A pinn-based pid control optimization method for water turbine regulating system
By combining a physical information neural network (PINN) with a PID controller in a turbine regulating system, a multi-state mechanism model and data-driven joint training are constructed, solving the adaptability and reliability problems of traditional PID controllers under complex operating conditions, and realizing high-precision and stable intelligent control.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTHWEST A & F UNIV
- Filing Date
- 2026-04-16
- Publication Date
- 2026-06-05
AI Technical Summary
Traditional PID controllers rely on engineering experience for tuning in hydro turbine regulation systems, making them difficult to adapt to complex operating conditions. Existing intelligent algorithms ignore physical laws, resulting in insufficient generalization ability and reliability, and the models rely on high-quality data acquisition, which is costly.
By employing a Physical Information Neural Network (PINN) in collaboration with a PID controller, a closed-loop intelligent control system is constructed through the joint training of a multi-state mechanism model and data-driven approaches, combined with physical constraints and data fitting. This system includes a bypass verification loop to monitor the model's accuracy and health status.
It significantly improves the control accuracy and reliability of the turbine regulation system under nonlinear and time-varying conditions, realizes active sensing and adaptive regulation, and enhances the intelligence level of the hydropower station.
Smart Images
Figure CN122148480A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of water conservancy regulation, specifically relating to a PID control optimization method for a turbine regulation system based on PINN. Background Technology
[0002] The turbine regulating system is the core control system of a hydropower station, primarily responsible for maintaining stable unit speed and responding quickly to load changes. Its control performance directly affects grid frequency quality, power station operational stability, and equipment safety. In this field, the proportional-integral-derivative (PID) controller is widely used due to its simple structure, high reliability, and clear physical meaning of parameters, making it the standard control configuration for turbine regulating systems.
[0003] However, the performance of traditional PID controllers is highly dependent on the precise tuning of Kp, Ki, and Kd, relying heavily on engineering experience or offline simulation. This makes them ill-suited for dynamic adjustment under complex operating conditions, and on-site parameter tuning requires significant manpower, resources, and time. Meanwhile, existing intelligent algorithms such as fuzzy neural networks, particle swarm optimization, and weed invasion algorithms, while not requiring precise mechanistic models and capable of learning system dynamics from historical operating data, are mostly black-box optimizations. They treat the system as a purely data mapping, ignoring the inherent physical laws and mechanisms of the turbine system. This neglect of physics leads to a lack of physically consistent explanations for their predictive behavior, making their generalization ability and reliability difficult to guarantee under operating conditions outside the training data coverage. Furthermore, the accuracy of the model heavily depends on massive amounts of high-quality training data, which is often costly to obtain or contains noise in real-world industrial scenarios. Moreover, ignoring known fundamental physical laws of the system, such as energy conservation and momentum theorem, may result in predictions that violate basic physical principles, posing risks in practical applications.
[0004] Physical Information Neural Networks (PINNs), as an emerging fusion modeling method, successfully combines the interpretability of mechanistic models with the flexibility of data-driven models by embedding physical laws such as control equations and boundary conditions as constraints into the training process of the neural network. Currently, PINN technology has shown great potential in solving forward and inverse problems in fields such as fluid mechanics and structural mechanics; however, research on its application in the control optimization of turbine regulating systems, especially in its synergistic application with the traditional mainstream industrial controller PID, remains largely unexplored. Summary of the Invention
[0005] The purpose of this invention is to propose a PID control optimization method for a turbine regulating system based on PINN, in order to solve the problems in the prior art.
[0006] Therefore, this invention provides a PID control optimization method for a turbine regulating system based on PINN, comprising: Construct and train a physical information neural network until the physical information neural network converges; The trained physical information neural network is deployed to the online inference system, and a PID controller is equipped as the main control loop of the turbine regulation system. The PID controller outputs a real control signal and performs state evolution based on a multi-state mechanism model, and outputs a real-time state feedback signal. The real-time state feedback signal is input into the physical information neural network for state recursion prediction and the predicted state feedback signal is output. The real-time state feedback signal and the predicted state feedback signal are continuously compared to evaluate the accuracy and reliability of the physical information neural network under the current operating conditions, and the evaluation results are output. Based on the evaluation results, a bypass verification loop is constructed, which together with the main control loop forms a closed-loop intelligent control system.
