Micro-gas turbine control method based on state space model in combination with BP neural network algorithm and genetic algorithm

By combining state space model, BP neural network and genetic algorithm in micro-gas turbine control, the local optimal solution and slow convergence problems of micro-gas turbine control algorithm are solved, and more efficient and stable operation is achieved.

CN119066987APending Publication Date: 2024-12-03ZHEJIANG UNIV CITY COLLEGE

Patent Information

Application Number
CN202411554197.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-11-04
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The existing micro-gas turbine control algorithms have problems of local optimal solutions and slow convergence, making it difficult to achieve efficient, stable and safe operation under various operating conditions.

Method used

The micro-gas engine control method based on state space model combined with BP neural network algorithm and genetic algorithm is adopted. By establishing a mathematical mechanism model, fitting transfer functions, initializing a state space model, building a multi-step prediction model of BP neural network, training and computing errors, Feedback correction and optimize control sequences using genetic algorithms to avoid local optimal solutions and slow convergence.

Benefits of technology

Improve the accuracy of control, avoid the problems of local optimal solution and slow convergence, and achieve more efficient and stable micro-gass engine operation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066987A_ABST
    Figure CN119066987A_ABST
Patent Text Reader

Abstract

The invention provides a micro-gas turbine control method based on a state space model in combination with a BP neural network algorithm and a genetic algorithm, and the method comprises the following specific steps: S1, building a mathematical mechanism model of a micro-gas turbine, determining the input quantities as air flow, fuel quantity and air-fuel ratio, and determining the output quantities as exhaust gas temperature and output power; s2, fitting a transfer function through simulation, and linearizing the transfer function to serve as a state space model; s3, initializing the state space model in the step S2; s4, establishing and training a BP neural network multi-step prediction model; s5, calculating an error E between the actual output of the controlled object and the output of the BP neural network multi-step prediction model; s6, performing feedback correction on prediction output of the BP neural network multi-step prediction model; and S7, obtaining a group of optimal control sequences by using a genetic algorithm.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of micro gas engine control, and relates to a micro gas engine control method based on a state space model combined with a BP neural network algorithm and a genetic algorithm. Background Art

[0002] A micro gas turbine is a miniaturized gas turbine, usually with a power of no more than 1000KW, and is mainly used in distributed power generation, drones, small power plants, and some specific industrial applications. A micro gas turbine control system is a set of electronic components designed specifically to control and manage the operation of a micro gas turbine. The design and function of these control systems are key to ensuring efficient, stable, and safe operation of gas turbines under various operating conditions. Existing control algorithms for micro gas turbines have problems with local optimal solutions and slow convergence. Summary of the invention

[0003] The present invention provides a micro-turbine control method based on a state space model combined with a BP neural network algorithm and a genetic algorithm, which improves the accuracy of control and avoids the problems of falling into a local optimal solution and slow convergence.

[0004] The technical solution adopted by the present invention is: The micro-turbine control method based on the state space model combined with the BP neural network algorithm and the genetic algorithm has the following specific steps: S1, establish the mathematical mechanism model of the micro-turbine, determine the input quantity as air flow, fuel quantity and air-fuel ratio, and the output quantity as exhaust temperature and output power; S2, through simulation, fit the transfer function and linearize it as the state space model; S3, initializing the state space model of step S2; S4, establish and train the BP neural network multi-step prediction model; S5, calculating the error E between the actual output of the controlled object and the output of the BP neural network multi-step prediction model; S6, performing feedback correction on the prediction output of the BP neural network multi-step prediction model; S7, using genetic algorithm to obtain a set of optimal control sequences.

[0005] Furthermore, the method further includes step S8, determining whether the simulation time is less than a preset simulation time, if yes, proceeding to step S3, if no, outputting the result.

[0006] Furthermore, the mathematical mechanism model in step S1 includes a compressor module, a combustion chamber module, a turbine module, a rotor module, a generator module and a volume inertia module.

