JPH08303210A - Autonomous adaptive optimizing control system for thermal power plant - Google Patents

Abstract

PURPOSE: To carry out monitoring evaluation for an operation restricting factor by providing a schedule execution controlling means which starts and controls a thermal power plant according to a starring schedule formed by an autonomous managing means. CONSTITUTION: In a schedule execution controlling means 300, a process condition accompanying a start of a thermal power plant 3000 is controlled according to a starting schedule made by an autonomous managing means 200. In the autonomous managing means 200, an employed ratio for a schedule, which is made on the basis of a correction amount obtained from a schedule correction amount computing means 600 based on fuzzy inference, and an employed ratio for a schedule, which is made on the basis of a correction amount obtained from an adaptive knowledge acquiring means 800, are variable. As starring is repeated, the latter ratio is increased, so that adaptive ability against a change in an operating condition and an autonomous ability can be increased. In this way, a starring schedule minimizing a starting time can be automatically formed and can be executed.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a thermal power plant control system and, more particularly, to a start-up schedule suitable for starting up a plant in a short time while satisfying various operational restriction conditions. The present invention relates to a control system for a thermal power plant.

[0002]

2. Description of the Related Art When a thermal power plant is to be started up rapidly, many problems are operating restriction conditions, nonlinear characteristics between input and output, and the presence of time delay. That is, the startup schedule to be created needs to satisfy not only the operation amount such as the fuel flow rate and the water supply flow rate, but also many process state values such as thermal stress generated in equipment and NOx discharged from the boiler to satisfy the operation restriction condition. . In addition, there is a large non-linearity and time delay between the turbine acceleration pattern and load increase pattern, that is, the turbine acceleration rate or load increase rate, which is the operation amount that defines the startup schedule, and the process state value. It was impossible to realize rapid start-up while satisfying the operation limiting condition with high accuracy by merely feedback control. For this reason, in the first conventional method, as described in Japanese Patent Laid-Open No. 63-94009, a dynamic system model of the entire system of the plant is built into the control system, and this is done before the actual plant startup. The start-up schedule was optimized by predicting the start-up characteristics by repeating the start-up simulation using. That is, the non-linearity was modeled as table information or a physical formula and used to create the activation schedule. The second conventional method is disclosed in Japanese Patent Laid-Open No. 3-164
As described in Japanese Patent Publication No. 804, there is a means for evaluating the start-up characteristics based on the actual operation result and calculating the correction amount of the start-up schedule by fuzzy reasoning from this evaluation result. The neural network is learned, and the startup schedule is optimized by correcting the schedule at the previous startup with the value output from the neural network when the driving target is input to the neural network. In other words, the control system does not have a dynamic characteristic model of the entire plant system, and the start schedule is created by utilizing the qualitative knowledge of the expert and the learning ability of the neural network by fuzzy reasoning.

[0003]

However, according to the above-mentioned conventional technique, the following problems occur.

(1) According to the first conventional method, the entire system dynamic characteristic model used for predicting the starting characteristic takes into consideration operating conditions such as the temperature, humidity, water temperature surrounding the plant and the temperature condition of the plant equipment before starting. It is necessary to make it something that has been done, it takes a lot of time to make it, and it has to be made for each target plant. In addition, there are many uncertainties such as heat transfer in the boiler and denitration reaction in the denitration equipment, which limits the accuracy of dynamic characteristics prediction. Further, in the second conventional method, it is difficult to construct a knowledge base for fuzzy reasoning in consideration of fluctuations in operating conditions and uncertainties because of the complexity of the causal relationship, and it takes a lot of time. But the reliability of the knowledge base becomes a problem.

(2) According to the first conventional method, in order to obtain the optimum start schedule within the practical time by using the whole system dynamic characteristic model of the plant, a high processing capacity and a large memory capacity are provided. You need a computer system.

(3) If the operating conditions such as the initial state of the plant before starting and the temperature are different, it is difficult to utilize the starting schedule created in the past, and the optimum solution must be obtained each time. In other words, the control system does not have the ability to adapt to changes in operating conditions, so it is necessary to fully operate the high-performance computer each time it is started. Further, in the second conventional method, as described above, it is difficult to construct the knowledge base for fuzzy inference in consideration of the fluctuation of the operating conditions and the uncertain factors. Therefore, it is the same as the first conventional method. Has no ability to adapt to changes in operating conditions.

The object of the present invention is to (1) be able to monitor and evaluate operational limiting factors without using a complex large-scale dynamic system model of a power plant. In other words, use the minimum necessary dynamic characteristic model to monitor the driving restriction factors, (2)
It is possible to optimize the startup schedule in the process of repeating actual startup without requesting the optimal startup schedule from the first startup. (3) Even if the operating conditions such as the plant initial state and temperature are different, the startup schedule created in the past can be (4) It is to provide a thermal power plant autonomous adaptive optimization control system capable of adaptive utilization, and (4) capable of efficiently adding adaptive capacity with a small amount of calculation in the process of repeating startup.

[0008]

A thermal power plant autonomous adaptive optimization control system of the present invention for achieving the above object is driven by a boiler for generating steam and steam generated by the boiler. A control system for a thermal power plant including a steam turbine, comprising the following means.

(1) Basic schedule creating means for creating a start-up schedule from ignition of the boiler to completion of start-up based on a temperature state before starting the thermal power plant, (2) in a starting process of the thermal power plant Operation restriction factor monitoring means for monitoring predetermined operation restriction factors,
(3) margin value evaluation means for evaluating a margin value for the operation restriction condition of the operation restriction factor obtained by the operation restriction factor monitoring means, (4) based on the margin value obtained by the margin value evaluation means, Schedule modification amount calculation means for calculating the modification amount of the startup schedule by using fuzzy inference, (5) when the past startup schedule and operating conditions of the thermal power plant are given as input values to the neural network, as output values In order to obtain the schedule correction amount calculated by the schedule correction amount calculation means, adaptive knowledge acquisition means for storing the input / output correspondence, (6) the schedule correction amount calculated by the schedule correction amount calculation means Based on the first start-up schedule created based on the above, and the schedule modification amount obtained by the adaptive knowledge acquisition means. And an autonomous coefficient λ that defines an adoption ratio of the first and second startup schedules when the actual startup schedule is created using the created second startup schedule. Is increased as the startup is repeated, thereby increasing the adoption ratio of the second startup schedule, (7) Startup control of the thermal power plant according to the startup schedule created by the autonomous management means. Schedule execution control means.

Further, the schedule modification amount calculation means is based on a knowledge base that stores a plurality of fuzzy rules that define a schedule modification amount of the start schedule corresponding to the size of the margin value obtained by the margin value evaluation means. Alternatively, the schedule modification amount may be calculated by incorporating the fuzzy rule.

Further, in (1) to (7), learning management means for learning the input / output correspondence by determining the internal connection state of the neural network in the adaptive knowledge acquisition means may be added.

A gas turbine and an exhaust heat recovery boiler for generating steam using the exhaust gas of the gas turbine,
Even in the control system of the combined cycle power plant having the steam turbine driven by the steam generated in the exhaust heat recovery boiler, by providing the means (1) to (7) or the learning management means of the present invention, The purpose can be achieved.

The predetermined operation limiting factor may be at least one of stress in the steam header of the superheater outlet of the boiler, stress in the steam turbine, and NOx discharged from the boiler, In the combined cycle power plant, at least one of the stress of the steam header at the superheater outlet of the exhaust heat recovery boiler, the stress of the steam turbine, and the NOx discharged from the exhaust heat recovery boiler may be used.

Further, the start-up schedule in the combined cycle power plant comprises a start-up schedule of the gas turbine and a start-up schedule of the steam turbine, and the start-up schedule of the gas turbine is a predetermined parameter defining the start-up schedule. And the predetermined parameter is a speed increase rate,
It is at least one of a rated speed holding time, an initial load, an initial load holding time, a load increase rate, and a load holding time, and the steam turbine start-up schedule is configured by a predetermined parameter that defines the start-up schedule. The predetermined parameters are the high-pressure turbine bypass valve operating speed,
It may be at least one of a medium pressure turbine bypass valve operating speed, a low pressure turbine bypass valve operating speed, a high pressure adjusting valve operating speed, an intermediate pressure adjusting valve operating speed, and a low pressure adjusting valve operating speed.

The predetermined operation limiting factors are the superheater outlet steam header stress of the exhaust heat recovery boiler, the steam turbine stress, and NOx exhausted from the exhaust heat recovery boiler, and the margin value evaluating means is For the superheater outlet steam header stress of the exhaust heat recovery boiler and the steam turbine stress, a minimum margin value for the operation limit value is obtained for each section obtained by dividing the starting process of the thermal power plant into a plurality of sections, and the exhaust For NOx discharged from the heat recovery boiler, an instantaneous value and a moving average value are obtained to obtain a minimum margin value for the operation limit value for each section obtained by dividing the starting process of the thermal power plant into a plurality of sections. May be

Furthermore, the schedule correction amount calculation means is a steam turbine main plan based on the minimum margin value of the exhaust heat recovery boiler superheater outlet steam header stress and the minimum margin value of the steam turbine stress obtained from the margin value evaluation means. (ST
Fuzzy inference for PS) and gas turbine global adjustment (GTG)
Fuzzy inference for T) and NOx exhausted from the exhaust heat recovery boiler
The fuzzy inference for the gas turbine master plan (GTPS) and the steam turbine local adjustment (STLT) based on the minimum margin value of STST is performed, and the STPS compares the STLT with the wider range of schedule parameters, The GTPS may perform more precise correction as compared with the GTGT, and the GTGT may perform more global correction as compared with the GTPS.

