Method for optimizing self-adaptive predictive controller of hydroelectric generating set based on PSO and GPC

Abstract

The invention provides a hydroelectric generating set adaptive predictive controller optimization method based on PSO and GPC, and relates to the technical field of safe and stable operation of hydroelectric generating sets. A non-linear water turbine model is constructed based on a BP neural network, an Octave tool is adopted to construct a non-linear water turbine adjusting system model, then a particle swarm optimization algorithm is adopted to obtain a steady-state control signal of the non-linear water turbine adjusting system model, a non-linear compensator is adopted for correction, and a GPC, a PSO and the compensator are integrated to form a self-adaptive controller; system input and output are comprehensively considered, more efficient and more stable operation of the hydroelectric generating set under different working conditions is realized, so that the service life of equipment is prolonged, the maintenance cost is reduced, the electric energy generation efficiency is improved, and the control error under a nonlinear working condition is greatly reduced; through multi-model cooperative control, the dynamic characteristic fitting accuracy of the water turbine is improved, the adjusting time is shortened, the energy consumption is reduced, and the service life of the water turbine is prolonged.

Description

Technical Field

[0001] The present invention relates to the technical field of safe and stable operation of hydropower units, and in particular to a method for optimizing a hydropower unit adaptive predictive controller based on PSO and GPC. Background Art

[0002] Hydropower generators play a vital role in modern power systems, with their operating efficiency and stability directly impacting the reliability and affordability of power supply. However, hydropower generators operate in a complex and dynamic environment, characterized by significant nonlinearity, time-varying behavior, and strong coupling. Traditional control methods, particularly PID controllers, while simple and widely used, have significant limitations when applied to complex hydropower generator systems.

[0003] Traditional PID controllers regulate system output through three steps: proportional, integral, and derivative. These controllers suffer from the following limitations: 1. Parameter rigidity: PID controllers rely on fixed parameters and are unable to adapt to the nonlinearities (such as flow-torque characteristics) and time-varying operating conditions (such as head fluctuations) of hydropower units. 2. Insufficient anti-interference capability: PID controllers respond slowly to dynamic events such as grid frequency disturbances and sudden load changes, with overshoots generally exceeding 10%. 3. Hydropower unit systems have significant hysteresis effects, and PID controllers are often sluggish when dealing with these systems, making it difficult to quickly track changes in target values. More importantly, PID controllers lack anti-interference capability. PID controllers are not sensitive enough to external disturbances in the system, making it difficult to maintain system stability and performance in complex environments.

[0004] To overcome the shortcomings of traditional PID controllers, many advanced control methods have been proposed, such as fuzzy control, neural network control, and model predictive control (MPC). While these methods have improved system control performance to a certain extent, they still have some limitations. Many advanced control methods rely on precise system models, but the complexity and uncertainty of hydropower units make it difficult to establish accurate models. Furthermore, some advanced control methods (such as MPC) require extensive computing resources, making real-time control difficult to achieve, especially under rapidly changing operating conditions. Although these methods have good robustness in theory, in practice, system uncertainties and disturbances can still lead to reduced control performance.

[0005] Given the limitations of traditional PID controllers and advanced control methods in that they do not adequately consider system inputs and outputs, it is necessary to propose a design method for an adaptive predictive controller for hydropower units based on particle swarm optimization algorithm and generalized predictive control. This method comprehensively considers factors such as system inputs and outputs, and enables more efficient and stable operation of hydropower units under different working conditions, thereby extending equipment life, reducing maintenance costs, and improving power generation efficiency. Summary of the Invention

[0006] The main purpose of the present invention is to provide a method for optimizing the adaptive predictive controller of a hydropower unit based on PSO and GPC, so as to solve the technical problems in the prior art such as lack of dynamic compensation mechanism, lack of multi-objective optimization, insufficient anti-interference ability and high computational complexity.