[0007] Furthermore, the micro-turbine in step S1 adopts a Capston centrifugal gas turbine and is simulated through simulink.

[0008] Furthermore, the BP neural network multi-step prediction model in step S4 adopts a recursive multi-step BP neural network to construct a single neural network model for multi-step prediction. The input and output of the neural network are as follows: ; In the formula, u(k), u(k-1), …, u(k-m+1) represent the input variables or control variables of the system at different time steps; y(k), …, y(k-n+1) represent the historical data of the output variables or controlled variables of the system at different time steps; Represents the prediction result of the model in the next Np time steps; m and n represent the order of the input and output of the nonlinear system respectively; Np is the prediction time domain or prediction step, X is the input vector of the BP neural network, and Y is the output vector of the BP neural network.

[0009] Further, the BP neural network training process of the BP neural network multi-step prediction model in step S4 includes: S41, initializing the neural network: confirming the input and output variables, and presetting the number of nodes of the input layer, hidden layer and output layer; among which, confirming the input parameters: air flow, fuel amount, air-fuel ratio; output parameters: output power, exhaust temperature; S42, after confirming the input parameters, the input parameters are normalized and the deviation is normalized, which is a linear transformation of the original data so that the values ​​are mapped to [0-1,0]. The conversion function is as follows: ; Among them, max is the maximum value of the sample data, min is the minimum value of the sample data range, x is the sampling value of the sample data, and x* is the data value after linear transformation; S43, after normalization processing, starting to construct a basic neural network, determining the number of neural network layers and the number of neurons in each layer; S44, after confirming the number of neurons in each layer, set the activation function f(x): ; Among them, e is a natural constant; Set the bias value of each neuron, that is, the default weight value of each neuron; S45, using a genetic algorithm to adjust the weight coefficients between the initial neurons of the BP neural network so that the error function E is minimized; S46, through multiple rounds of learning, until the training is stopped.

[0010] Furthermore, in step S41, it is also necessary to set the learning rate and activation function, and set the weights between layers and the thresholds within the layers; Hidden layer calculation: The hidden layer output H is calculated according to the following equation, where f is the activation function, ; In the formula, ω ij is the connection weight from the i-th node in the input layer to the j-th node in the hidden layer, x i is the input value of the i-th node in the input layer, a j is the bias of the jth node in the hidden layer, n represents the number of nodes in the input layer, and l represents the number of nodes in the hidden layer; Output layer calculation: According to the hidden layer output H and weight ω of the BP neural network jk And threshold b calculate the predicted value O k , ; In the formula, ω jk is the weight from the jth node in the hidden layer to the kth node in the output layer, b k is the bias of the kth node in the output layer, L represents the number of nodes in the output layer, and m represents the number of nodes in the output layer.

[0011] Further, the neuron basis function in step S43 is: ; Among them, U represents the neuron function.

[0012] Furthermore, in step S45, the weight training steps of adjusting the initial neurons of the BP neural network by the genetic algorithm are as follows: S451, initialize the population, use roulette to obtain the population, the population size can be determined based on empirical values, the basic principle is to perform sampling calculations on the samples, and achieve the purpose of optimization through crossover operations; S452, the mutation operation and mutation rate are adjusted according to the empirical value and determined through experiments; S453, cyclically performing selection, crossover, mutation, and fitness value calculation operations until the number of evolutions is reached and the optimal initial weight is obtained; S454, obtaining the initial weights of the neural network, initializing the weights and thresholds using a genetic algorithm, and when the genetic algorithm completes one generation, reinserting the weights of the new generation into the neural network; S455, judging whether the error reaches an acceptable range by calculating the error, and the training ends if it is acceptable; S456, continuously adjusts the weights of each layer and its threshold; S457, neural network training is completed.

[0013] Further, the error E calculation formula of step S5 is as follows: ; Among them, t j is the expected value, y j is the output value of each neuron, E p is the error sum of a neural layer, E is the error sum of the entire neural network, p represents the sample number, P represents the total number of samples used for training or testing, j represents the node number in the output layer, and m represents the number of nodes in the output layer.