Furthermore, the schedule correction amount calculation means is for the STPS, STLT, GTPS, GT
From the plurality of correction amounts for the same schedule parameter obtained from each fuzzy inference for GT, the priority value determining means selects low values for the speed-up rate, the load change rate, and the valve operation speed, and the speed holding time, A high value may be selected as the parameter related to the load holding time.

Furthermore, the neural network used in the adaptive knowledge acquisition means has a multi-layer structure composed of at least an input layer and an output layer by collecting a plurality of units having a signal conversion function, and has an output unit of the preceding layer unit. The internal connection strength provided in the connection part with the input part of the next layer unit can be set, the number of units corresponding to the startup schedule parameters and operating conditions is arranged in the input layer, and the startup schedule is provided in the output layer. The number of units corresponding to the parameters may be arranged, and the signal input to the input layer may be preset by the learning management means, converted according to the internal connection strength, and output from the output layer.

Furthermore, the learning management means includes a cumulative correction amount calculation means for cumulatively calculating a new correction amount obtained by the schedule correction amount calculation means each time the thermal power plant is repeatedly started, A cumulative correction amount storage means for storing the cumulative correction amount obtained by the cumulative correction amount calculation means, a schedule storage means for storing the start schedule applied to the actual start in the past, and an actual start time in the past Operating condition storing means for storing the operating conditions, the schedule storing means, and the operating condition storing means, respectively, and retrieves the past activation schedule and operating conditions, and inputs them into the neural network in the adaptive knowledge acquisition means. The output value obtained from the output section when input to the section is stored in the cumulative correction amount storage means. And learning means for determining the internal bond strength of the neural network to match the cumulative correction amount that is, may have a.

Further, the learning management means has a learning means for preferentially learning the output with respect to the input having a large deviation from the stored cumulative correction amount in the learning process. May be.

Furthermore, the operating conditions stored in the operating condition storage means are the shutdown period of the thermal power plant, plant equipment temperature, pressure, condenser cooling water temperature, atmospheric temperature, humidity, plant operation restriction conditions. It may be at least one of the above.

Furthermore, the autonomy managing means inputs a starting schedule X ₁ determined by the correction amount obtained from the schedule correction amount calculating means as a first starting schedule, and inputs X ₁ to the neural network together with the operating conditions. The correction amount ΔX obtained from the output unit when
A value obtained by multiplying the first startup schedule X ₁ by the complement (1-λ) of the autonomous coefficient when the startup schedule X ₂ obtained by adding _OUT and X ₁ is used as the second startup schedule. (1-λ) X ₁ and the second startup schedule X
_The starting schedule X _R obtained by adding ₂ to the value λX ₂ obtained by multiplying the autonomous coefficient λ may be adopted as the actual starting schedule and transferred to the schedule execution control means.

Furthermore, the autonomy managing means has an autonomic coefficient generating means for varying the autonomic coefficient λ,
At the time of initial activation, λ = 0, and 0 ≦ λ as the number of activations is increased, corresponding to the number of activations i or the schedule modification amount obtained by the schedule modification amount calculation means.
It is also possible to increase λ within the range of ≦ 1.

Further, the operation limiting factor is at least the expansion difference between the rotor and the casing of the steam turbine, the stress generated in the drum of the exhaust heat recovery boiler, and the SO or CO exhausted from the exhaust heat recovery boiler. It may be.

Furthermore, the thermal power generation plant autonomous adaptive optimization control system having the above configuration may be used as a start control system for a gasification combined cycle power generation plant, a normal pressure and a pressurized fluidized bed boiler power generation plant.

Furthermore, the above-mentioned series of means for creating a start schedule in the thermal power plant autonomous adaptive optimization control system of the present invention may be used as guidance presenting means for creating or modifying a start schedule in a plant operation training simulator. Good.

Furthermore, the thermal power plant autonomous adaptive optimization control system of the present invention may be provided with means for displaying the actual starting schedule.

[0028]

According to the present invention, the first activation schedule created based on the schedule modification amount obtained by the schedule modification amount calculation means and the schedule modification amount obtained by the adaptive knowledge acquisition means. Second
Is used to create an actual startup schedule, and the autonomous coefficient λ that defines the adoption ratio of the first and second startup schedules when creating the actual startup schedule is increased as the startup is repeated. Accordingly, since the autonomous management means for expanding the adoption ratio of the second activation schedule is provided, the adaptive knowledge acquisition means rather than the conclusion from the fuzzy inference as the activation is repeated when determining the actual activation schedule. The degree depending on the conclusion obtained from the neural network can be strengthened, and a suitable start-up schedule corresponding to changes in the operating conditions of the plant can be created.

That is, the adaptive knowledge acquisition means has a neural network for storing the correspondence between the schedule and the operating conditions of the plant and the correction amount obtained by the schedule correction amount determining means every time the plant is started. When creating an actual startup schedule, a first schedule created by the modification amount from the schedule modification amount determining means by fuzzy reasoning and a second schedule created by the modification amount from the adaptive knowledge acquisition means If the latter ratio of the adoption ratio (autonomous coefficient λ) is increased as the start-up is repeated, the actual start-up schedule becomes higher in the second schedule created corresponding to the operating conditions of the plant. As the number of startups increases, it is possible to create a suitable startup schedule that responds to changes in plant operating conditions. And are possible.

The functions of the above-mentioned means (1) to (7), the knowledge base and the learning management means of the present invention will be described below.

(1) In the basic schedule creating means, paying attention to the stop period having a relatively high correlation with the temperature state of the plant, the central power feeding command is selected from a plurality of schedules prepared in advance as table information for each stop period. The corresponding schedule is selected according to the start time commanded by the office, and this is used as the basic schedule. This creates a schedule that, although not necessarily optimal, can safely start the plant.

(2) In the operation restriction factor monitoring means, in the process of starting the plant by the schedule execution control means,
The process state that can be directly measured is monitored by the signal obtained from the measuring instrument, and the process state that is difficult to be directly measured is input to the dynamic characteristic model to estimate the behavior of the operation restriction factor. To do.

(3) The margin value evaluation means evaluates the margin value with respect to the limit value for the operation limit factor obtained by the operation limit factor monitoring means in the entire starting process.

(4) The schedule correction amount calculation means calculates the correction amount of the schedule according to the size of the margin value obtained by the margin value evaluation means. At this time, it is calculated by fuzzy inference using a knowledge base described later.

(5) The adaptive knowledge acquisition means acquires adaptive knowledge by using a neural network. Here, the adaptive knowledge is not only the schedule correction measure expressed by the fuzzy rule, but also the knowledge having the adaptability to deal with the change in the plant operation condition. In order to acquire this knowledge, the neural network stores the correspondence between the schedule and the plant operating conditions and the correction amount obtained by the schedule correction amount determining means each time the plant is started.

(6) In determining the actual activation schedule, the autonomy management means depends on the conclusion from the neural network having the ability to acquire adaptive knowledge rather than the conclusion from the fuzzy inference as the activation is repeated. It works to strengthen the degree. That is, the adoption ratio (autonomous coefficient λ) of the schedule created by the modification amount from the schedule modification amount determination means by fuzzy reasoning and the schedule created by the modification amount from the adaptive knowledge acquisition means is made variable, and the latter is repeated as the activation is repeated. By increasing the ratio of the above, it works to enhance adaptability and autonomy to changes in operating conditions.

(7) The schedule execution control means controls the process states necessary for starting the plant, such as turbine speed-up control, load increase control, steam pressure control, etc., according to the start-up schedule created by the autonomy control means described later. To do.

The knowledge base is an aggregate of a plurality of rules for which a schedule correction policy is determined based on the causal relationship with the margin value obtained by the margin value evaluation means, and each rule corresponds to the size of the margin value. It functions as a piece of knowledge that expresses the expert's knowledge about the method of modifying the schedule into a computer processable expression.

Further, the learning management means causes the adaptive knowledge acquisition means to learn the correspondence. For that purpose, the correspondence between the schedule and the plant operating conditions at each plant start-up and the correction amount obtained by the schedule correction amount determining means is sequentially presented as a learning sample to the neural network for learning. That is, the learning management means functions to determine the internal connection state of the neural network so that the latter is output when the former is input to the neural network. Here, when determining the internal connection state, it works to preferentially modify the internal connection strength corresponding to the learning parameter including the schedule parameter or the schedule parameter in which the deviation of the output value from the target value is large.