[0007] To solve the above technical problems, the technical solution adopted by the present invention is: a method for optimizing a hydropower unit adaptive predictive controller based on PSO and GPC, comprising the following steps: S1: Constructing nonlinear turbine model based on BP neural network; S2: Octave tool is used to construct a nonlinear turbine regulation system model, which includes a water diversion system model, a turbine model, a generator and power grid model, a speed governor and a servo system model; S3: Using particle swarm optimization algorithm to obtain the steady-state control signal of the nonlinear turbine regulation system model; S4: Construct a nonlinear characteristic compensator for the adaptive predictive controller; S5: Determine hyperparameters in the generalized predictive controller; S6: Integrate the generalized predictive controller, particle swarm optimization algorithm and compensator to obtain an adaptive controller.

[0008] In a preferred embodiment, the step S1 further comprises the following steps: S11: Acquire turbine operation data and perform data preprocessing; S12: Based on the BP neural network structure, a nonlinear turbine model is constructed. The BP neural network uses the ReLU activation function and the input includes the unit speed. n 11 and guide vane opening Y ,in: The number of nodes in the input layer and the output layer is equal to the number of corresponding input parameters and output parameters; The number of hidden layers is 1, and the number of neurons can be determined according to the empirical formula: (1); Where: N i is the number of neurons in the input layer, N o is the number of neurons in the output layer, N h is the number of neurons in the hidden layer, Z is a constant between 0 and 10; S13: The nonlinear turbine model uses the preprocessed data to perform model training and save the optimal nonlinear turbine model.

[0009] In the preferred solution, model training is specifically as follows: Using forward propagation, that is, the input data is calculated through the network output; The loss function uses mean square error and performs back propagation to optimize the model; Repeat the above steps until the loss converges.

[0010] In the preferred solution, the data preprocessing includes: removing outliers, filling missing values ​​and normalization, specifically: Based on isolation forest, outlier detection is performed and missing values ​​are repaired using dynamic sliding windows; The operating intervals are divided according to the water head and guide vane opening, and each interval is independently standardized and normalized using the Sigmoid function.

[0011] In the preferred solution, in S2, the water diversion system model adopts a rigid water hammer model, which is mathematically described as: (2); Where, is the hydraulic time constant; The generator and grid model considers an infinite grid, and the generator and load transfer functions are: (3); Where, T a is the unit inertia time constant, e g is the generator load self-regulation coefficient.

[0012] In the preferred solution, the generalized predictive controller in S5 implements rolling optimization based on the CARIMA model and the Diophantine equation, and performs feedback correction, specifically: 1) Forecasting model: Based on the CARIMA model, the formula is: (4) ; Where, 、 、 are the output signal, input signal and white noise of the system respectively; d is the system delay order; Δ =1- z -1 is the difference operator; polynomial A ( z -1 ), B ( z -1 )and C ( z-1 )satisfy: (5); Where, n a 、 n b and n c Polynomials A 、 B and C The order of like , formula (4) can be expressed as: (6); like , polynomial Center front The coefficient of the term is zero, that is, , multiply both sides of formula (4) by Δ This can be simplified to: (7); Where, ; According to the prediction theory, in order to predict ahead of time j Step output, introduce the Diophantine equation: (8); In the formula, the polynomial E ( z -1 ), F ( z -1 ), G ( z -1 )satisfy: (9); Combining CARIMA expression with Diophantine equation, we can get the controlled object k + j Moment prediction equation: (10); Since the forecast of future output noise is unknown, the impact of future noise can be ignored, and the optimal prediction is: (11); 2) Scrolling optimization Variance of prediction error As the GPC objective function, the prediction model shown in formula (6) can be obtained. Minimum optimal prediction error: (12); Where: is the solution of the Diophantine equation; In the GPC rolling optimization, the following equation (10) is calculated online: E 、 F 、 G , we can further calculate ; Among them, the initial value of formula (8) is the same as formula (13); 、 and The formulas for the coefficient iteration process are: (13); (14); (15); (16); 3) Feedback correction: Recursive least squares method is used to timely correct the uncertainty of the prediction model output through feedback correction. The current output of the prediction model is: (17); Where, and satisfy: (18); The iterative formula of recursive least squares method is: (19); Where, λ For the forgetting factor.

