Optimization method of flow characteristic function of steam turbine valve based on GA-LSTM

The GA-LSTM method optimizes the flow characteristic function of the steam turbine valve, which solves the problem that the valve flow characteristics are not ideal and affects the frequency regulation accuracy, and achieves more efficient flow characteristic curve optimization, which improves the grid frequency stability.

CN113962140BActive Publication Date: 2024-11-19STATE GRID HEBEI ELECTRIC POWER RES INST +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111072154.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-09-14
Publication Date
2024-11-19
Estimated Expiration
2041-09-14

AI Technical Summary

Technical Problem

The flow characteristics of the steam turbine valve are not ideally affected by the accuracy of the primary frequency regulation, resulting in the linearity of the flow characteristics, and the flow characteristic curve of the valve needs to be corrected and optimized.

Method used

Using a GA-LSTM-based method, by collecting historical operation data, building an LSTM neural network model and optimizing model parameters using genetic algorithms, establishing a steam turbine valve flow characteristic function, and drawing the optimized flow characteristic curve.

Benefits of technology

Avoid the time-consuming valve flow characteristic test in traditional methods, improve the accuracy and level of frequency regulation at one time, provide an accurate, adaptable and fast optimization method, reduce the workload of staff, and improve the valve flow characteristics. Model accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113962140B_ABST
    Figure CN113962140B_ABST
Patent Text Reader

Abstract

The invention relates to an optimization method for a turbine valve flow characteristic function based on GA-LSTM, comprising the following steps: collecting historical operation data; determining a training set and a verification set of the data; constructing a turbine valve flow characteristic function optimization model based on LSTM; constructing a turbine valve flow characteristic model based on a GA-LSTM neural network; optimizing the turbine valve flow characteristic model through the GA-LSTM neural network, and drawing an optimized turbine valve flow characteristic curve; the invention avoids a time-consuming valve flow characteristic test, reduces the workload of staff, obtains an optimal fitting function through deep learning of massive historical data, and further optimizes the relationship between a turbine comprehensive valve position instruction and an actual steam inlet amount to obtain an optimal turbine valve flow characteristic curve, thereby providing support for turbine valve flow characteristic analysis and parameter optimization thereof.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a method for optimizing a flow characteristic function of a steam turbine valve based on GA-LSTM. Background Art

[0002] With the continuous progress of social economy, electrical appliances have played an important role in people's lives, and people's electricity consumption has also increased. This requires power plants to provide high-quality and stable electricity to ensure people's living needs. Grid frequency is an important factor affecting the quality of power, and primary frequency regulation is also an important means to maintain the stability of grid frequency. The unsatisfactory flow characteristics of the steam turbine valve will affect the accuracy of the primary frequency regulation, resulting in the deterioration of the linearity of the valve flow characteristics. Therefore, in order to improve the accuracy and level of primary frequency regulation, it is of great significance to correct and optimize the flow characteristic curve of the valve.

[0003] Genetic Algorithm (GA) is an algorithm that simulates the biological evolution mechanism in nature. It introduces concepts such as chromosomes, populations, offspring, and parents to simulate natural evolution. It is often used to solve global optimal problems and search the global environment to avoid the problem that other algorithms that use gradient descent are prone to fall into local optimal solutions.

[0004] Long Short-Term Memory (LSTM) neural network is an improved network of recurrent neural network, which can effectively avoid the gradient vanishing and gradient exploding problems in ordinary neural networks. The LSTM network structure adds a structure called memory unit to remember past information, and adds input gate, forget gate and output gate to control the transmission of historical information. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide an optimization method for the turbine valve flow characteristic function based on GA-LSTM, optimize the relationship between the turbine comprehensive valve position instruction and the actual steam intake amount to obtain the best turbine valve flow characteristic curve, and provide support for the turbine valve flow characteristic analysis and parameter optimization.