[0014] The beneficial effects of the present invention are as follows: by establishing a mathematical mechanism model of a micro gas turbine and performing simulation operations through simulink, the characteristics of a BP neural network are used as a prediction model, and a genetic algorithm is introduced to roll-optimize the control process, thereby avoiding falling into a local optimal solution and slow convergence problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the process of the present invention.

[0016] Figure 2 It is a schematic diagram of the model of the micro-turbine of the present invention.

[0017] Figure 3 It is a control flow diagram of the present invention.

[0018] Figure 4 It is a schematic diagram of the BP neural network model structure of the present invention.

[0019] Figure 5 It is a schematic diagram of the single-model recursive neural network structure of the present invention. DETAILED DESCRIPTION

[0020] The present invention is further described below in conjunction with specific embodiments, but the present invention is not limited to these specific embodiments. Those skilled in the art should recognize that the present invention covers all possible alternatives, improvements and equivalents within the scope of the claims.

[0021] Technical terms: Main components and functions of micro gas turbine: 1. Control unit: usually includes one or more microprocessors or digital signal processors (DSPs), responsible for implementing control algorithms, data processing and decision making.

[0022] 2. Sensors: used to monitor various parameters of the gas turbine, such as temperature, pressure, speed and flow, to provide real-time data to the control unit.

[0023] 3. Actuator: According to the instructions of the control unit, the actuator adjusts the fuel supply, adjusts the valve position, etc. to control the operating status of the gas turbine.

[0024] 4. Human Machine Interface (HMI): Allows the operator to monitor the status of the gas turbine, set parameters, and perform manual intervention when necessary.

[0025] 5. Communication interface: enables the control system to communicate with external systems (such as monitoring systems, other mechanical control systems, etc.) for data exchange and remote control.

[0026] Micro gas turbine control strategies and algorithms: • PID control: One of the most common control algorithms used to maintain gas turbine operating parameters (such as speed, temperature, etc.) near preset target values.

[0027] • Advanced control strategies: such as improved PID control algorithm, fuzzy logic control, neural network control, adaptive control, etc., which are used to deal with nonlinear systems or to improve control accuracy and stability in environments with large parameter changes.

[0028] • Fault diagnosis and management: Control systems also include fault detection and diagnostic capabilities, as well as the ability to take preventive action when abnormal operation or failure is detected.

[0029] See also Figure 1-Figure 5 The present invention provides a micro-turbine control method based on a state space model combined with a BP neural network algorithm and a genetic algorithm, and the specific steps are as follows: S1, establish the mathematical mechanism model of the micro-turbine, determine the input as air flow, fuel amount and air-fuel ratio, and the output as exhaust temperature and output power; the mathematical mechanism model includes compressor module, combustion chamber module, turbine module, rotor module, generator module and volume inertia module. The micro-turbine adopts Capston centrifugal gas turbine, and the parameters required by the model are all matched with this model, including but not limited to rated power, etc., and simulated through simulink.

[0030] S2, through simulation, fit the transfer function and linearize it as the state space model; S3, initializing the state space model of step S2; for example, initializing various parameters such as system simulation time; S4, establishing and training a BP neural network multi-step prediction model; offline training the BP neural network multi-step prediction model according to system training data and using test data to test the accuracy of the prediction model; The BP neural network multi-step prediction model adopts a recursive multi-step BP neural network to construct a single neural network model for multi-step prediction. The input and output of the neural network are as follows: ; In the formula, u(k), u(k-1), …, u(k-m+1) represent the input variables or control variables of the system at different time steps; y(k), …, y(k-n+1) represent the historical data of the output variables or controlled variables of the system at different time steps; Represents the prediction result of the model in the next Np time steps; m and n represent the order of the input and output of the nonlinear system respectively; Np is the prediction time domain or prediction step, X is the input vector of the BP neural network, and Y is the output vector of the BP neural network.