[0013] In the preferred solution, the servo system model and the turbine model in S2 are specifically: The servo system contains nonlinear links such as delay, saturation, speed limit, and dead zone. The transfer function formula of its linear part is: (20); in: K y is the comprehensive amplifier coefficient; T y1 and T y are the reaction time constants of the intermediate servomotor and the main servomotor respectively; In the turbine model, a flow characteristic neural network model and a torque characteristic neural network model are constructed based on the neural network, with unit speed and guide vane opening as input and turbine flow and torque as output respectively.

[0014] In the preferred solution, the objective function formula used in S3 is: (twenty one); Where, f M is the turbine torque characteristic function constructed based on BP neural network, n 11 is the unit speed, u N is the steady-state control signal of the nonlinear turbine regulation system model, Y 0 is the initial guide vane opening, D is the runner inlet diameter, H is the water head, ω is the rotation speed, P c is the power disturbance value; Among them, the particles in the particle swarm optimization algorithm i The update speed and position formula is: (twenty two); (twenty three); Where, ; ; n represents the number of particles, K represents the dimension of space, represents the position of the particle, w Indicates that the inertia weight is a constant, and represents the learning factor, and Represented as a random number between 0 and 1 Represents particles i The best location, represents the best position of all particles i in the entire group.

[0015] In the preferred embodiment, the S4 is specifically: for the linear model, the steady-state control signal u L = p c / e y , combined with the determined u L and u N , we can get the nonlinear object control signal correction coefficient as a nonlinear characteristic compensator : (twenty four).

[0016] In the preferred solution, a GPC controller, a particle swarm optimization algorithm, a nonlinear sub-model and a linear sub-model are integrated to design an adaptive predictive controller for a hydropower unit based on generalized predictive control, wherein the nonlinear characteristic compensator corrects the control signal of the generalized predictive controller according to the output of the particle swarm optimization algorithm.

[0017] This invention provides a method for optimizing a hydropower unit's adaptive predictive controller based on PSO and GPC. This method constructs a nonlinear turbine model using a BP neural network and a nonlinear turbine regulation system model using the Octave tool. A particle swarm optimization algorithm is then used to obtain the steady-state control signal of the nonlinear turbine regulation system model. This signal is then corrected using a nonlinear characteristic compensator. The GPC, PSO, and compensator are then integrated to form an adaptive controller. The PSO algorithm dynamically optimizes the steady-state control signal, which is corrected using a nonlinear characteristic compensator. This significantly reduces control errors under nonlinear operating conditions. Multi-model collaborative control improves the accuracy of turbine dynamic characteristic fitting, shortens regulation time, reduces energy consumption, and extends the turbine's service life. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The present invention will be further described below with reference to the accompanying drawings and examples: Figure 1 It is a flow chart of the optimization method of the present invention; Figure 2 It is a schematic diagram of the simulation process of the water wheel regulating system model of the present invention; Figure 3 It is a flow chart of the particle swarm optimization algorithm of the present invention; Figure 4 It is a schematic diagram of the overall structure of the adaptive controller of the present invention. DETAILED DESCRIPTION

[0019] Example 1 like Figure 1-4 As shown, a method for optimizing a hydropower unit adaptive predictive controller based on PSO and GPC includes the following steps: S1: Construct a nonlinear turbine model based on BP neural network.

[0020] S2: Octave tool is used to construct a nonlinear water turbine regulation system model containing a nonlinear water turbine. The nonlinear water turbine regulation system model includes a water diversion system model, a water turbine model, a generator and power grid model, a speed governor and a servo system model.

[0021] S3: PSO is used to obtain the steady-state control signal of the nonlinear turbine regulation system model.