[0006] The technical solution adopted by the present invention is:

[0007] A method for optimizing a flow characteristic function of a steam turbine valve based on GA-LSTM comprises the following steps: S100: collecting historical operation data; S200: determining a training set and a validation set of the data; S300: constructing an optimization model for a flow characteristic function of a steam turbine valve based on LSTM; S400: using a GA algorithm to determine the optimal parameters of the LSTM model, and constructing a flow characteristic model of a steam turbine valve based on a GA-LSTM neural network; S500: optimizing the flow characteristic model of a steam turbine valve through the GA-LSTM neural network, and drawing an optimized flow characteristic curve of a steam turbine valve.

[0008] Furthermore, when S100 collects historical data, historical operation data with a time span of 4 months is selected throughout the year, and the data sampling time interval is 5s.

[0009] Furthermore, the collected parameters include load P * , Comprehensive valve position command R f , main steam flow Q1, main steam pressure P, main steam temperature T, regulating stage pressure P a , High pressure cylinder exhaust pressure P c , High pressure cylinder exhaust temperature T C , reheat steam pressure P d , Reheat steam temperature T d .

[0010] Furthermore, in S200, the collected data is normalized to map the samples to [0, 1] to obtain a data set D, and its normalization formula is:

[0011]

[0012] In formula 1, X ※ is the normalized data, X min , X max is the minimum and maximum value of the collected data, and X is the original sample data;

[0013] The comprehensive valve position command R f , main steam pressure P, regulating stage pressure P a The input variables of the model constitute the input vector X(k)=[R f ,P,P a ], and the actual steam intake amount Q of the turbine is taken as the output variable Y(k)=[Q] of the model.

[0014] Furthermore, the actual steam inlet amount of the steam turbine is calculated as follows:

[0015]

[0016] In formula 2, P a0is the rated regulating stage pressure, P a is the regulating stage pressure, P c0 is the rated high pressure cylinder exhaust pressure, P c is the high pressure cylinder exhaust pressure, T C0 is the rated high pressure cylinder exhaust temperature, T C is the exhaust temperature of the high-pressure cylinder. The data standardized by formula 1 are randomly selected for 20 days each month as the training set data D train , the remaining 10 or 11 days of each month are used as the validation set data D test .

[0017] Furthermore, the steps for constructing the turbine valve flow characteristic function optimization model based on LSTM are as follows:

[0018] S310: Initialize the network structure, number of hidden layers and number of output layers of the LSTM neural network, and initialize the training step size; set the connection matrix U from the input layer to the hidden layer, the connection matrix W from the hidden layer to the hidden layer, and the connection weight V from the hidden layer to the output layer of the LSTM neural network to all be random real numbers;

[0019] S320: Set the activation function of the LSTM neural network and set the input gate i t 、Forget gate t and output gate o t , the unit state activation function is the tanh function, and the input vector of the neural network is X=(X1,X2,X3,X4,X5……X t-1 ,X t ), the state of the hidden layer H=(h1,h2,h3,h4,h5……h t-1 ,h t ), the output vector is Y=(Y1,Y2,Y3,Y4,Y5……Y t-1 ,Y t );

[0020] S330: The preprocessed training set data D in S200 train The comprehensive valve position command R f , main steam pressure P, regulating stage pressure P a As the input variable of the model, X(k) = [R f ,P,P a ] and the actual steam inlet volume Q of the turbine are used as the output variable Y(k)=[Q] of the model and are respectively sent to the LSTM neural network models initialized by S310 and S320 to obtain the turbine valve flow characteristic function optimization model based on LSTM.

[0021] Furthermore, the calculation method of each unit gating in S320 is as follows:

[0022]

[0023] In formula 3, i t The input gate mainly determines how much current information is retained and input to the unit at the current moment, f t The forget gate is mainly used to decide whether to retain the unit state c at the previous moment. t-1 Information to the current unit state c t In, o t The output gate is mainly used to determine how many outputs the unit state has at the current moment;

[0024] Formula h t =o t ·tanh(c t ) is represented as the final output of the network, The current input unit status.