[0031] The BP neural network training process of the BP neural network multi-step prediction model includes: S41, initializing the neural network: confirming the input and output variables, and presetting the number of nodes of the input layer, hidden layer and output layer; among which, confirming the input parameters: air flow, fuel amount, air-fuel ratio; output parameters: output power, exhaust temperature; In step S41, it is also necessary to set the learning rate and activation function, and set the weights between layers and the thresholds within the layers; Hidden layer calculation: The hidden layer output H is calculated according to the following equation, where f is the activation function, ; In the formula, ω ij is the connection weight from the i-th node in the input layer to the j-th node in the hidden layer, x i is the input value of the i-th node in the input layer, a j is the bias of the jth node in the hidden layer, n represents the number of nodes (neurons) in the input layer, and l represents the number of nodes (neurons) in the hidden layer; Output layer calculation: According to the hidden layer output H and weight ω of the BP neural network jk And threshold b calculate the predicted value O k , ; In the formula, ω jk is the weight from the jth node in the hidden layer to the kth node in the output layer, b k is the bias of the kth node in the output layer, L represents the number of nodes (neurons) in the output layer, and m represents the number of nodes (neurons) in the output layer.

[0032] S42, after confirming the input parameters, the input parameters are normalized and the deviation is normalized, which is a linear transformation of the original data so that the values ​​are mapped to [0-1,0]. The conversion function is as follows: ; Among them, max is the maximum value of the sample data, min is the minimum value of the sample data range, x is the sampling value of the sample data, and x* is the data value after linear transformation; S43, after normalization processing, starting to construct a basic neural network, determining the number of neural network layers and the number of neurons in each layer; The neuron basis function is: ; Among them, w i represents the weight of the i-th input from the input layer to the hidden layer or output layer, x i represents the input value of the i-th node in the input layer, n represents the number of nodes in the input layer, and β represents the bias, which is a parameter used to adjust the calculation results.

[0033] S44, after confirming the number of neurons in each layer, set the activation function f(x): ; Among them, e is a natural constant; Set the bias value of each neuron, that is, the default weight value of each neuron; S45, using a genetic algorithm to adjust the weight coefficients between the initial neurons of the BP neural network so that the error function E is minimized; The steps of training the weights between the initial neurons of the BP neural network by using the genetic algorithm are as follows: S451, initialize the population, use roulette to obtain the population, the population size can be determined based on empirical values, the basic principle is to perform sampling calculations on the samples, and achieve the purpose of optimization through crossover operations; S452, mutation operation and mutation rate are adjusted according to the empirical value, and the new population y=(y1…y2) is determined through experiments; ; In the formula, rand represents a random number between 0 and 1, β represents the distribution index, and p m is the mutation rate, a k , b k are the upper and lower bounds of the optimization variable value; S453, cyclically performing selection, crossover, mutation, and fitness value calculation operations until the number of evolutions is reached and the optimal initial weight is obtained; S454, obtaining the initial weights of the neural network, initializing the weights and thresholds using a genetic algorithm, and when the genetic algorithm completes one generation, reinserting the weights of the new generation into the neural network; S455, judging whether the error reaches an acceptable range by calculating the error, and the training ends if it is acceptable; S456, continuously adjusts the weights of each layer and its threshold; S457, neural network training is completed.

[0034] S46, through multiple rounds of learning, until the training is stopped.

[0035] S5, calculating the error E between the actual output of the controlled object and the output of the BP neural network multi-step prediction model; The error E is calculated as follows: ; Among them, t j is the expected value, y j is the output value of each neuron, E p is the error sum of a neural layer, E is the error sum of the entire neural network, p represents the sample number, P represents the total number of samples used for training or testing, j represents the node number in the output layer, and m represents the number of nodes in the output layer.

[0036] S6, performing feedback correction on the prediction output of the BP neural network multi-step prediction model; S7, using genetic algorithm to obtain a set of optimal control sequences.