[0040]

Embodiments of the present invention will be described below.

FIG. 1 shows the basic configuration of a thermal power plant autonomous adaptive optimization control system 1000 of the present invention. The present system operates upon receipt of a start time T _Ri (i indicates the next start, for example, i = 1 indicates the first start) which is a start command from the central power feeding command station 2000, and schedule execution control means in the system. 300 thermal power plant 30
00 is actually started.

The system 1000 includes a basic schedule creating means 100, a schedule modifying means 150, a schedule execution control means 300, and an operation restriction factor monitoring means 40.
0, margin value evaluation means 500, schedule correction amount calculation means 600, knowledge base 700, correction amount storage means 650,
It comprises adaptive knowledge acquisition means 800, learning management means 900, and autonomy management means 200.

The basic schedule creating means 100 selects a corresponding schedule from a plurality of schedules prepared in advance for each stop period as table information, corresponding to the start time T _Ri instructed by the central power feeding command station,
This is the basic schedule X _0i . Where X
Is a vector quantity consisting of a plurality of parameter values that define the schedule, and details will be described later. This creates a schedule that is not optimal, but safe and reliable for plant startup.

In the schedule correction means 150, the schedule correction amount calculation means 60 using fuzzy inference described later.
A new schedule X _i is obtained by using the correction amount ΔX _i-1 obtained from 0 and stored in the correction amount storage means 650 and the previous activation schedule X _Ri-1 stored in the learning management means 900 described later. create. Here, at the first startup (i =
In 1), ΔX ₀ = 0 and X ₀ = X ₀₁ .

The schedule execution control means 300 controls the process speed associated with the start-up of the turbine speed, load, steam pressure, etc. in the thermal power plant 3000 according to the start-up schedule X _Ri created by the autonomous management means 200 described later. .

In the operation restriction factor monitoring means 400, in the process in which the plant is started by the schedule execution control means 300, for the process status that is difficult to be directly measured, the measurable process status is input to the dynamic characteristic model to restrict the operation. Estimate and monitor the behavior of factors. In addition, the process state where direct measurement is possible is monitored by the signal obtained from the measuring instrument.

The allowance value evaluation means 500 evaluates the allowance value for the operation restriction factor obtained by the operation restriction factor monitoring means 400 in the entire starting process.

In the schedule correction amount calculation means 600,
The schedule correction amount ΔX _i is calculated according to the size of the margin value obtained by the margin value evaluation means 500. At this time,
It is calculated by fuzzy inference using the knowledge base 700 described below.

The knowledge base 700 includes the margin value evaluation means 50.
It is a set of multiple rules that define the schedule modification policy based on the causal relationship with the margin value obtained by 0, and each rule uses a computer to collect the expert's knowledge about the method of modifying the schedule according to the size of the margin value. It functions as a piece of knowledge that can be processed and expressed.

In the correction amount storage means 650, the correction amounts ΔX _{1 to} ΔX _i obtained from the schedule correction amount calculation means 600.
Is stored in association with each past plant startup.

The adaptive knowledge acquisition means 800 acquires adaptive knowledge using a neural network. Here, the adaptive knowledge is not only the schedule correction measure expressed by the fuzzy rule, but also the knowledge having the adaptability to deal with the change in the plant operation condition. In order to acquire this knowledge, the neural network stores the correspondence between the schedule and the plant operating conditions each time the plant is started, and the correction amounts ΔX _{1 to} ΔX _i obtained by the schedule correction amount calculation means 600. Here, as the plant operation conditions, in the embodiment of the present invention, the stop period T _Si (which means the elapsed time between the previous stop time and the command start time T _Ri ) and the temperature T _Ai that most affect the start-up temperature state of the plant. , Humidity H _Ai and water temperature T _Wi are used. In addition, Z _i will be used hereinafter as a general term for these plant operating conditions.

The learning management means 900 causes the adaptive knowledge acquisition means 800 to learn the correspondence relationship. for that reason,
The correspondence relationship between the schedule X _Ri and the plant operating condition Z _{i at} each plant startup and the cumulative value of the correction amount ΔX _i obtained by the schedule correction amount calculation means 600 is sequentially presented as a learning sample and learned. That is, the learning management unit 900 uses X _Rk (k = 1 to i) in the input layer of the neural network used as the adaptive knowledge acquisition unit 800.
−1) and Z _k (k = 1 to i−1) are input, the output value ΔX _OUTk from the output layer is the cumulative correction amount ΔX _Qk (k of ΔX _i
= 1 to i-1), the connection state in the neural network is determined. Cumulative correction amount ΔX
_Qk will be described later in detail.

In determining the actual activation schedule, the autonomy management means 200 functions to strengthen the degree of dependence on the conclusion ΔX _OUTi from the neural network rather than the conclusion ΔX _i from the fuzzy inference as the activation is repeated. To do. That is, the adoption ratio of the schedule created by the modification amount obtained from the schedule modification amount calculation means 600 by fuzzy reasoning and the schedule created by the modification amount obtained from the adaptive knowledge acquisition means 800 is represented by the autonomous coefficient λ, which is variable. By increasing the ratio of the latter as the activation is repeated, it works to increase adaptability and autonomy to changes in operating conditions.

Hereinafter, as a thermal power plant to which the present invention is applied, a combined cycle power plant comprising a gas turbine, an exhaust heat recovery boiler and a steam turbine will be described as an example.

FIG. 2 shows a combined cycle power plant 3000.
3 shows the relationship between the device configuration and the schedule execution control means 300. The schedule execution control means 300 further comprises a gas turbine control system 330 and a steam turbine control system 360. In the gas turbine equipment 3200, the energy generated by combustion by injecting the fuel 22 into the combustor 23 and pressurizing the combustion air 24 by the compressor 20 is converted into mechanical energy by the gas turbine 26, whereby the common shaft is provided. It drives a generator 35 connected to 28 to convert it into electric energy, and part of the electric power serves as driving force for the compressor 20. At the time of startup, the fuel control valve 21 is operated by the fuel control valve opening command 311 from the gas turbine control system 330, and the fuel flow rate is controlled to control the gas flow rate and the steam turbine facility 3300 connected to the common shaft 28 and the steam turbine facility 3300. The machine 35 is accelerated.

Further, the exhaust gas 27 from the gas turbine 26
Is guided to the exhaust heat recovery boiler facility 3100, and the thermal energy of the exhaust gas 27 is recovered. At this time, the exhaust gas 27 causes the fluid in the various heat exchangers arranged in the flue 14 of the exhaust heat recovery boiler facility 3100 to receive heat, evaporate and be superheated.
The exhaust heat recovery boiler equipment 3100 of this embodiment is composed of a steam system having three pressure levels, and the steam generated from each of them is a high pressure steam 341, an intermediate pressure steam 342, and a low pressure steam 34.
It is 3. The high-pressure turbine 31, the intermediate-pressure turbine 32, and the low-pressure turbine 33 are driven in the steam turbine equipment 3300 by the heat energy of these steams, and play a role in power generation by the generator 35 connected to the common shaft 28.

At start-up, the exhaust heat recovery boiler equipment 3100
By controlling the high pressure main steam 341, the intermediate pressure main steam 342, and the low pressure main steam 343, which are generated from each of them, via the high pressure bypass valve 334, the intermediate pressure bypass valve 335, and the low pressure bypass valve 336, the individual pressures are controlled to predetermined values. In addition, by operating the high pressure control valve 331, the medium pressure control valve 332, and the low pressure control valve 333, the output of the high pressure turbine 31, the intermediate pressure turbine 32, and the low pressure turbine 33 is increased. Therefore, the flow rates of the high-pressure bypass steam 344, the medium-pressure bypass steam 345, and the low-pressure bypass steam 346 increase as the opening degrees of the respective bypass valves 334, 335, 336 increase, and the opening degrees of the control valves 331, 332, 333. When is increased, the amount of steam flowing into the steam turbine increases, and the amount decreases. Opening commands 324, 32 to these bypass valves
5,326 and opening degree commands 321,322,3 to the regulator valve
All of 23 are output from the steam turbine control system 360. Further, when the temperature deviation between the medium-pressure superheated steam 1 and the high-pressure turbine exhaust 2 is within a predetermined value during the startup of the plant, the medium-pressure stop valve 337 is opened. Operation signal 32 at this time
7 is also commanded by the steam turbine control system 360.
Condensate 37 from the condenser 34 is the low-pressure feed pump 1
6, by medium pressure water supply pump 17, high pressure water supply pump 18,
The flow rates of the low-pressure feed water 3, the medium-pressure feed water 4, and the high-pressure feed water 5 are controlled so that the water levels of the low-pressure drum 6, the medium-pressure drum 7, and the high-pressure drum 8 are kept within predetermined values.