[0025] Furthermore, the specific steps of constructing the turbine valve flow characteristic model based on the GA-LSTM neural network in S400 are as follows:

[0026] S410: Initialize various parameters of the population, and use the number of hidden layer neurons m, learning rate lr and training times n in the LSTM network as initialization objects of the GA algorithm;

[0027] S420: Use the validation set data D determined in S200 test The comprehensive valve position command R f , main steam pressure P, regulating stage pressure P a As input variable, that is, X(k) = [R f ,P,P a ], imported into the LSTM-based turbine valve flow characteristic function optimization model established by S330, the output value of the model can be obtained A network model estimate of the steam inlet to the turbine is represented by With the validation set data D test The root mean square error between the actual steam inlet volume Q of the steam turbine in is used as the size of the individual fitness. The choice of fitness function directly affects the effect of optimizing the flow characteristics of the steam turbine valve, and the fitness function is defined as follows:

[0028]

[0029] In formula 4, f i is the fitness function value, N is the validation set data D test The total amount of data, is the network prediction value of the steam inlet of the turbine output by the LSTM network model, Q nThe validation set data D test The actual value of the steam inlet to the turbine;

[0030] S430: Use the roulette wheel selection method as the selection strategy of GA to select individuals with better adaptability in the current population as parents, and pass the genetic information to the offspring for selection processing; use the shuffle crossover algorithm as the crossover operator, and use the random sorting function to perform a shuffle operation on the parent generation before crossover; use the rand() algorithm to generate a random number between (0,1) that is less than the given mutation rate, and then mutate;

[0031] S440: Calculate the fitness f of individuals in the population i , the fitness calculation function is in S430, the smaller the fitness in the genetic algorithm, the more the individual should be retained, otherwise the individual will be eliminated, if the genetic termination condition is not met, it returns to S430, if the termination condition of the genetic algorithm is met, the optimal parameters obtained by the GA algorithm are used as the number of hidden layer neurons m, the learning rate lr and the number of training times n of the LSTM network model;

[0032] S450: The optimal parameters of hidden layer neuron number m, learning rate lr and training times n obtained in S440 are introduced into the network to obtain an optimized model of turbine valve flow characteristic optimization based on GA-LSTM. The input variable of the model is the comprehensive valve position command R f , main steam pressure P, regulating stage pressure P a , that is, X(k)=[R f ,P,P a ], the output variable of the model is the steam inlet volume Q of the turbine, that is, Y(k) = [Q].

[0033] Further, in S410, the interval range of the hidden layer neuron m is set to [10, 50], the interval range of the learning rate lr is set to [0.001, 0.1], and the interval range of the number of training times n is set to [500, 5000].

[0034] Furthermore, in S430, the specific method of the roulette wheel selection method is as follows:

[0035] S431: The formula for calculating the probability of an individual being selected is as follows:

[0036]

[0037] Where n is the population size, f i is the fitness of the individual, P i is the probability of individual i being selected, calculated in generation t and

[0038] S432: Generate a random number rand() of {0,1}, and calculate S=rand()*f sum ;

[0039] S433: Find out The smallest k in , then it means that the kth individual is selected;

[0040] S434: Perform S432 and S433 N times to obtain N individuals, which become the t=t+1th generation population.

[0041] The positive effects of the present invention are: compared with traditional optimization methods, the present invention avoids the time-consuming valve flow characteristic test, reduces the workload of staff, obtains the optimal fitting function through deep learning of massive historical data, and then optimizes the relationship between the turbine comprehensive valve position instruction and the actual steam intake volume to obtain the best turbine valve flow characteristic curve, providing support for turbine valve flow characteristic analysis and parameter optimization.

[0042] The present invention proposes to use a GA-LSTM network model to establish a turbine valve flow characteristic model by deep learning of massive historical data, thereby optimizing the relationship between the comprehensive valve position instruction of the turbine and the actual steam intake, avoiding the problems of harsh test conditions, low precision, unit shutdown, etc. in the traditional field flow characteristic test of the power plant to obtain the flow characteristics of the turbine regulating valve, and providing an accurate, adaptable and fast optimization method for optimizing the flow characteristics of the turbine valve. It can be seen from step S100 and step S200 that the present invention is based on the massive historical operation data of the thermal power unit to obtain the unit operation information, and on this basis, the valve flow characteristic parameters of the unit are optimized, thereby avoiding the drawbacks of the traditional test being limited by the unit load operation range during the parameter optimization process. The turbine valve flow characteristic function optimization model based on LSTM established in step S300, because the LSTM network introduces a forget gate, can fully consider the time series characteristics of the historical data of the thermal power unit, and has a higher modeling accuracy than the traditional modeling method. To ensure the modeling accuracy, step S400 proposes to use the GA algorithm to optimize the number of hidden layer neurons m, the learning rate lr and the number of training times n of the LSTM network model, so as to avoid problems such as low precision and low efficiency caused by the traditional LSTM network selecting model parameters based on experience, and improve the accuracy of the turbine valve flow characteristic model. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a flow chart of the steps of the present invention;