[0022] S4: Construct a nonlinear characteristic compensator for the adaptive predictive controller.

[0023] S5: Determine hyperparameters in the generalized predictive controller.

[0024] S6: Integrate the generalized predictive controller, particle swarm optimization algorithm and nonlinear characteristic compensator to obtain an adaptive controller.

[0025] In this embodiment, PSO is a particle swarm optimization algorithm (PSO), GPC is a generalized predictive control (GPC), BP neural network is a back propagation neural network (BP neural network), TCNN is a torque characteristic neural network model (TCNN), and DCNN is a discharge characteristic neural network model (DCNN).

[0026] like Figure 1 As shown in the figure, this embodiment constructs a nonlinear turbine model based on a BP neural network and a nonlinear turbine control system model using the Octave tool. Then, a particle swarm optimization algorithm is used to obtain the steady-state control signal of the nonlinear turbine control system model. This signal is corrected using a nonlinear characteristic compensator. Finally, GPC, PSO, and the compensator are integrated to form an adaptive controller. The PSO algorithm dynamically optimizes the steady-state control signal, and the nonlinear characteristic compensator performs corrections, significantly reducing control errors under nonlinear operating conditions. Multi-model collaborative control improves the accuracy of turbine dynamic characteristic fitting, shortens regulation time, reduces energy consumption, and extends the turbine's service life.

[0027] Step S1 specifically includes data collection and preparation, design of BP neural network structure, model training, and model verification and optimization.

[0028] In the preferred embodiment, step S1 further comprises the following steps: S11: Acquire turbine operation data and perform data preprocessing.

[0029] S12: Based on the BP neural network structure, a nonlinear turbine model is constructed. The BP neural network uses the ReLU activation function and the input includes the unit speed. n 11 and guide vane opening Y ,in: The number of nodes in the input layer and the output layer is equal to the number of corresponding input parameters and output parameters.

[0030] The number of hidden layers is 1, and the number of neurons can be determined according to the empirical formula: (1); Where: N i is the number of neurons in the input layer, N o is the number of neurons in the output layer, N h is the number of neurons in the hidden layer, Z A constant between 0 and 10.

[0031] S13: The nonlinear turbine model uses the data after data preprocessing to perform model training and save the optimal nonlinear turbine model.

[0032] In this embodiment, steps S11-S13 are described in detail as follows: 1) Data collection and preparation Data collection: Obtain turbine operating data (such as water head, water flow, guide vane opening, speed, power, etc.).

[0033] In the preferred solution, the data preprocessing includes: removing outliers, filling missing values ​​and normalization, specifically: Outlier detection is performed based on isolation forest, and a dynamic sliding window is used to repair missing values.

[0034] The operating intervals are divided according to the water head and guide vane opening, and each interval is independently standardized and normalized using the Sigmoid function.

[0035] 2) Design BP neural network structure Input layer: The number of nodes is equal to the number of input parameters.

[0036] Output layer: The number of nodes is equal to the number of output parameters.

[0037] Hidden layer: The number of layers is 1, the number of neurons is determined according to the empirical formula, and the ReLU activation function is selected.

[0038] 3) Model training, specifically: Initialization parameters: Randomly initialize weights and biases.

[0039] Forward propagation: Input data passes through the network to calculate the output.

[0040] Calculate loss: Use mean squared error (MSE).

[0041] Backpropagation: Calculate gradients and update parameters.

[0042] Training: Repeat the above steps until the loss converges.

[0043] 4) Model validation and optimization Cross-validation: Evaluate the generalization ability of the model.

[0044] Parameter tuning: Adjust the network structure and hyperparameters (such as learning rate).

[0045] 5) Model testing Model testing: Evaluate model performance using an independent test set.