[0037] S8, determine whether the simulation time is less than the preset simulation time, if yes, go to step S3, if not, output the result.

[0038] The purpose of entering step S3 is to retrain the prediction model or optimize the control strategy. In the framework of genetic algorithms, this means re-performing a round of optimization process based on genetic operations (selection, crossover, mutation). Through the iteration and selection of the genetic algorithm, the prediction ability or control effect of the model is gradually improved, so that the model can better adapt to the dynamic changes of the system. Therefore, the core purpose of entering step S3 is to optimize the prediction model or control strategy in each simulation time step until the required performance is achieved. The output result is the control strategy or control sequence optimized by the genetic algorithm. During the simulation process, the genetic algorithm generates a set of optimal parameters or strategies through continuous iteration, which are used to control the actual system.

[0039] The present invention establishes a mathematical mechanism model of a micro gas turbine and performs simulation operations through simulink, uses the characteristics of a BP neural network as a prediction model, and introduces a genetic algorithm to roll-optimize the control process to avoid falling into a local optimal solution and slow convergence.

Claims

1. A micro-turbine control method based on a state space model combined with a BP neural network algorithm and a genetic algorithm, characterized in that: The specific steps are as follows: S1, establish the mathematical mechanism model of the micro-turbine, determine the input quantity as air flow, fuel quantity and air-fuel ratio, and the output quantity as exhaust gas temperature and output power; S2, through simulation, fit the transfer function and linearize it as the state space model; S3, initializing the state space model of step S2; S4, establish and train the BP neural network multi-step prediction model; S5, calculating the error E between the actual output of the controlled object and the output of the BP neural network multi-step prediction model; S6, performing feedback correction on the prediction output of the BP neural network multi-step prediction model; S7, using genetic algorithm to obtain a set of optimal control sequences.

2. The micro-turbine control method based on the state space model combined with the BP neural network algorithm and the genetic algorithm according to claim 1 is characterized in that: The method further includes step S8, determining whether the simulation time is less than a preset simulation time, if yes, proceeding to step S3, if no, outputting the result.

3. The micro-turbine control method based on the state space model combined with the BP neural network algorithm and the genetic algorithm according to claim 1 is characterized in that: The mathematical mechanism model in step S1 includes a compressor module, a combustion chamber module, a turbine module, a rotor module, a generator module and a volume inertia module.

4. The micro-turbine control method based on the state space model combined with the BP neural network algorithm and the genetic algorithm according to claim 3 is characterized in that: The micro-turbine in step S1 adopts a Capston centrifugal gas turbine and is simulated through simulink.

5. The micro-turbine control method based on the state space model combined with the BP neural network algorithm and the genetic algorithm according to claim 1 is characterized in that: The BP neural network multi-step prediction model in step S4 adopts a recursive multi-step BP neural network to construct a single neural network model for multi-step prediction. The input and output of the neural network are as follows: ; In the formula, u(k), u(k-1), …, u(k-m+1) represent the input variables or control variables of the system at different time steps; y(k), …, y(k-n+1) represent the historical data of the output variables or controlled variables of the system at different time steps; Represents the prediction result of the model in the next Np time steps; m and n represent the order of the input and output of the nonlinear system respectively; Np is the prediction time domain or prediction step, X is the input vector of the BP neural network, and Y is the output vector of the BP neural network.