[0044] Figure 2 It is a flow chart of the system of the present invention. DETAILED DESCRIPTION

[0045] As attached Figure 1 , 2 As shown, the present invention comprises the following steps:

[0046] Step 1: Collection of historical operation data. Obtain a large amount of historical operation data of the units through the distributed control system of the thermal power plant. Select historical operation data with a time span of 4 months (one month per quarter) throughout the year, and the data sampling interval is 5s. Within this data collection time span, there should be no less than 10 times of large-scale load increase and decrease data per month, and within this range, the thermal power generating units should have no faults or shutdown processes. The main parameters collected include load P * , Comprehensive valve position command R f , main steam flow Q1, main steam pressure P, main steam temperature T, regulating stage pressure P a , High pressure cylinder exhaust pressure P c , High pressure cylinder exhaust temperature T C , reheat steam pressure P d , Reheat steam temperature T d ;

[0047] Step 2: Normalize the collected data and map the samples to [0,1] to obtain the data set D. The normalization formula is:

[0048]

[0049] In formula 1, X * is the normalized data, X min , X max is the minimum and maximum value of the collected data, and X is the original sample data;

[0050] The comprehensive valve position command R f , main steam pressure P, regulating stage pressure P a The input variables of the model constitute the input vector X(k)=[R f ,P,P a ], and the actual steam inlet amount Q of the steam turbine is used as the output variable of the model Y(k)=[Q]. The actual steam inlet amount of the steam turbine is calculated as follows:

[0051]

[0052] In formula 2, P a0 is the rated regulating stage pressure, P a is the regulating stage pressure, P c0 is the rated high pressure cylinder exhaust pressure, P c is the high pressure cylinder exhaust pressure, T C0 is the rated high pressure cylinder exhaust temperature, T Cis the exhaust temperature of the high-pressure cylinder. The data standardized by formula 1 are randomly selected for 20 days each month as the training set data D train , the remaining 10 or 11 days of each month are used as the validation set data D test ;

[0053] Step 3: Construct the LSTM neural network model. The specific model construction steps are as follows:

[0054] Step 3-1: Initialize the network structure, number of hidden layers and output layers of the LSTM neural network, and initialize the training step size. Set the connection matrix U from the input layer to the hidden layer, the connection matrix W from the hidden layer to the hidden layer, and the connection weight V from the hidden layer to the output layer of the LSTM neural network to all random real numbers;

[0055] Step 3-2: Set the activation function of the LSTM neural network and set the input gate i t 、Forget gate t and output gate o t , the unit state activation function is the tanh function, and the input vector of the neural network is X=(X1,X2,X3,X4,X5……X t-1 ,X t ), the state of the hidden layer H = (h1,h2,h3,h4,h5 ... h t-1 ,h t ), the output vector is Y=(Y1,Y2,Y3,Y4,Y5……Y t-1 ,Y t ), the calculation method of each unit gate is as follows:

[0056]

[0057] In formula 3, i t The input gate mainly determines how much current information is retained and input to the unit at the current moment, f t The forget gate is mainly used to decide whether to retain the unit state c at the previous moment. t-1 Information to the current unit state c t In, o t The output gate is mainly used to determine how many outputs the current cell state has. t =o t ·tanh(c t ) is represented as the final output of the network, The current input unit status.