[0046] In this embodiment, a numerical simulation platform for a hydraulic turbine regulating system with a nonlinear hydraulic turbine is constructed using the Octave tool. The simulation model includes a water diversion system model, a hydraulic turbine model, a generator and power grid model, a generalized predictive controller, and a servo system model. The hydraulic turbine is a core component, and its nonlinear dynamic characteristics determine the complexity of the system. Therefore, a nonlinear hydraulic turbine model is constructed based on a BP neural network to accurately simulate the dynamic behavior of the hydraulic turbine.

[0047] 1) Water diversion system model In the preferred solution, in step S2, the water diversion system model adopts a rigid water hammer model, which is mathematically described as: (2); Where, is the hydraulic time constant.

[0048] 2) Generator and grid model Generator and grid model Considering an infinite grid, the generator and load transfer functions are: (3); Where, T a is the unit inertia time constant, e g is the generator load self-regulation coefficient.

[0049] 3) GPC controller model The GPC algorithm comprises three fundamental elements: a predictive model, rolling optimization, and feedback correction. The core of this method lies in describing the system's dynamic response and optimization objectives through mathematical equations, thereby enabling accurate calculation of the control variable. By continuously updating the polynomials, the system achieves real-time optimal adjustment, better adapting to complex engineering environments and changing operating conditions.

[0050] In the preferred embodiment, the GPC controller in step S5 implements rolling optimization based on the CARIMA model and the Diophantine equation, and performs feedback correction, specifically: 3.1) Prediction model: For hydropower units, the GPC prediction function is based on the CARIMA model, and the formula is: (4) ; Where, 、 、 are the output signal, input signal and white noise of the system respectively; d is the system delay order; Δ =1- z -1 is the difference operator.

[0051] Polynomial A ( z -1 ), B ( z -1 )and C ( z -1 )satisfy: (5); Where, n a 、 n b and n c Polynomials A 、 B and C The order of .

[0052] like , formula (4) can be expressed as: (6).

[0053] like , polynomial Center front The coefficient of the term is zero, that is, , multiply both sides of formula (4) by Δ This can be simplified to: (7); Where, .

[0054] According to the prediction theory, in order to predict ahead of time j Step output, introduce the Diophantine equation: (8); In the formula, the polynomial E ( z -1 ), F (z -1 ), G ( z -1 )satisfy: (9); Combining CARIMA expression with Diophantine equation, we can get the controlled object k + j Moment prediction equation: (10).

[0055] Since the forecast of future output noise is unknown, the impact of future noise can be ignored, and the optimal prediction is: (11).

[0056] 3.2) Rolling Optimization Variance of prediction error As the GPC objective function, the prediction model shown in formula (6) can be obtained. Minimum optimal prediction error: (12); Where: is the solution of the Diophantine equation.

[0057] In the GPC rolling optimization, the following equation (10) is calculated online: E 、 F 、 G , we can further calculate ; Among them, the initial value of formula (8) is the same as formula (13); 、 and The formulas for the coefficient iteration process are: (13); (14); (15); (16).

[0058] 3.3) Feedback correction: Recursive least squares method is used to timely correct the uncertainty of the prediction model output through feedback correction. The current output of the prediction model is: (17); Where, and satisfy: (18).

[0059] The iterative formula of recursive least squares method is: (19); Where, λ For the forgetting factor.

[0060] like Figure 2 As shown in FIG, the simulation process of the water turbine regulation system model based on the final optimized GPC controller model.

[0061] In the preferred solution, the servo system model and the turbine model in step S2.

[0062] 4) Servo system model The servo system contains nonlinear links such as delay, saturation, speed limit, and dead zone. The transfer function formula of its linear part is: (20); in: K y is the comprehensive amplifier coefficient; T y1 and T y are the reaction time constants of the intermediate servomotor and the main servomotor, respectively.

[0063] 5) Turbine model To ensure the accuracy of the model, a DCNN (DCNN, Q 11 = Q 11 ( n 11 , Y )) and TCNN (TCNN, M 11 = M 11 ( n 11 , Y )).

[0064] In this embodiment, the control signal of the nonlinear turbine regulation system model in steady state is identified based on PSO.