Here, the operation limiting factors are the stress generated in the rotors of the high-pressure turbine 31 and the intermediate-pressure turbine 32, and the stress generated in the headers at the outlets of the high-pressure superheater 11 and the intermediate-pressure superheater 12. , NOx emissions from the flue outlet 15 of the exhaust heat recovery boiler facility 3100 to the atmosphere. Each stress is
It appears as a result of a dynamic process with a large time delay of heat transfer from the gas turbine exhaust gas 27 to the metal of the heat exchanger, heat transfer from the metal to the internal fluid, and heat transfer from the internal fluid to the metal of interest. In addition, centrifugal force on the rotor,
Mechanical stress corresponding to the vapor pressure is added to the header. Hereinafter, unless otherwise specified, these are collectively referred to as stress. The NOx emission amount from the boiler also largely depends on the NOx emission characteristic of the gas turbine itself and the temperature characteristic of the denitration device 13 installed in the flue. Further, the various operation limiting factors are greatly influenced by the plant stop period T _Si , the temperature T _Ai , the humidity H _Ai , and the water temperature T _Wi which are the operating conditions Z _i . That is, the temperature and humidity affect the exhaust gas temperature of the gas turbine and the exhausted NOx, and as a result, affect the stress characteristics of the boiler and the steam turbine and the amount of NOx exhausted to the atmosphere after passing through the denitration device. Further, the water temperature affects the condensate performance of the steam turbine exhaust, and changes in the plant efficiency due to changes in plant efficiency have an effect. This affects both the stress and NOx characteristics as well as the effect of the stop time. Therefore, in order to accurately manage these operation limiting factors, it is necessary to coordinate and coordinate each control operation. Therefore, as described above, in the thermal power plant autonomous adaptive optimization control system 1000 of the present invention, the operation restriction factor monitoring means 400 and the margin value evaluation means 500 for monitoring and evaluating these dynamic behaviors are provided.

FIG. 3 shows the relationship between the starting process and the starting schedule parameter of the combined cycle power plant to which the present invention is applied.

As shown in the figure, the schedule parameters related to the gas turbine are the speed-up rate DN, the rated speed holding time DT _NL , the initial load L _I , the initial load holding time DT _LI , the first load increase rate DL ₁ , and the load. The holding time DT _HL , the second load increase rate DL ₂ , and the third load increase rate DL ₃ . The gas turbine is started by adjusting the opening of the fuel control valve 21, which is the operating end, with the start schedule defined by these as the control target. In addition, the operating end of the steam turbine is, as already described, the high pressure bypass valve (HPBV) 334, the intermediate pressure bypass valve (IPBV) 335, the low pressure bypass valve (LP
BV) 336, high pressure control valve (HPCV) 331, medium pressure control valve (IPCV) 332, low pressure control valve (LPCV) 33
3, there is an intermediate pressure stop valve (ISHV) 337, and high pressure bypass valve operation speed DA _HBV , intermediate pressure bypass valve operation speed DA _IBV , low pressure bypass valve operation speed DA _LBV , low pressure bypass valve operation waiting time DT as schedule parameters _{It is controlled according to LBV} , high pressure control valve first operation speed DA _HCV1 , high pressure control valve second operation speed DA _HCV2 , and low pressure control valve operation speed DA _LCV . The control targets and operation timings other than these schedule parameters are as illustrated. Here, as steam conditions, T _MS is high-pressure main steam temperature, P _MS is high-pressure main steam pressure, P _CRP is high-pressure turbine exhaust pressure, P _IS is medium-pressure main steam pressure, P _LS is low-pressure main steam pressure, and ΔT is This is the deviation between the medium pressure superheater steam temperature and the high pressure turbine exhaust temperature. Also,
In the figure, the% display of each valve indicates the opening. Next, the detailed configuration of the control system and each means will be described step by step with reference to FIG. 4 and the subsequent figures.

FIG. 4 shows a detailed configuration of a thermal power plant autonomous adaptive optimization control system for a combined cycle power plant which is an embodiment of the present invention.

The present system operates in response to a start time T _Ri (i indicates the next start, for example, i = 1 indicates the first start) which is a start command from the central power feeding command station 2000,
The combined cycle power plant 3000 is actually started by the schedule execution control means 300 in the system. The system 1000 includes a basic schedule creating means 100,
Schedule correction means 150, schedule execution control means 300, operation restriction factor monitoring means 400, margin value evaluation means 500, schedule correction amount calculation means 600, knowledge base 700, correction amount storage means 650, adaptive knowledge acquisition means 800, learning management means. 900 and an autonomy management means 200. Hereinafter, each means will be described step by step.

FIG. 5 shows a basic schedule creating method in the basic schedule creating means 100. Here, when the start time T _Ri is commanded from the central power feeding command station 2000, first, the stop period T _Si , which is the elapsed time from the previous plant stop time T _STPi ,

[0064]

[ _Equation 1] T _Si = T _Ri −T _STPi ( _Equation 1) Next, a table corresponding to the suspension period T _Si is selected from a plurality of schedules prepared in advance for each suspension period as table information, and this is set as the basic schedule X _0i . Therefore, X _0i is a vector quantity composed of a plurality of parameter values that define the schedule.

In the schedule correction means 150, the schedule correction amount calculation means 60 using fuzzy inference described later is used.
A new schedule X _i is obtained by using the correction amount ΔX _i-1 obtained from 0 and stored in the correction amount storage means 650 and the previous activation schedule X _Ri-1 stored in the learning management means 900 described later.

[0066]

[Formula 2] X _i = X _Ri-1 + ΔX _i-1 (Formula 2) Here, at the first startup (i = 1), Δ
X ₀ = 0 and X ₀ = X ₀₁ .

In the schedule execution control means 300, the combined cycle power plant 300 is _operated according to the start schedule X _Ri created by the autonomy management means 200 described later.
The control of the process state associated with startup such as turbine speed, load, steam pressure at 0 is carried out by the above-mentioned method.

In the operation restriction factor monitoring means 400, in the process of starting the plant by the schedule execution control means 300, the dynamic characteristics of the measurable process states of steam turbine stress and boiler stress, which are process states that are difficult to measure directly, are set. Estimate and monitor dynamic behavior by inputting into the model. Further, NOx emitted from the exhaust heat recovery boiler, which is a process state in which direct measurement is possible, is monitored by a signal obtained from a measuring instrument.

FIG. 6 shows a method of monitoring the steam turbine stress in the operation limiting factor monitoring means 400. The stress generated on the rotor surface and rotor bore of the high pressure turbine and the intermediate pressure turbine, which are points to be monitored as the operation limiting factor, are stresses. It was constructed as a dynamic characteristic model for. That is, the steam condition (temperature, pressure) inside the turbine and the heat transfer coefficient on the rotor surface are estimated from the turbine inlet steam condition, speed and load, and the unsteady temperature distribution inside the rotor metal is calculated to determine the rotor surface and bore. Thermal stress was calculated and centrifugal stress was added to obtain total stress. The stress generated in the exhaust heat recovery boiler is also monitored by the same method. That is, the internal heat transfer coefficients of the high pressure superheater outlet header and the medium pressure superheater outlet header, which are the points to be monitored as an operation restriction factor, are estimated from the steam conditions (temperature, pressure) and the flow rate, and the same as in the case of the steam turbine. To find the unsteady temperature distribution inside the metal,
The thermal stress of the inner and outer surfaces of the header metal was calculated, and the stress due to steam pressure was added to obtain the total stress.

Next, the margin value evaluation means 500 for evaluating the margin value with respect to the limit value of the operation restriction factor obtained by the operation restriction factor monitoring means 400 will be described.

FIG. 7 shows a margin value evaluation method for the limiting value of the steam turbine stress. First, the time zone until a predetermined time elapses after activation completion from the turbine start-up (t ₁ ~t ₆₎
Is divided into a plurality of sections (in this example, the case of 5 divisions is shown),
The minimum stress margin value m (i) in the i-th section is calculated. Here, the stress margin value m is defined by the following equation, where the limit value is S _L and the value obtained by the operation limit factor monitoring means 400 is S.

[0072]

Equation 3] m = S _L -S ... (Equation 3) Note that, the stress value is positive when the tensile stress, if negative means compressive stress. As described above, the stress points are actually the rotor surface and the bore of the high-pressure turbine and the rotor surface and the bore of the medium-pressure turbine. Therefore, there are four minimum stress margin values to be obtained for each section. These are described below.

M _HS (i): Minimum stress margin value of high pressure turbine rotor surface in section i m _HB (i): Minimum stress margin value of high pressure turbine rotor bore in section i m _IS (i): Medium pressure turbine rotor in section i Surface minimum stress margin value m _IB (i): Medium pressure turbine rotor bore minimum stress margin value in section i The exhaust heat boiler stress margin value evaluation method is basically the same as the turbine stress margin value evaluation method. However, two points to be noted are the header outer surface at the high pressure superheater outlet and the medium pressure superheater outlet. This is because the stress generated on the outer surface of the header is larger than that on the inner surface. Therefore, there are two minimum stress allowances to be obtained for each section, and they are as follows.