[0058] Step 3-3: The preprocessed training set data D in step 2 train The comprehensive valve position command R f , main steam pressure P, regulating stage pressure Pa As the input variable of the model, X(k) = [R f ,P,P a ] and the actual steam inlet amount Q of the steam turbine are used as the output variable Y(k)=[Q] of the model and are respectively sent to the LSTM neural network model initialized in step 3-1 and step 3-2, and the flow characteristic function optimization model of the steam turbine valve based on LSTM can be obtained;

[0059] Step 4: Construct a genetic network model and use the GA algorithm to optimize the LSTM neural network. The specific model construction steps are as follows:

[0060] Step 4-1: Initialize the parameters of the population, and use the number of hidden layer neurons m, learning rate lr and training times n in the LSTM network as the initialization objects of the GA algorithm. Set the interval range of hidden layer neurons m to [10,50], the interval range of learning rate lr to [0.001,0.1], and the interval range of training times n to [500,5000];

[0061] Step 4-2: Use the validation set data D determined in step 2 test The comprehensive valve position command R f , main steam pressure P, regulating stage pressure P a As input variable, that is, X(k) = [R f ,P,P a ], imported into the LSTM-based turbine valve flow characteristic function optimization model established in step 3-3, the output value of the model can be obtained A network model estimate of the steam inlet to the turbine is represented by With the validation set data D test The root mean square error between the actual steam inlet volume Q of the steam turbine in is used as the size of the individual fitness. The choice of fitness function directly affects the effect of optimizing the flow characteristics of the steam turbine valve, and the fitness function is defined as follows:

[0062]

[0063] In formula 4, f i is the fitness function value, N is the validation set data D test The total amount of data, is the network prediction value of the steam inlet of the turbine output by the LSTM network model, Q n The validation set data D test The actual value of the steam inlet to the turbine;

[0064] Step 4-3: Use the roulette wheel selection method as the selection strategy of GA to select individuals with better adaptability as parents in the current population, and pass the genetic information to the offspring for selection. Use the shuffle crossover algorithm as the crossover operator, and use the random sorting function to shuffle the parent generation before crossover. Use the rand() algorithm to generate a random number between (0,1) that is less than the given mutation rate, and then mutate. The specific method of the roulette wheel selection method is as follows:

[0065] Step 4-3-1: Calculate the probability of an individual being selected. The formula is as follows:

[0066]

[0067] Where n is the population size, f i is the fitness of the individual, P i is the probability of individual i being selected. Calculated in the tth generation and

[0068] Step 4-3-2: Generate a random number rand() from {0,1}, and calculate S = rand()*f sum ;

[0069] Step 4-3-3: Find The smallest k in , then it means that the kth individual is selected;

[0070] Step 4-3-4: Perform steps 4-3-2 and 4-3-3 N times to obtain N individuals, which become the t=t+1 generation population.

[0071] Step 4-4: Calculate the fitness f of individuals in the population i , the fitness calculation function in step 4-2, the smaller the fitness in the genetic algorithm, the more the individual should be retained, otherwise the individual will be eliminated. If the genetic termination condition is not met, it returns to step 4-3. If the termination condition of the genetic algorithm is met, the optimal parameters obtained by the GA algorithm are used as the number of hidden layer neurons m, the learning rate lr and the number of training times n of the LSTM network model;

[0072] Step 4-5: Bring the optimal parameters of hidden layer neuron number m, learning rate lr and training times n obtained in step 4-4 into the network to obtain the optimized model of turbine valve flow characteristic optimization based on GA-LSTM. The input variable of the model is the comprehensive valve position command R f , main steam pressure P, regulating stage pressure P a , that is, X(k)=[R f ,P,P a ], the output variable of the model is the steam inlet volume Q of the steam turbine, that is, Y(k) = [Q];

[0073] Step 5: Through steps 3 and 4, a model for optimizing the flow characteristic function of a steam turbine valve based on GA-LSTM is obtained. Through this model, the input variable X(k)=[R f ,P,P a ]The mapping relationship between the steam inlet volume Q of the steam turbine is obtained through nonlinear transformation, that is, the optimal fitting function Therefore, the output variable steam turbine inlet volume optimized by the model can be obtained The optimized steam inlet volume of the steam turbine As the ordinate and the comprehensive valve position command R f By plotting the turbine valve flow characteristic curve as the horizontal axis, the turbine valve flow characteristic curve optimized by the GA-LSTM network model can be obtained.