[0065] In the preferred solution, the objective function formula used in step S3 is: (twenty one); Where, f M is the turbine torque characteristic function constructed based on BP neural network, n 11 is the unit speed, uN is the steady-state control signal of the nonlinear turbine regulation system model, Y 0 is the initial guide vane opening, D is the runner inlet diameter, H is the water head, ω is the rotation speed, P c is the power disturbance value.

[0066] like Figure 3 As shown in Figure 1, PSO is an optimization algorithm based on swarm intelligence. Its optimization concept is inspired by the collective behavior of swarming organisms such as birds and fish. The algorithm simulates the collaboration and information sharing of individuals in a swarm to find the optimal solution. The specific concept involves designing a swarm of particles with two properties: speed and position. These particles adjust their speed and position to search for the optimal solution. Each particle's speed and position are updated based on the guidance of the individual optimal solution and the global optimal solution.

[0067] Among them, the update speed and position formula of particle i in the particle swarm optimization algorithm is: (twenty two); (twenty three); Where, ; ; n represents the number of particles, K represents the dimension of space, represents the position of the particle, w represents the inertia weight as a constant, and represents the learning factor, and Represented as a random number between 0 and 1 represents the optimal position of particle i, represents the best position of all particles i in the entire group.

[0068] In the preferred solution, step S4 is specifically as follows: for the linear model, the steady-state control signal u L = p c / e y , combined with the determined u L and u N , we can get the nonlinear object control signal correction coefficient as a nonlinear characteristic compensator : (twenty four).

[0069] In this embodiment, under normal operating conditions, a sensitivity study of the generalized predictive controller to its hyperparameters is conducted, and the optimal hyperparameter values ​​are determined based on the study results.

[0070] Combining the above models, an adaptive predictive controller for hydropower units based on generalized predictive control is designed.

[0071] like Figure 4 The following is a system block diagram of a hydropower unit adaptive predictive controller based on particle swarm optimization (PSO) and generalized predictive control (GPC). The main modules and signal flows are as follows: Power fluctuation related: the power fluctuation value Input the PSO module, which calculates the steady-state control signal of the nonlinear turbine regulation system model based on a specific formula u N .

[0072] Control signal generation: Initial power and The difference divided by P r get p c ,Will p c Input GPC controller, combined with the nonlinear compensator output (involving calculation), generating control signals u ( k ).

[0073] System response path: The control signal acts on the nonlinear servo system, nonlinear turbine and linear water diversion system, affecting the generator output; at the same time, there are branches of the linear servo system, linear turbine and water diversion system, which receive real-time feedback of linearization-related parameters and are finally connected to the infinite power grid.

[0074] In the preferred solution, a GPC controller, a particle swarm optimization algorithm, a nonlinear sub-model and a linear sub-model are integrated to design an adaptive predictive controller for a hydropower unit based on generalized predictive control, wherein the nonlinear characteristic compensator corrects the control signal of the generalized predictive controller according to the output of the particle swarm optimization algorithm.

[0075] The design method of the adaptive predictive controller for a hydropower unit based on the particle swarm optimization algorithm and generalized predictive control in this embodiment can fully consider the impact of system input and output on the regulation quality of the turbine regulation system, and effectively ensure the safe and stable operation of the hydropower unit.

[0076] As the core equipment for power generation in hydropower stations, the safe and stable operation of hydropower units has become an increasingly important focus in the industry. This paper addresses the limitations of traditional PID controllers and advanced control methods, which often fail to adequately consider system inputs and outputs. By designing an adaptive predictive controller for hydropower units based on a particle swarm optimization algorithm and generalized predictive control, this paper proposes a method for designing an adaptive predictive controller for hydropower units.

[0077] Beneficial effects of this embodiment: 1) While ensuring the conventional regulation requirements of the hydropower unit, the designed adaptive predictive controller fully considers the system input and output, effectively ensuring the safe and stable operation of the hydropower unit.