_{M HHD} (i): Minimum stress margin value of outer surface of high pressure superheater header in section i m _IHD (i): Minimum stress margin value of outer surface of medium pressure superheater header in section i FIG. A margin value evaluation method is shown. This method is also similar to the case of turbine stress.
The time zone (t _{1 to} t ₇ ) from the start of the gas turbine start to the elapse of a predetermined time after the completion of the plant start is divided into a plurality of sections (in this example, the case of 6 divisions is shown), and the minimum emission NOx in the i-th section. Instantaneous value margin value m _PS (i) and minimum emission NOx
The average value margin value m _PA (i) is calculated. Here, the discharge NOx instantaneous value margin value m _PS and the discharge NOx average value margin value m _PA are defined by the following equations, where the respective limit values are P _SL and P _AL and the measured values are P _S and P _A.

[0075]

[Equation 4] m _PS = P _SL −P _S (Equation 4)

[0076]

[Equation 5] m _PA = P _AL −P _A ( _Equation 5) FIG. 9 shows the evaluation results obtained from the margin value evaluation means 500 m _HS , m _HB , m _IS , m _IB , m _HHD , and m _IHD. , M _PS , m _PA
In response, the overall outline of the schedule correction amount calculation method in the schedule correction amount calculation means 600 will be shown. In the schedule correction amount calculation means 600, the margin value evaluation means 50
The schedule correction amount ΔX _i is calculated according to the size of the margin value obtained at 0. The schedule correction amount calculating means 600 further includes a steam turbine stress adjusting means 610, a boiler stress adjusting means 620, an exhaust NOx adjusting means 630,
It comprises a priority value determining means 640. Each adjusting means each have a low value selection means 611, 621, 631, low value selection means 611 margin value _{m HS, m HB, m IS} , m
_{From the IB} , the low value selecting means 621 determines the margin values m _HHD and m _IHD.
From among the above, the low value selecting means 631 selects the minimum margin values from the margin values m _PS and m _PA , and sets these as m _T , m _B and m _P , respectively. That is, m _T , m _B , and m _P are defined by the following equations.

[0077]

M _T = Min (−m _HS , m _HB , −m _IS , m _IB ) (Equation 6)

[0078]

[ _Equation 7] m _B = Min (m _HHD , m _IHD ) ... ( _Equation 7)

[0079]

[Equation 8] m _P = Min (m _PS , m _PA ) ... (Equation 8) Next, fuzzy inference 612, 6 using m _T , m _B , m _P
Schedule correction amounts ΔX _T (i), ΔX _B (i), and ΔX _P (i) are calculated from 13, 622, 623, 632, and 633. Here, the subscripts T, B, and P indicate that the correction amounts are determined according to the steam turbine stress, the boiler stress, and the margin value of exhausted NOx, respectively.

FIG. 10 shows the membership function used in the fuzzy rule at this time. The membership function shown in FIG. 10A is a steam turbine main plan (ST) for adjusting the stress by modifying the schedule of the steam system from the turbine stress margin value m _T and the boiler stress margin value m _B.
It defines a correction coefficient k _S for PS) and a schedule correction coefficient k _G for gas turbine global adjustment (GTGT) for adjusting stress by correcting the schedule of the gas turbine system. FIG.
The membership function shown in (b) is the emission NOx allowance m
Main plan of gas turbine for adjusting NOx emission by modifying the schedule of gas turbine system from _P
The correction coefficient k _G for (GTPS) and the schedule correction coefficient k _S for steam turbine local adjustment (STLT) for adjusting exhaust NOx by correcting the schedule of the steam system are defined. More specifically, the upper part of FIG. 10A shows the membership function used in the conditional part of the fuzzy rule for stress evaluation, which consists of four functions (NS, ZO, PS, PB), and Indicates the membership function for determining the schedule correction coefficient used in the conclusion part, and the five functions (NB, NS, Z
O, PS, PB). Here, the meaning of each membership function is NB: Negative Big, NS: Negati
ve Small, ZO: Zero, PS: Positive Small, PB:
Positive Big. The upper part of FIG. 10 (b) shows the membership function used in the conditional part of the fuzzy rule for the emission NOx evaluation, and four functions (NS, ZO, PS, PB) are shown.
The following shows the membership function for determining the schedule correction coefficient used in the conclusion section, and the five functions (NB, N
S, ZO, PS, PB). Here, the meaning of each membership function is the same as in the case of FIG.

Next, a fuzzy rule using a membership function will be described.

FIG. 11 shows a turbine stress adjustment rule used in the steam turbine main plan STPS. Here, as the four correction schedule parameters, the high pressure bypass valve operating speed DA _HBV , the medium pressure bypass valve operating speed DA _IBV , the high pressure adjusting valve first operating speed DA _HCV1 , and the high pressure adjusting valve second operating speed DA _HCV2 are shown in FIG. The second of the five sections of interest shown in
A fuzzy rule for correction in relation to the turbine stress margin value m _T in sections 3, 4, and 5 is shown. That is, the second
And a rule table defining the correction amounts of DA _HCV1 and DA _{HCV2 in} the relationship between the margin values m _T (2) and m _T (3) in the third section, and the margin value m _T (in the third and fourth sections) 3) and m _T (4)
The rule table that defines the correction amounts of DA _HCV1 and DA _{HCV2 based on} the relationship of the above, and the margin value m _T (4) in the fourth and fifth sections
And m _T (5), DA _HBV , DA _IBV , DA _HCV1 and D
A _HCV2 consists of a rule table defining the modification amount. However, the blank portion of the rule table means that the causal relationship between each stress allowance value and the schedule parameter is small or almost nonexistent.

FIG. 12 shows the gas turbine wide area adjustment (GTGT).
Shows the turbine stress adjustment rule used in. Here, the six acceleration schedule parameters DN,
Turbine stress in the five target sections shown in FIG. 7 showing the initial load holding time DT _LI , the first load rising rate DL ₁ , the load holding time DT _HL , the second load rising rate DL ₂ , and the third load rising rate DL ₃ . A fuzzy rule for correction in relation to the margin value m _T is shown. That is, the margin value m in the first and second sections
A rule table defining the correction amounts of DN, DT _LI, and DL ₁ in the relationship between _T (1) and m _T (2), and margin values m _T (2) and m _T (3 in the second and third sections ) DT _LI , DL ₁ ,
A rule table defining the correction amounts of DT _HL and DL ₂ , and margin values m _T (3) and m _T (4) in the third and fourth sections
And the rule table defining the correction amounts of DL ₁ , DT _HL , DL ₂ and DL ₃ in the relation of DT _HL and the margin values m _T (4) and m _T (5) in the fourth and fifth sections. , DL ₂ and DL
_It consists of a rule table that defines the correction amount of ₃ . Also in this rule table, the blank part means that the causal relationship between each stress allowance value and the schedule parameter is small or almost nonexistent.

As described above, the fuzzy rules shown in FIGS. 11 and 12 are for STPS and GTG based on the boiler stress margin value.
It is also used for T.

FIG. 13 shows the emission NOx adjustment rule used in the gas turbine main plan GTPS. Here, as the eight correction schedule parameters, the speed increase rate DN, the rated speed holding time DT _NL , the initial load L _I , and the initial load holding time D
T _LI, first load increase rate DL _1, the load retention time DT _HL, second load increase rate DL _2, third load increase rate DL _3, NOx emission allowance m _P in six interest sections shown in FIG. 8 A fuzzy rule to correct in relation to is shown.

That is, the margin value m in the first and second sections
A rule table defining the correction amounts of DN, DT _NL , L _I, and DT _{LI based on} the relationship between _P (1) and m _P (2), and the margin values m _P (2) and m in the second and third sections _{Because of P} (3), DT _NL ,
A rule table defining correction amounts of L _I , DT _LI, and DL ₁ and margin values m _P (3) and m in the third and fourth sections
_{Due to} the relationship of _P (4), DT _LI , DL ₁ , DT _HL , DL ₂ and DL
D is the relationship between the rule table defining the correction amount of ₃ and the margin values m _P (4) and m _P (5) in the fourth and fifth sections.
It is composed of a rule table defining the correction amounts of L ₁ , DT _HL , DL ₂ and DL ₃ . In this rule table, the blank part is
This means that the causal relationship between each stress margin value and the schedule parameter is small or almost nonexistent.

FIG. 14 shows the steam turbine local adjustment (STL
The emission NOx adjustment rule used in T) is shown below. here,
High pressure bypass valve operating speed DA _HBV and medium pressure bypass valve operating speed DA as three correction schedule parameters.
A fuzzy rule for correcting the _IBV and the first operation speed DA _{HCV1 of} the high-pressure control valve in relation to the exhaust NOx margin value m _{P in the first, second,} and third sections among the six sections of interest shown in FIG. 8 is shown. . That is, the margin value m in the first and second sections
DA _HBV , DA _IBV and DA _{HCV1 due to} the relationship between _P (1) and m _P (2)
DA _HBV , which is a rule table defining the correction amount of, and the relationship between the margin values m _P (2) and m _P (3) in the second and third sections.
It consists of a rule table that defines the correction amount of DA _IBV and DA _HCV1 .