[0074] The present invention proposes a method for optimizing the flow characteristic function of a steam turbine valve based on GA-LSTM, which includes data collection, determination of input and output variables, optimization of LSTM network parameters using GA, construction of a GA-LSTM network model, and optimization of the valve flow characteristic function. The method proposes to use the historical operation data of the unit, and to form a set of input and output variables by comprehensive valve position instructions, main steam pressure, regulating stage pressure, and actual steam inlet flow parameters of the steam turbine, and to use the GA algorithm to optimize the key hyperparameters of the LSTM network, to obtain the optimal combination of the number of neurons, learning rate, and number of training times in the LSTM network, and to construct a GA-LSTM network model to optimize the valve flow characteristic function. The present invention is of great significance for improving the accuracy of primary frequency modulation and the safe and stable operation of steam turbines.

Claims

1. An optimization method for turbine valve flow characteristic function based on GA-LSTM, Features It includes the following steps: S100: Collect historical operation data; S200: Determine a training set and a validation set of data; S300: Constructing a steam turbine valve flow characteristic function optimization model based on LSTM; S400: Use the GA algorithm to determine the optimal parameters of the LSTM model and build a turbine valve flow characteristic model based on the GA-LSTM neural network; S500: Optimize the turbine valve flow characteristic model through the GA-LSTM neural network and draw the optimized turbine valve flow characteristic curve; The steps for constructing the turbine valve flow characteristic function optimization model based on LSTM are as follows: S310: Initialize the network structure, number of hidden layers and number of output layers of the LSTM neural network, and initialize the training step size; set the connection matrix U from the input layer to the hidden layer, the connection matrix W from the hidden layer to the hidden layer, and the connection weight V from the hidden layer to the output layer of the LSTM neural network to all be random real numbers; S320: Set the activation function of the LSTM neural network and set the input gate i t 、Forget gate t and output gate o t , the unit state activation function is the tanh function, and the input vector of the neural network is X=(X 1 , X 2 , X 3 , X 4 , X 5 ...X t-1 , X t ), the state of the hidden layer H=(h 1 ,h 2 ,h 3 ,h 4 ,h 5 ……h t-1 ,h t ), whose output vector is And=(And 1 ,AND 2 ,AND 3 ,AND 4 ,AND 5 ……AND t-1 ,AND t ); S330: The preprocessed training set data D in S200 train The comprehensive valve position command R f , main steam pressure P, regulating stage pressure P a As the input variable of the model, X(k) = [R f ,P,P a ] and the actual steam inlet amount Q of the steam turbine are used as the output variable Y(k)=[Q] of the model and are respectively sent to the LSTM neural network models initialized by S310 and S320 to obtain the flow characteristic function optimization model of the steam turbine valve based on LSTM; The specific steps of constructing the turbine valve flow characteristic model based on GA-LSTM neural network in S400 are as follows: S410: Initialize various parameters of the population, and use the number of hidden layer neurons m, learning rate lr and training times n in the LSTM network as initialization objects of the GA algorithm; S420: Use the validation set data D determined in S200 test The comprehensive valve position command R f , main steam pressure P, regulating stage pressure P a As input variable, that is, X(k) = [R f ,P,P a ], imported into the LSTM-based turbine valve flow characteristic function optimization model established by S330, the output value of the model can be obtained A network model estimate of the steam inlet to the turbine is represented by With the validation set data D test The root mean square error between the actual steam inlet volume Q of the steam turbine in is used as the size of the individual fitness. The choice of fitness function directly affects the effect of optimizing the flow characteristics of the steam turbine valve, and the fitness function is defined as follows: In formula 4, f i is the fitness function value, N is the validation set data D test The total amount of data, Q is the network prediction value of the steam inlet of the turbine output by the LSTM network model, n The validation set data D test The actual value of the steam inlet to the turbine; S430: Use the roulette wheel selection method as the selection strategy of GA to select individuals with better adaptability in the current population as parents, and pass the genetic information to the offspring for selection processing; use the shuffle crossover algorithm as the crossover operator, and use the random sorting function to perform a shuffle operation on the parent generation before crossover; use the rand() algorithm to generate a random number between (0, 1) that is less than the given mutation rate, and then mutate; S440: Calculate the fitness f of individuals in the population i , the fitness calculation function is in S430, the smaller the fitness in the genetic algorithm, the more the individual should be retained, otherwise the individual will be eliminated, if the genetic termination condition is not met, it returns to S430, if the termination condition of the genetic algorithm is met, the optimal parameters obtained by the GA algorithm are used as the number of hidden layer neurons m, the learning rate lr and the number of training times n of the LSTM network model; S450: The optimal parameters of hidden layer neuron number m, learning rate lr and training times n obtained in S440 are introduced into the network to obtain an optimized model of turbine valve flow characteristic optimization based on GA-LSTM. The input variable of the model is the comprehensive valve position command R f , main steam pressure P, regulating stage pressure P a , that is, X(k)=[R f ,P,P a ], the output variable of the model is the steam inlet volume Q of the turbine, that is, Y(k) = [Q].