[0078] 2) In view of the limitations of traditional PID controllers and advanced control methods in not taking system inputs and outputs into consideration, a design method for an adaptive predictive controller for hydropower units based on particle swarm optimization and generalized predictive control is proposed. This method comprehensively considers system inputs and outputs, allowing hydropower units to achieve more efficient and stable operation under different operating conditions, thereby extending equipment life, reducing maintenance costs, and improving power generation efficiency.

[0079] 3) The steady-state control signal is dynamically calculated through the PSO algorithm, and the GPC output is corrected in combination with the compensation coefficient, thereby reducing the control error under nonlinear conditions.

[0080] The above embodiments are merely preferred technical solutions of the present invention and should not be construed as limiting the present invention. The scope of protection of the present invention shall be the technical solutions set forth in the claims, including equivalent alternatives to the technical features of the technical solutions set forth in the claims. In other words, equivalent alternatives and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for optimizing a hydropower unit adaptive predictive controller based on PSO and GPC, characterized in that: The following steps are involved: S1: Constructing nonlinear turbine model based on BP neural network; S2: Octave tool is used to construct a nonlinear turbine regulation system model, which includes a water diversion system model, a turbine model, a generator and power grid model, a speed governor and a servo system model; S3: Using particle swarm optimization algorithm to obtain steady-state control signals of nonlinear turbine regulation system model; S4: Construct a nonlinear characteristic compensator for the adaptive predictive controller; S5: Determine hyperparameters in the generalized predictive controller; S6: Integrate the generalized predictive controller, particle swarm optimization algorithm and compensator to obtain an adaptive controller.

2. The method for optimizing the adaptive predictive controller of a hydropower unit based on PSO and GPC according to claim 1 is characterized in that: Said S1 further comprises the following steps: S11: Acquire turbine operation data and perform data preprocessing; S12: Based on the BP neural network structure, a nonlinear turbine model is constructed. The BP neural network uses the ReLU activation function and the input includes the unit speed. n 11 and guide vane opening Y ,in: The number of nodes in the input layer and the output layer is equal to the number of corresponding input parameters and output parameters; The number of hidden layers is 1, and the number of neurons can be determined according to the empirical formula: (1)； Where: N i is the number of neurons in the input layer, N o is the number of neurons in the output layer, N h is the number of neurons in the hidden layer, Z is a constant between 0 and 10; S13: The nonlinear turbine model uses the preprocessed data to perform model training and save the optimal nonlinear turbine model.

3. The method for optimizing the adaptive predictive controller of a hydropower unit based on PSO and GPC according to claim 2 is characterized in that: Model training is specifically as follows: Using forward propagation, that is, the input data is calculated through the network output; The loss function uses mean square error and performs back propagation to optimize the model; Repeat the above steps until the loss converges.

4. The method for optimizing the adaptive predictive controller of a hydropower unit based on PSO and GPC according to claim 2 is characterized in that: The data preprocessing includes: removing outliers, filling missing values ​​and normalization, specifically: Based on isolation forest, outlier detection is performed and missing values ​​are repaired using dynamic sliding windows; The operating intervals are divided according to the water head and guide vane opening, and each interval is independently standardized and normalized using the Sigmoid function.

5. The method for optimizing the adaptive predictive controller of a hydropower unit based on PSO and GPC according to claim 1 is characterized in that: In S2, the water diversion system model adopts a rigid water hammer model, which is mathematically described as: (2)； Where, is the hydraulic time constant; The generator and grid model considers an infinite grid, and the generator and load transfer functions are: (3)； Where, T a is the unit inertia time constant, e g is the generator load self-regulation coefficient.