Compared to STLT, the above-mentioned STPS targets a wider range of schedule parameters, GTPS performs more precise correction than GTGT, and GTGT is more comprehensive than GTPS. It works to make various corrections.

FIG. 15 shows a schedule modification amount calculation method by fuzzy inference. This figure shows calculation of the steam system schedule correction coefficient k _S as an example. In this example, for a certain schedule parameter, the membership functions and membership values as conclusions from the four rules are (NS, 0.6), (ZO, 0.8),
The case where (PS, 0.4) and (PB, 0.2) are obtained is shown. The comprehensive evaluation value is defined by the weight W (i) of the trapezoidal portion determined by each membership value and the center of gravity k _SG of the correction coefficient k _S (i). That is, W (1) = 0.168, W (2) = 0.09
6, W (3) = 0.096, W (4) = 0.09, and k
_S (1) =-0.2, k _S (2) = 0, k _S (3) = 0.15, k _S
Since (4) = 0.35, k _SG is calculated as follows.

[0090]

[Equation 9]

Therefore, the modification amount k _SG for this activation schedule parameter is 0.0273. Using this result, the schedule correction amount ΔX _Ti is calculated according to the following equation.

[0092]

[Expression 10] ΔX _T = k _SG (X _Max −X _Min ) ( _Expression 10) Here, X _Max and X _Min represent upper and lower limit values of X.

According to the method described above, the STP
A plurality of correction amounts are calculated for the same schedule parameter by fuzzy inference of S, GTGT, STPS, STLT. Therefore, in order to surely keep all the operation limiting factors within the limit values, the priority value determining means 64 shown in FIG.
At 0, the final correction amount ΔX (i) is determined. That is, the rate of speed increase,
The low value on the safe side is selected by the low value selection means LVG in the figure for the parameters relating to the load change rate and the valve operation speed.
The high value selecting means HVG selects a high value on the safe side for the parameters relating to the speed holding time and the load holding time.

In the correction amount storage means 650, the correction amounts ΔX _{1 to} ΔX _i obtained from the schedule correction amount calculation means 600.
Is stored in association with each past plant startup.

In the learning management means 900, the schedule X _Rk (k = 1 to i-1) for each plant start and the operation condition Z _k (k = _k ) for the adaptive knowledge acquisition means 800 composed of a neural network are used. 1 to i-1) and the cumulative correction amount ΔX which is the cumulative value of the schedule correction amount ΔX _i.
Correspondences with _Qk (k = 1 to i-1) are sequentially presented as learning samples and learned. However, in order to prevent the learning samples from increasing endlessly as the plant accumulates start-up records, old samples are excluded and only r samples (for example, r = 15) from the newest one are stored for learning. It is a matter of course to set it, and the following description will also adopt this measure unless otherwise noted.

FIG. 16 shows a learning sample presented to the neural network. If the startup is completed up to the nth time, the cumulative correction amount ΔX _Qi as teacher data
Is a value defined by the following equation.

[0097]

[Equation 11]

That is, ΔX _Qi is calculated by the cumulative correction amount calculation means 910 as a new correction amount obtained each time the activation is repeated.
It is obtained by performing cumulative calculation according to equation 11. The cumulative correction amount ΔX _Qi is stored in the cumulative correction amount storage means 920.
Stored in.

The learning management means 900 uses X _Rk (k = 1 to i−1) and Z _k (k = _k ) in the input layer of the neural network.
1 to i-1), the output value ΔX from the output layer
_OUTk (k = 1 to i−1) is the cumulative correction amount ΔX _Qk (k = 1 to i
It functions to determine the connection state in the neural network so as to match -1).

Next, the neural network and its learning method will be described.

This neural network has an input layer,
It is composed of a three-layer structure consisting of an intermediate layer and an output layer. Regarding this basic structure, the MIT Press and Neurocomputing Foundations of
Research, 1988, pages 318-362 (The M
IT Press, Neurocomputing Foundations of research,
1988, pp318-362). Further, in this paper, when a certain input signal pattern is given to the input layer, the output layer and the output layer are matched with the desired signal pattern, that is, the teacher data, so that the intermediate layer and the output layer are adjusted according to the error between them. A learning algorithm (referred to as backpropagation) for correcting the connection strength of the input section to each unit of is shown. Also in the embodiment of the present invention, the structure of the neural network and the learning algorithm itself basically use those shown in the paper.

FIG. 17 shows the relationship between the structure of the neural network and the input / output variables. Input variables are the activation schedule parameter X _IN and the operating condition Z. X _IN is a learning sample X _Rk (k = 1) read from the schedule storage means 950 that stores past activation schedules during learning.
~ I-1), the schedule parameter X _i is input when the activation schedule is created. Further, Z is an operation condition storage means 94 that stores past operation conditions at the time of learning.
Learning sample Z _k (k = 1 to i-1) read from 0
However, Z _i is input when the activation schedule is created. When X _i and Z _i are input to the neural network, it is determined by the internal connection strength determined by learning Δ
X _OUTi is output as an output variable. When learning, X _Rk
And Z _k are input, ΔX _OUTk (k = 1 to i−1) is output. ΔX _OUTk (k = 1 to i−1) is ΔX _Qk (k = 1
To i-1), learning is performed by modifying the internal connection strength of the neural network.

In the embodiment relating to this learning, as shown by the following equation, the learning sample is presented to the neural network until the learning error E _MAX becomes smaller than the target value E _R (until it converges), and the connection strength is corrected.

[0104]

[ _Equation 12] E _MAX = M _AX (E _MAXk ) <E _R (Equation 12) Here, E _MAXk is defined by the following equation.

[0105]

[ _Equation 13] E _MAXk = M _AX | ΔX _Qk −ΔX _OUTk | / | ΔX _Qk | ( _Equation 13) That is, E _MAXk is the error corresponding to the scheduling parameter that maximizes the learning error at the k-th learning sample. Yes,
E _MAX indicates the maximum value among E _MAXk (k = 1 to i−1).

Further, in the present invention, the target value E _R is obtained by this learning.
In order to speed up the convergence to, the upper N samples (referred to as learning target samples) with a larger E _MAXk value are given priority for learning. That is, the connection strength is corrected only for the upper N samples extracted at each learning time. Furthermore, in the present invention, in order to speed up the convergence, the N extracted by this method is
As for the schedule parameter for each sample, the upper M parameters (referred to as learning target parameters) are preferentially learned from the one having the largest error.
That is, the connection strength is corrected only for the upper M pieces extracted for each learning. In addition, in the present learning method, the predetermined error E
_S (Learning exclusion parameter selection standard error. However, E _S
Parameters that are less than or equal to <E _R ) are excluded from the correction target of the connection strength to shorten the overall convergence time. The number N of samples to be learned, the number M of parameters, and the learning exclusion parameter selection reference error E _S do not have to be fixed, but can be varied according to the progress of learning, for example, the average error or the maximum error. , More efficient learning is possible.

Next, the autonomy management means 200 will be specifically described. When determining the actual activation schedule X _Ri , the autonomy management means 200 functions to strengthen the degree of dependence on the conclusion ΔX _OUTi from the neural network rather than the conclusion ΔX _i from the fuzzy inference as the activation is repeated. . That is, the autonomous coefficient λ, which is the adoption ratio of the schedule created by the modification amount ΔX _i obtained from the schedule modification amount calculation means 600 by fuzzy inference and the schedule created by the modification amount ΔX _OUTi obtained from the adaptive knowledge acquisition means 800, It is variable and increases the ratio of the latter as the activation is repeated, thereby increasing adaptability and autonomy to changes in operating conditions.
This autonomous coefficient λ is determined by the autonomous coefficient generating means 210. Therefore, the activation schedule X _Ri defined by the following equation is adopted for the actual activation.

[0108]

X _Ri = λ (X _i + ΔX _{OUT i} ) + (1−λ) X _i ( _Equation 14) Here, assuming that 0 ≦ λ ≦ 1, the autonomous coefficient λ is X when λ = 0. _{When Ri} = X _i and λ = 1,

[0109]

[ _Expression 15] X _Ri = X _i + ΔX _OUTi ( _Expression 15)

That is, the former is an activation schedule in which only the conclusions from the fuzzy inference are adopted and the output from the neural network is not used. The latter is a startup schedule in which the basic schedule is corrected by the output from the neural network, and as described above, the adaptability to changes in operating conditions is fully utilized. Therefore, as the value of λ increases, the control system becomes more autonomous. However, if the autonomic coefficient λ is increased before the neural network gains sufficient start-up experience (before providing sufficient adaptability by learning), the optimumness and high reliability of the start-up schedule created cannot be expected. Therefore, how to select this autonomous coefficient λ is important. The following method is applied to the autonomy management means 200 in the present embodiment.