2. The optimization method of the flow characteristic function of a steam turbine valve based on GA-LSTM according to claim 1, Features When S100 collects historical data, historical operation data with a time span of 4 months is selected throughout the year, and the data sampling time interval is 5s.

3. The optimization method of the flow characteristic function of a steam turbine valve based on GA-LSTM according to claim 2, Features The collected parameters include load P * , Comprehensive valve position command R f , Main steam flow Q 1 , main steam pressure P, main steam temperature T, regulating stage pressure P a , High pressure cylinder exhaust pressure P c , High pressure cylinder exhaust temperature T C , reheat steam pressure P d , Reheat steam temperature T d .

4. The optimization method of the flow characteristic function of a steam turbine valve based on GA-LSTM according to claim 1, Features In S200, the collected data is normalized to map the samples to [0, 1] to obtain a data set D, whose normalization formula is: In formula 1, X ※ is the normalized data, X min , X max is the minimum and maximum value of the collected data, and X is the original sample data; The comprehensive valve position command R f , main steam pressure P, regulating stage pressure P a The input variables of the model constitute the input vector X(k)=[R f ,P,P a ], and the actual steam intake amount Q of the turbine is taken as the output variable Y(k)=[Q] of the model.

5. The optimization method of the flow characteristic function of a steam turbine valve based on GA-LSTM according to claim 4, Features The actual steam inlet amount of the steam turbine is calculated as follows: In formula 2, P a0 is the rated regulating stage pressure, P a is the regulating stage pressure, P c0 is the rated high pressure cylinder exhaust pressure, P c is the high pressure cylinder exhaust pressure, T C0 is the rated high pressure cylinder exhaust temperature, T C is the exhaust temperature of the high-pressure cylinder. The data standardized by formula 1 are randomly selected for 20 days each month as the training set data D train , the remaining 10 or 11 days of each month are used as the validation set data D test .

6. The optimization method of the flow characteristic function of a steam turbine valve based on GA-LSTM according to claim 1, Features The calculation method of each unit gate in S320 is as follows: In formula 3, i t The input gate mainly determines how much current information is retained and input to the unit at the current moment, f t The forget gate is mainly used to decide whether to retain the unit state c at the previous moment. t-1 Information to the current unit state c t In, o t The output gate is mainly used to determine how many outputs the unit state has at the current moment; Formula h t =o t ·tanh(c t ) is represented as the final output of the network, The current input unit status.

7. The optimization method of the flow characteristic function of a steam turbine valve based on GA-LSTM according to claim 1, Features In S410, the interval range of the hidden layer neuron m is set to [10, 50], the interval range of the learning rate lr is set to [0.001, 0.1], and the interval range of the training times n is set to [500, 5000].

8. The method for optimizing the flow characteristic function of a steam turbine valve based on GA-LSTM according to claim 1, Features In S430, the specific method of the roulette wheel selection method is as follows: S431: The formula for calculating the probability of an individual being selected is as follows: Where n is the population size, f i is the fitness of the individual, P i is the probability of individual i being selected, calculated in the tth generation and S432: Generate a random number rand() of {0, 1}, and calculate S=rand()*f sum ; S433: Find out The smallest k in , then it means that the kth individual is selected; S434: Perform S432 and S433 N times to obtain N individuals, which become the t=t+1th generation population.