6. The method for optimizing the adaptive predictive controller of a hydropower unit based on PSO and GPC according to claim 1 is characterized in that: The generalized predictive controller in S5 implements rolling optimization based on the CARIMA model and the Diophantine equation, and performs feedback correction, specifically: 1) Forecasting model: Based on the CARIMA model, the formula is: (4) ； Where, 、 、 are the output signal, input signal and white noise of the system respectively; d is the system delay order; Δ =1- z -1 is the difference operator; Polynomial A ( z -1 ), B ( z -1 )and C ( z -1 )satisfy: (5)； Where, n a 、 n b and n c Polynomials A 、 B and C The order of like , formula (4) can be expressed as: (6) ； like , polynomial Center front The coefficient of the term is zero, that is, , multiply both sides of formula (4) by Δ This can be simplified to: (7)； Where, ; According to the prediction theory, in order to predict ahead of time j Step output, introduce the Diophantine equation: (8)； In the formula, the polynomial E ( z -1 ), F ( z -1 ), G ( z -1 )satisfy: (9)； Combining CARIMA expression with Diophantine equation, we can get the controlled object k + j Moment prediction equation: (10)； Since the forecast of future output noise is unknown, the impact of future noise can be ignored, and the optimal prediction is: (11)； 2) Scrolling optimization Variance of prediction error As the GPC objective function, the prediction model shown in formula (6) can be obtained. Minimum optimal prediction error: (12)； Where: is the solution of the Diophantine equation; In the GPC rolling optimization, the following equation (10) is calculated online: E 、 F 、 G , we can further calculate ; Among them, the initial value of formula (8) is the same as formula (13); 、 and The formulas for the coefficient iteration process are: (13)； (14)； (15)； (16)； 3) Feedback correction: Recursive least squares method is used to timely correct the uncertainty of the prediction model output through feedback correction. The current output of the prediction model is: (17)； Where, and satisfy: (18)； The iterative formula of recursive least squares method is: (19)； Where, λ For the forgetting factor.

7. The method for optimizing the adaptive predictive controller of a hydropower unit based on PSO and GPC according to claim 1, characterized in that: The servo system model and turbine model in S2 are specifically: The servo system contains nonlinear links such as delay, saturation, speed limit, and dead zone. The transfer function formula of its linear part is: (20)； in: K y is the comprehensive amplifier coefficient; T y1 and T y are the reaction time constants of the intermediate servomotor and the main servomotor respectively; In the turbine model, a flow characteristic neural network model and a torque characteristic neural network model are constructed based on the neural network, with unit speed and guide vane opening as input and turbine flow and torque as output respectively.

8. The method for optimizing the adaptive predictive controller of a hydropower unit based on PSO and GPC according to claim 1 is characterized in that: The objective function formula used in S3 is: (21)； Where, f M is the turbine torque characteristic function constructed based on BP neural network, n 11 is the unit speed, u N is the steady-state control signal of the nonlinear turbine regulation system model, Y 0 is the initial guide vane opening, D is the runner inlet diameter, H is the water head, ω is the rotation speed, P c is the power disturbance value; Among them, the update speed and position formula of particle i in the particle swarm optimization algorithm is: (22)； (23)； Where, ; ; n represents the number of particles, K represents the dimension of space, represents the position of the particle, w represents the inertia weight as a constant, and represents the learning factor, and Represented as a random number between 0 and 1 represents the optimal position of particle i, represents the best position of all particles i in the entire group.

9. The method for optimizing the adaptive predictive controller of a hydropower unit based on PSO and GPC according to claim 1, characterized in that: The S4 is specifically: For the linear model, the steady-state control signal u L = p c / e y , combined with the determined u L and u N , we can get the nonlinear object control signal correction coefficient as a nonlinear characteristic compensator : (24)。 10. The method for optimizing the adaptive predictive controller of a hydropower unit based on PSO and GPC according to claim 1, characterized in that: An adaptive predictive controller for hydropower units based on generalized predictive control is designed by integrating the GPC controller, particle swarm optimization algorithm, nonlinear sub-model and linear sub-model. The nonlinear characteristic compensator modifies the control signal of the generalized predictive controller according to the output of the particle swarm optimization algorithm.

Images