FIG. 18 shows an autonomous management system for determining the autonomous coefficient λ in the autonomous management means 200. In the present embodiment, as shown in the figure, 3
The method can be applied. The first method is the leap-growth type autonomy management method shown in (a), in which λ = 0 is set until the number of activations i reaches a predetermined value i _A , and the neural network continuously learns to improve adaptability. Aside, i
By setting λ = 1 when ≧ i _A , the autonomy is increased at once. The second method is the linear continuous growth type autonomy management method shown in (b), where λ = 0 at i <i _B , the neural network has a certain degree of adaptive ability enhanced by learning, and i _B ≦ i ≦ In i _C , the autonomy is gradually increased at a constant speed while increasing the adaptive ability, and λ = 1 when i ≧ i _C. The third method is a non-linear continuous growth type autonomy management method shown in (c), in which λ is a function of the correction amount ΔX _i obtained from the schedule correction amount calculation means 600 by fuzzy inference. This method uses the correction amount ΔX _i
It is a method of increasing λ according to the decrease of. The function that defines λ is

[0112]

[Expression 16] λ = 1−‖ΔX _i ‖ / ‖ΔX ₁ ‖ (Equation 16) Alternatively, the same function can be achieved by the following equation.

[0113]

[Formula 17] λ = 1-M _AX (| ΔX _i | / (X _MAX −X _MIN )) / M _AX (| ΔX ₁ | / (X _MAX −X _MIN )) (Formula 17) FIG. 7 shows the effect of improving the optimal value convergence by the autonomous adaptive optimization control system of the invention. (A) is a case without autonomous control, that is, a case where the activation schedule is optimized only by fuzzy inference, and (b) is a case with autonomous adaptive control, that is, by combining fuzzy inference and neural network. This is a case where the adaptability is provided to optimize the startup schedule. All of them show how the initial schedule converges toward the optimum value as the plant is repeatedly started. However, in the case of (a), there is a non-convergence region because there is no adaptability to changes in operating conditions as described above, and convergence beyond the convergence limit cannot be expected. on the other hand,
In the case of (b), since the neural network has the adaptive ability of acquiring the schedule correction method including the causality of the operating conditions and the activation schedule, the neural network can converge to an optimum value without limit. This is evaluated from another angle as follows.

FIG. 20 shows the effect of improving the starting characteristics of the plant by the autonomous adaptive optimization control system of the present invention, which is shown in comparison with the case where the autonomous adaptive optimization control is not used. As described above, when the autonomous adaptive optimization control is not used, the non-convergence region exists, and thus the created activation schedule has a range. Therefore, the stress and the exhausted NOx, which are operation limiting factors, also have a range. Therefore, in order to satisfy all the limiting conditions, it is necessary for the created schedule to have a sufficient margin in advance, and the start-up time becomes longer accordingly. On the other hand, in the case of the autonomous adaptive optimization control, it is possible to flexibly adapt to changes in operating conditions as described above, so that an optimal schedule can always be created. Therefore, it is possible to start in the shortest necessary time by the start schedule that can satisfy the operation restriction condition with high accuracy.

The present invention can also be applied to other plants, for example, a normal power plant including a boiler, a steam turbine and a generator, a coal gasification power plant, an atmospheric pressure or a pressurized fluidized bed boiler power plant. Of course. Also, as fuel used, coal, oil, LNG
It is also clear that the above is not limited.

Further, in the embodiment of the present invention, the stress of the steam turbine and the exhaust heat recovery boiler header and the exhausted NOx are treated as the operation limiting factors, but depending on the characteristics of the applied plant,
It is also possible to adopt a method that takes into consideration other factors, for example, the difference in expansion between the rotor of the steam turbine and the casing, the stress of other parts such as the drum of the exhaust heat recovery boiler, and the exhaust SOx and CO. Further, it is clear that even if the stress is not necessarily estimated as a dynamic characteristic model, indirect limit value management such as the rate of change or the range of change of the steam temperature or the metal temperature can be performed without changing the essence of the present invention. Is.

Further, in the embodiment of the present invention, the plant stop period, temperature, humidity, and water temperature are adopted as the operating conditions, but the plant start characteristics such as atmospheric pressure, wind speed, solar radiation, equipment temperature, and pressure are also included. It is also possible to add various state quantities to implement the essence of the present invention without changing the basic principle. Further, when the plant is operated while changing the operation restriction condition at each start, the operation restriction condition can be handled as a part of the operation condition without changing the basic principle of the present invention.

Further, in the embodiment of the present invention, the starting schedule parameters to be modified are eight gas turbine-related and four steam turbine-related, but the invention is not necessarily limited to these, and FIG. It is obvious that the present invention can be implemented without changing the basic principle as long as it is a parameter that defines the plant starting pattern such as the other parameters shown, the opening / closing timing condition of the regulator valve, the set value for steam pressure control, and the like. In the embodiment of the present invention, the start start time commanded by the central power feeding command station is exactly followed,
In addition, it is possible to create a start-up schedule of the shortest time, but instead of the start-up completion time instructed by the central power supply command center, the reference time is shifted even if it is gas turbine ignition time, load combination time, target load arrival time, etc. It is obvious that the present invention can be implemented without changing the principle of the present invention.

[0119]

The first effect of the present invention is a start-up control system for a thermal power plant, which has a plurality of operation restriction conditions such as device life standards and emission NOx, which are not possible with conventional methods, and device protection standards and environmental regulation values. It is possible to automatically create and execute a startup schedule that minimizes the startup time while satisfying both. As a result, the burden on the operator is significantly reduced, and the energy loss accompanying the shortening of the start-up time can be reduced, so that the plant operation cost can be significantly reduced.

A second effect of the present invention is a start control system for a thermal power plant, which simultaneously satisfies a plurality of operation restriction conditions such as equipment life standards and emission NOx, equipment protection standards, environmental regulation values, etc. It is possible to complete the startup at the time specified by. This enables stable and accurate power supply to the power system, which requires frequent start and stop of the plant due to fluctuations in power demand.

A third effect of the present invention is a start control system for a thermal power plant, in which the operation limit value is changed or the start command time is changed from the central power supply command station.
It is possible to recreate the boot schedule and run it. This enables flexible and safe plant operation and power system operation.

A fourth effect of the present invention is a start control system for a thermal power plant, which has adaptability to changes in operating conditions. Therefore, a large-scale complex dynamic characteristic model for predicting start characteristics of the plant is constructed. , It does not have to be built into the startup control system. As a result, the system design and construction time can be significantly reduced, the computer load can be reduced, and the startup schedule can be created faster.

A fifth advantage of the present invention is that a start control system for a thermal power plant realizes a start schedule optimizing function having adaptability to fluctuations in operating conditions by utilizing the features of fuzzy inference and neural networks. It is possible to simplify the knowledge base used in fuzzy inference. This makes it possible to significantly reduce the time required to build the knowledge base.

The sixth effect of the present invention is that the starting control system of the thermal power plant can efficiently perform learning management in the neural network. This allows
It is possible to significantly reduce the learning time for acquiring adaptive knowledge.

[Brief description of drawings]

FIG. 1 is a block diagram of a thermal power plant autonomous adaptive optimization control system of the present invention.

FIG. 2 is a block diagram showing the relationship between the equipment configuration of the combined cycle power plant to which the present invention is applied and the startup schedule execution control.

FIG. 3 is a system diagram showing a relationship between a start process and a start schedule parameter of a combined cycle power plant to which the present invention is applied.

FIG. 4 is a block diagram of a thermal power plant autonomous adaptive optimization control system for a combined cycle power plant that is an embodiment of the present invention.

FIG. 5 is an explanatory diagram showing a basic schedule creation method in the embodiment of the present invention.

FIG. 6 is a block diagram showing a method of monitoring steam turbine stress as an operation limiting factor of the combined cycle power plant according to the embodiment of the present invention.

FIG. 7 is a characteristic diagram showing a steam turbine stress margin value evaluation method as an operation limiting factor of the combined cycle power plant according to the embodiment of the present invention.

FIG. 8 is a characteristic diagram showing an exhaust NOx margin value evaluation method as an operation limiting factor of the combined cycle power plant according to the embodiment of the present invention.

FIG. 9 is a flowchart showing an overall outline of a schedule correction amount calculation method in consideration of all operation limiting factors in the embodiment of the present invention.

FIG. 10 is an explanatory diagram showing a membership function used in a fuzzy rule in the embodiment of the present invention.

FIG. 11 is an explanatory diagram showing a steam turbine stress adjustment rule used for a steam turbine main plan (STPS) by the schedule correction amount calculation means in the embodiment of the present invention.

FIG. 12 is an explanatory diagram showing a steam turbine stress adjustment rule used for gas turbine wide area adjustment (GTGT) by the schedule correction amount calculation means in the embodiment of the present invention.

FIG. 13 is an explanatory diagram showing an emission NOx adjustment rule used for a gas turbine main plan (GTPS) by the schedule correction amount calculation means in the embodiment of the present invention.

FIG. 14 is an explanatory diagram showing an emission NOx adjustment rule used for steam turbine local adjustment (STLT) by the schedule correction amount calculation means in the embodiment of the present invention.

FIG. 15 is an explanatory diagram showing a schedule correction amount calculation method by fuzzy inference according to the embodiment of the present invention.