6. The micro-turbine control method based on the state space model combined with the BP neural network algorithm and the genetic algorithm according to claim 5 is characterized in that: The BP neural network training process of the BP neural network multi-step prediction model in step S4 includes: S41, initializing the neural network: confirming the input and output variables, and presetting the number of nodes of the input layer, hidden layer and output layer; among which, confirming the input parameters: air flow, fuel amount, air-fuel ratio; output parameters: output power, exhaust temperature; S42, after confirming the input parameters, the input parameters are normalized and the deviation is normalized, which is a linear transformation of the original data so that the values ​​are mapped to [0-1,0]. The conversion function is as follows: ; Among them, max is the maximum value of the sample data, min is the minimum value of the sample data range, x is the sampling value of the sample data, and x* is the data value after linear transformation; S43, after normalization processing, starting to construct a basic neural network, determining the number of neural network layers and the number of neurons in each layer; S44, after confirming the number of neurons in each layer, set the activation function f(x): ; Among them, e is a natural constant; Set the bias value of each neuron, that is, the default weight value of each neuron; S45, using a genetic algorithm to adjust the weight coefficients between the initial neurons of the BP neural network so that the error function E is minimized; S46, through multiple rounds of learning, until the training is stopped.

7. The micro-turbine control method based on the state space model combined with the BP neural network algorithm and the genetic algorithm according to claim 6 is characterized in that: In step S41, it is also necessary to set the learning rate and activation function, and set the weights between layers and the thresholds within the layers; Hidden layer calculation: The hidden layer output H is calculated according to the following equation, where f is the activation function, ; In the formula, ω ij is the connection weight from the i-th node in the input layer to the j-th node in the hidden layer, x i is the input value of the i-th node in the input layer, a j is the bias of the jth node in the hidden layer, n represents the number of nodes in the input layer, and l represents the number of nodes in the hidden layer; Output layer calculation: According to the hidden layer output H and weight ω of the BP neural network jk And threshold b calculate the predicted value O k , ; In the formula, ω jk is the weight from the jth node in the hidden layer to the kth node in the output layer, b k is the bias of the kth node in the output layer, L represents the number of nodes in the output layer, and m represents the number of nodes in the output layer.

8. The micro-turbine control method based on the state space model combined with the BP neural network algorithm and the genetic algorithm according to claim 6 is characterized by: The neuron basis function in step S43 is: ; Among them, U represents the neuron function.

9. The micro-turbine control method based on the state space model combined with the BP neural network algorithm and the genetic algorithm according to claim 6 is characterized by: The training steps of adjusting the weights between the initial neurons of the BP neural network by the genetic algorithm in step S45 are as follows: S451, initialize the population, use roulette to obtain the population, the population size can be determined based on empirical values, the basic principle is to perform sampling calculations on the samples, and achieve the purpose of optimization through crossover operations; S452, the mutation operation and mutation rate are adjusted according to the empirical value and determined through experiments; S453, cyclically performing selection, crossover, mutation, and fitness value calculation operations until the number of evolutions is reached and the optimal initial weight is obtained; S454, obtaining the initial weights of the neural network, initializing the weights and thresholds using a genetic algorithm, and when the genetic algorithm completes one generation, reinserting the weights of the new generation into the neural network; S455, judging whether the error reaches an acceptable range by calculating the error, and the training ends if it is acceptable; S456, continuously adjusts the weights of each layer and its threshold; S457, neural network training is completed.

10. The micro-turbine control method based on state space model combined with BP neural network algorithm and genetic algorithm according to claim 1 is characterized in that: The error E calculation formula of step S5 is as follows: ; Among them, t j is the expected value, y j is the output value of each neuron, E p is the error sum of a neural layer, E is the error sum of the entire neural network, p represents the sample number, P represents the total number of samples used for training or testing, j represents the node number in the output layer, and m represents the number of nodes in the output layer.

Citation Information

Patent Citations

  • Heating, ventilating and air conditioning energy-saving control method based on genetic algorithm and depth BP neural network algorithm

    CN106765959A

  • Multi-step prediction method of parking based on optimized wavelet neural network

    CN108091135A

  • Uncertain system model predictive control parameter setting method based on machine learning

    CN112327625A

  • Integrated energy system energy management method based on improved deep reinforcement learning

    CN115409645A

  • Heavy duty gas turbine system control method based on model predictive control

    CN117270447A

Cited By

  • Flight control method and system based on nonlinear model

    CN120353238A