[0007] In some embodiments, constructing and training the physical information neural network includes: A multi-state mechanism model is constructed, and a numerical solution algorithm is used to solve the multi-state mechanism model to obtain a high-fidelity physical simulation dataset; A physical information neural network is constructed based on an artificial neural network framework. The state variables and control signals of the turbine regulation system are used as the input units of the physical information neural network, and the state prediction values of the turbine regulation system are used as the output units of the physical information neural network. The physical information neural network is jointly trained using the multi-state mechanism model as physical driving information and the high-fidelity physical simulation dataset as data driving information. Based on the physical driving information and the data driving information, a multi-objective loss function for physical information is constructed. The physical information neural network is trained using the high-fidelity physical simulation dataset as training data. The physical information multi-objective loss function is used to ensure training accuracy and optimize the physical information neural network until the model converges.
[0008] In some embodiments, the multi-state mechanism equation model includes: The rotational speed dynamic equation is as follows: ; The guide vane servo equation is as follows: ; The flow dynamics equation is as follows: ;as well as The head dynamics equation is as follows: ; in, As an external input, no physical constraints are imposed; The system time constant; and Indicates the turbine characteristic coefficient; The damping coefficient; Rated head; This is the head loss coefficient; For reference traffic; For speed, For guide vane opening, For traffic, For water head, This represents the load torque.
[0009] In some embodiments, the input unit is a 7-dimensional vector [dt,n0,y0,q0,h0, The output unit is a five-dimensional vector [n, y, q, h, 0, u], The neural network includes a feedforward neural network structure with multiple hidden layers, and the activation function of the hidden layers is the tanh function.
[0010] In some embodiments, the physical information multi-objective loss function is: ; in This is the data loss term, used to ensure that the model output fits the training data. This is a physical loss term, used to force the model output to satisfy the physical equations of the turbine regulation system. and These are the corresponding data loss hyperparameters and physical loss hyperparameters.
[0011] In some embodiments, the physical loss item Including the sum of squared residuals from the multi-state mechanism equations of the turbine regulating system, specifically: ; Where F1, F2, F3, and F4 are the residuals of the dynamic equations for rotational speed, guide vane, flow rate, and head, respectively; N is the number of training data.
[0012] In some embodiments, the data loss item This includes the mean squared error between the neural network prediction and the target value in the high-fidelity physical simulation dataset, specifically:
[0013] in, The predicted value output by the neural network. For the target value in the simulation dataset, This represents the number of training data points.
[0014] Understandably, the purely data-driven neural networks commonly used in existing technologies have weak interpretability and are often regarded as black boxes. However, PINN in this application introduces physical equation constraints, making the prediction results more interpretable in terms of physical consistency and easier to diagnose the source of errors.
[0015] In some embodiments, the PID controller prevents the integral term from becoming too large by using anti-integral saturation, suppresses high-frequency noise by using derivative filtering, and ensures output stability by using amplitude limiting and rate of change limiting.
[0016] In some embodiments, when the physical information neural network performs state recursive prediction, the input is the current real state and control signal, and the output is a prediction of the state of the turbine regulation system at the next moment, and it is executed cyclically within a preset control cycle.
[0017] In some embodiments, the bypass verification loop is used to provide model inaccuracy warnings and health status information to trigger parameter self-tuning of the PID controller or online optimization of the control strategy.
[0018] Beneficial effects: 1. This invention adopts an innovative architecture that combines physical information neural networks with traditional PID controllers. It maintains the advantages of PID control in engineering practice, such as simple structure and high reliability, while improving the control quality of the system through intelligent prediction. This significantly improves the control accuracy and operational reliability of the turbine regulating system under nonlinear and time-varying conditions, providing an effective technical solution for the intelligent upgrading of hydropower stations.
[0019] 2. This invention embeds the multi-state mechanism equations of a water turbine into a neural network for training. The PINN model, while ensuring data fitting accuracy, strictly follows the physical laws of the system, significantly enhancing the generalization ability and prediction reliability under unknown operating conditions and small sample conditions. It effectively solves the problems of physical inconsistency and weak generalization of traditional black-box models in complex dynamic systems.
[0020] 3. This invention continuously monitors model accuracy and system health status through a PINN-based bypass verification loop. Once model inaccuracy or performance degradation is detected, it can trigger PID parameter self-tuning or control strategy optimization, achieving a leap from passive control to active sensing and adaptive regulation.
