural network re-learns it. Therefore, the plant control characteristics can be improved in a self-growing manner. (6) The control response is predicted in advance before the manipulated variable is output to the plant, and the acceleration signal is output based on the evaluation result to prevent the control disturbance due to an inappropriate manipulated variable. Therefore, safe and reliable control can be realized.

[Brief description of drawings]

FIG. 1 is a partially enlarged view of a control system diagram of a thermal power plant showing a characteristic configuration of a process control device of the present invention.

FIG. 2 is a control system diagram showing the overall equipment configuration of a thermal power plant to which the process control method according to the present invention is applied.

FIG. 3 is a control system diagram showing a configuration of a process control device of the thermal power plant of FIG.

FIG. 4 is a diagram showing a signal waveform of each part for explaining the operation of the speed circuit.

FIG. 5 is a control conceptual diagram showing an embodiment in which the acceleration circuit is configured by a neural network.

FIG. 6 is a control conceptual diagram showing a neuron unit model forming the neural network.

FIG. 7 is a control conceptual diagram illustrating the structure and operation of the neural network.

FIG. 8 is a conceptual control diagram explaining the structure and learning mechanism of the neural network.

FIG. 9 is a control conceptual diagram illustrating the configuration of another embodiment in which fuzzy inference is applied to the acceleration circuit.

FIG. 10 is a diagram showing an example of a membership function of the fuzzy inference.

FIG. 11 is a control conceptual diagram for explaining the configuration of another embodiment in which the acceleration circuit shown in FIG. 5 is further provided with an acceleration signal correction function.

FIG. 12 is a control conceptual diagram showing the configuration of an embodiment in which fuzzy inference is applied to the acceleration signal correction function of FIG.

FIG. 13 is a control conceptual diagram showing a configuration of another embodiment in which a response is evaluated in advance by using a prediction model before outputting the acceleration signal.

[Explanation of symbols]

 Ld load command value 700 load command setter 500, 510, 570 control amount detector 800, 860, 920 setter ET control deviation 400, 410, 420 acceleration circuit 430, 440, 450 adder 755, 830, 890 basic program Fwdo , Frdo, Fido Basic target command Fwd, Frd, Fid Demand signal AFwd, AFrd, AFid Acceleration signal 760, 850, 910 Controller △ AFrd Fuel acceleration signal correction signal 260 Fuel adjustment valve 270 Air amount adjustment damper 572 Fuel amount detector NN Neural network 11 Unit model of neural network LC learning unit TCHi teaching pattern input circuit TCHo teaching pattern output circuit LCCK1, 2 learning control circuit

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Masahide Nomura 4026 Kuji-cho, Hitachi City, Ibaraki Prefecture Hitachi Research Laboratory, Hitachi, Ltd.

Claims (12)

[Claims] 1. A basic target command is created from a load command of a process, a correction signal for correcting the basic target command is formed from a deviation between a detected process control amount and a control amount set value, and the load command is changed. Forming an acceleration signal to correct the basic target command based on the situation, these basic target command,
In a process control method for determining a manipulated variable of the process using a correction signal and an acceleration signal, the acceleration signal is used together with the change status of the load command, and the detected process control variable and control variable set value A process control method, wherein the deviation is minimized so that the deviation is minimized. 2. The neural network according to claim 1, wherein learning is performed using a neural network so as to minimize a deviation between the detected process control amount and a control amount set value, and based on the result, the load command is changed. A process control method characterized in that the acceleration signal is generated correspondingly. 3. The conformity degree and the rule according to claim 1, wherein the acceleration signal determination rule is used as a knowledge base to apply a deviation between the process control amount and the control amount set value to a membership function to calculate the conformity degree. A process control method characterized in that an acceleration signal is generated by fuzzy inference using. 4. The correction signal for the acceleration signal according to claim 2, wherein a correction signal for the acceleration signal is generated based on a response characteristic of the process control amount when the acceleration signal generated by the neural network is applied. A process control method comprising: modifying with the modification signal to create a teaching signal; and using the teaching signal to cause the neural network to learn an optimum acceleration signal. 5. The correction signal according to claim 4, wherein the rule for determining the correction signal is used as a knowledge base, and the deviation between the process control amount and the control amount set value and the response characteristic are applied to a membership function. The process control method is characterized by calculating the goodness of fit by means of fuzzy inference using the goodness of fit and the rule. 6. The response characteristic of the process control amount is estimated by using a prediction model based on the basic target command of the process corrected by the acceleration signal,
A process control method comprising modifying the acceleration signal based on the estimated response characteristic. 7. A basic program circuit for creating a basic target command for a process operation amount from a process load command, and a correction signal for the basic target command from a deviation between a detected process control amount and a control amount set value. In a process controller including a correction circuit and an acceleration circuit that creates an acceleration signal that corrects the basic target command based on a change in the load command, the acceleration circuit includes a load command, and further, the process control amount. And a deviation between a control amount set value and an input, the process control device is configured to generate an acceleration signal that minimizes the deviation between the process control amount and the control amount set value. 8. The acceleration circuit according to claim 7, wherein the acceleration circuit is configured by a neural network, and the neural network learns the acceleration signal that minimizes the deviation between the process control amount and the control amount set value. The process control device is configured to generate the acceleration signal corresponding to a change in the load command based on the above. 9. The acceleration circuit according to claim 7, wherein the acceleration circuit has an acceleration signal determination rule as a knowledge base, and the deviation between the process control amount and the control amount set value is applied to a membership function to calculate a goodness of fit. Then, the process control device is configured to generate the acceleration signal by fuzzy inference using the goodness of fit and the rule. 10. The circuit according to claim 8, further comprising a circuit that creates a correction signal of the acceleration signal based on a response characteristic of a process control amount when the acceleration signal created by the neural network is applied, A process control device characterized in that the neural network learns an optimum acceleration signal by using a signal obtained by correcting the acceleration signal with the correction signal as a teaching signal. 11. The circuit for calculating a correction signal of the acceleration signal according to claim 10, wherein a rule for determining the correction signal is provided as a knowledge base, and the deviation between the process control amount and the control amount set value and the A process control device characterized in that a response characteristic is applied to a membership function to calculate a goodness of fit, and the modified signal is calculated by fuzzy inference using the calculated goodness of fit and the rule. .. 12. The circuit according to claim 7, further comprising: a circuit that estimates a response characteristic of the process control amount using a prediction model based on a basic target command of the process corrected by the acceleration signal, and the estimated value. And a circuit that corrects the acceleration signal based on the response characteristic.