FIG. 16 is an explanatory diagram showing a learning sample for a neural network according to the embodiment of the present invention.

FIG. 17 is an explanatory diagram showing the structure of a neural network and the definition of input / output variables in the embodiment of the present invention.

FIG. 18 is an explanatory diagram showing an autonomy management system according to the embodiment of this invention.

FIG. 19 is an explanatory diagram showing the effect of improving the optimal value convergence by the autonomous adaptive optimization control system of the present invention.

FIG. 20 is an explanatory diagram showing the effect of improving the starting characteristics of the combined cycle power plant by the autonomous adaptive optimization control system of the present invention.

[Explanation of symbols]

100 ... Basic schedule creation means, 150 ... Schedule correction means, 200 ... Autonomy management means, 300 ... Schedule execution control means, 400 ... Operation restriction factor monitoring means, 500 ... Margin value evaluation means, 600 ... Schedule correction amount calculation means, 650 ... Correction amount storage means, 700 ... Knowledge base, 800 ... Adaptive knowledge acquisition means, 900 ... Learning management means, 1000 ... Autonomous adaptive optimization control system for thermal power plant, 2000 ... Central power supply command station.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Takao Akiyama 3-1-1 1-1 Saiwaicho, Hitachi-shi, Ibaraki Hitachi Ltd. Hitachi factory

Claims (11)

[Claims] 1. A control system for a thermal power plant having a boiler for generating steam and a steam turbine driven by the steam generated by the boiler, wherein a start-up schedule from ignition to completion of start-up of the boiler is set. , A basic schedule creating means created based on a temperature state before starting of the thermal power plant, an operating restriction factor monitoring means for monitoring a predetermined operating restriction factor in the startup process of the thermal power plant, and the operating restriction factor monitoring A margin value evaluating means for evaluating a margin value for the operation limiting condition of the operation limiting factor obtained by the means, and a schedule correction amount of the startup schedule corresponding to the size of the margin value obtained by the margin value evaluating means. And a knowledge base that stores a plurality of fuzzy rules that are stored in the knowledge base. When a schedule correction amount calculation means for calculating the schedule correction amount by fuzzy inference using a fuzzy rule, and the past start schedule and operating conditions of the thermal power plant are given as input values to the neural network, the output value And determining the internal connection state of the neural network in the adaptive knowledge acquisition means and the adaptive knowledge acquisition means so as to obtain the schedule correction amount obtained by the schedule correction amount calculation means. Learning management means for learning the input-output correspondence relationship by means of, a first start-up schedule created based on the schedule correction amount obtained by the schedule correction amount calculation means, and a schedule obtained by the adaptive knowledge acquisition means. The second created based on the correction amount A dynamic schedule is used to create an actual startup schedule, and the autonomous coefficient λ that defines the adoption ratio of the first and second startup schedules when creating the actual startup schedule increases as the startup is repeated. The autonomy management means for expanding the adoption ratio of the second start-up schedule, and the schedule execution control means for start-up control of the thermal power plant according to the start-up schedule created by the autonomy management means. Autonomous adaptive optimization control system for thermal power plants. 2. A gas turbine, and an exhaust heat recovery boiler for generating steam using exhaust gas of the gas turbine,
In a control system of a thermal power plant having a steam turbine driven by steam generated in the exhaust heat recovery boiler, a startup schedule from ignition of the gas turbine to completion of startup, a temperature state before startup of the thermal power plant. A basic schedule creating means created based on, and an operation restriction factor monitoring means for monitoring a predetermined operation restriction factor in the starting process of the thermal power plant,
A margin value evaluation means for evaluating a margin value for the operation restriction condition of the operation restriction factor obtained by the operation restriction factor monitoring means, and a start-up schedule corresponding to the size of the margin value obtained by the margin value evaluation means. A knowledge base for storing a plurality of fuzzy rules defining a schedule modification amount, a schedule modification amount calculation means for calculating the schedule modification amount by fuzzy inference using the fuzzy rules stored in the knowledge base, and the thermal power plant. The input / output correspondence is stored so that when the past start schedule and the operating conditions of the above are given as input values to the neural network, the schedule correction amount obtained by the schedule correction amount calculation means is obtained as the output value. Adaptive knowledge acquisition means and new information in the adaptive knowledge acquisition means Learning management means for learning the input / output correspondence by determining the internal connection state of the network, and a first startup schedule created based on the schedule correction amount obtained by the schedule correction amount calculation means, An actual startup schedule is created using the second startup schedule created based on the schedule modification amount obtained by the adaptive knowledge acquisition means, and the first and the first steps for creating the actual startup schedule are performed. Created by the autonomy management means for increasing the adoption ratio of the second startup schedule by increasing the autonomous coefficient λ that defines the adoption ratio of the second startup schedule as the startup is repeated. Schedule for controlling startup of the thermal power plant according to the startup schedule An autonomous adaptive optimization control system for a thermal power plant, characterized by comprising: 3. A neural network used in the adaptive knowledge acquisition means according to claim 1 or 2, wherein a plurality of units having a signal conversion function are collected to form a multilayer structure including at least an input layer and an output layer. It is possible to set the internal connection strength provided at the connection between the output section of the layer unit and the input section of the next layer unit, and the input layer is arranged with a number of units corresponding to the startup schedule parameters and operating conditions, and the output The number of units corresponding to the start-up schedule parameter is arranged in the layer, and the signal input to the input layer is preset by the learning management means, converted according to the internal connection strength, and output from the output layer Autonomous plant Adaptive optimization control system. 4. The cumulative correction according to claim 1, wherein the learning management means cumulatively calculates a new correction amount obtained by the schedule correction amount calculation means each time the thermal power plant is repeatedly started. An amount calculation means, a cumulative correction amount storage means for storing the cumulative correction amount obtained by the cumulative correction amount calculation means, and a schedule storage means for storing a start schedule applied to actual start in the past, The past operating schedule and operating conditions are respectively taken out from the operating condition storing means for storing the operating conditions at the time of actual start-up in the past, the schedule storing means and the operating condition storing means, and the adaptive knowledge acquiring means obtains them. The output value obtained from the output unit when input to the input unit of the neural network in Learning means and the thermal power plant autonomous adaptive optimization control system for determining the internal bond strength of the neural network to match the cumulative correction amount stored in the amount storage means. 5. The start schedule according to claim 1 or 2, comprising a start schedule of the gas turbine and a start schedule of the steam turbine, and the start schedule of the gas turbine defines a predetermined start schedule. The predetermined parameter is at least one of a speed-up rate, a rated speed holding time, an initial load, an initial load holding time, a load rising rate, and a load holding time, and the steam turbine startup schedule Is configured by a predetermined parameter that defines the startup schedule, and the predetermined parameter is
At least one of high-pressure turbine bypass valve operating speed, medium-pressure turbine bypass valve operating speed, low-pressure turbine bypass valve operating speed, high-pressure control valve operating speed, medium-pressure control valve operating speed, low-pressure control valve operating speed Autonomous adaptive optimization control system. 6. The operating conditions stored in the operating condition storage means according to claim 4, wherein the thermal power plant is stopped, the equipment temperature and pressure of the plant, the condenser cooling water temperature, the atmospheric temperature and humidity, A thermal power plant autonomous adaptive optimization control system that is at least one of plant operation restriction conditions. 7. The autonomy managing means according to claim 1 or 2, wherein an activation schedule X ₁ determined by a correction amount obtained from the schedule correction amount calculating means is a first activation schedule, and X ₁ is the neural network. When the activation schedule X ₂ obtained by adding the correction amount ΔX _OUT and X ₁ obtained from the output section when inputting together with the operating conditions into the second activation schedule is set to the first activation schedule X ₁ . Complement of the autonomous coefficient (1
1-?) Value obtained by multiplying the _(1-λ) X 1 and the second start scheduled X ₂ in the autonomous factor activation is obtained by adding the value .LAMBDA.x ₂ obtained by multiplying the lambda schedule X A thermal power plant autonomous adaptive optimization control system which adopts _R as an actual start schedule and transfers it to the schedule execution control means. 8. The autonomy managing means according to claim 1 or 2, wherein the autonomy managing means has an autonomic coefficient generating means for varying the autonomic coefficient .lamda.
Alternatively, in accordance with the schedule correction amount obtained by the schedule correction amount calculation means, a thermal power plant autonomous adaptive optimization control system in which λ is increased in the range of 0 ≦ λ ≦ 1 as the number of startups is increased. 9. The learning management according to claim 1, 2, 3 or 4, wherein the upper N samples from the one with the largest learning error among the plurality of learning samples are prioritized to correct the connection strength in the neural network. means. 10. The connection strength in a neural network according to claim 1, 2, 3 or 4, wherein priority is given to M schedule parameters having higher learning errors out of the schedule parameters having larger learning errors for each learning sample. Learning management means to do. 11. The learning management means according to claim 1, 2, 3, 4, 9 or 10, wherein schedule parameters that have already become less than or equal to a predetermined error are excluded from the correction target of the connection strength.