[0021] 4. Through the deep integration of PINN and PID, the control accuracy and operational stability of the turbine regulation system under complex operating conditions such as nonlinearity and time-varying loads are significantly improved, helping hydropower stations to transform and upgrade towards digitalization and intelligence. Attached Figure Description
[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0023] Figure 1 This is a flowchart illustrating the PID control optimization method for a turbine regulating system based on PINN provided by the present invention.
[0024] Figure 2 This is a flowchart illustrating the construction of a physical information neural network in the PID control optimization method for a turbine regulating system based on PINN provided by the present invention.
[0025] Figure 3 This is a schematic diagram of the PINN training and validation loss curves and loss ratio analysis of the present invention.
[0026] Figure 4 This is a schematic diagram of the closed-loop simulation and PINN prediction error analysis of the multi-state high-precision turbine model of the present invention.
[0027] Figure 5 This is a framework diagram of the PINN of the present invention.
[0028] Figure 6 For the present invention Figure 4 Adaptive variation of PID parameters under various state changes. Detailed Implementation
[0029] The invention will be more readily understood by referring to the following detailed description of preferred embodiments and included examples. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. In case of conflict, the definitions in this specification shall prevail.
[0030] like Figure 1-6As shown, a PID control optimization method for a turbine regulating system based on PINN includes: Construct and train a physical information neural network until it converges. The construction and training of the physical information neural network includes the following steps: A multi-state mechanism model was constructed and solved using a numerical algorithm to obtain a high-fidelity physical simulation dataset. Specifically, the multi-state mechanism model is a five-state model, with the five states including rotational speed, guide vane opening, flow rate, head, and load torque. Based on the structural parameters of the turbine regulating system, a fourth / fifth-order embedded Runge-Kutta method was used to solve the five-state mechanism equations, generating the high-fidelity physical simulation dataset.
[0031] The constructed 5-state mechanism equation model of the hydro-turbine regulation system includes the following dynamic equations: Rotational speed dynamics equation: ; Guide vane servo equation: ; Flow dynamics equation: ; Hydrodynamic equations: ; in, As an external input, no physical constraints are imposed; The system time constant; and Indicates the turbine characteristic coefficient; The damping coefficient; Rated head; This is the head loss coefficient; For reference traffic; For speed, For guide vane opening, For traffic, For water head, The load torque is used. Based on the structural parameters of the turbine regulating system, the mechanistic equations are solved using the fourth / fifth-order embedded Runge-Kutta method (RK45), generating a high-fidelity physical simulation dataset.
[0032] Then, a physical information neural network (PINN) is constructed based on an artificial neural network framework. The state variables and control signals of the turbine regulating system are used as input units, and the predicted state values of the turbine regulating system are used as output units. For example, the neural network (NN) part of the PINN is constructed using an artificial neural network (ANN) built with TensorFlow / Keras, thus confirming the input and output units of the neural network. The input unit of the neural network is a 7-dimensional vector [dt,n0,y0,q0,h0, [0,u], where dt is the time step, n0, y0, q0, h0, 0 represents the current system state variable, and u represents the control signal. The output unit is a 5-dimensional vector [n, y, q, h, ... ] represents the predicted system state at the next time step. The neural network part adopts a feedforward neural network structure with multiple hidden layers, and the activation function of the hidden layers is the tanh function.
[0033] A physical information neural network is jointly trained using a multi-state mechanism model as the physical driving information and a high-fidelity physical simulation dataset as the data driving information. For example, the high-fidelity physical simulation dataset is divided into a training set and a validation set in an 8:2 ratio. The training set is used for model parameter updates, while the validation set is used to monitor the training process and implement an early stopping mechanism. The early stopping mechanism monitors the total multi-objective loss of the validation set. For example, a tolerance value of 100 iterations and a tolerance of 0.0001 are set as trigger conditions. During training, the model weights with the minimum validation set loss are continuously saved. When the validation set loss does not decrease for 100 consecutive iterations and the decrease does not exceed the tolerance, training is immediately terminated and the optimal weights are loaded. This avoids model overfitting, balances training efficiency and generalization ability, and ensures that the model simultaneously satisfies the data fitting accuracy and the physical equation constraints of the turbine regulating system.
[0034] Simultaneously, based on the aforementioned physical-driven and data-driven information, a multi-objective loss function for physical information is constructed. The formula is: ; in This is the data loss term, used to ensure that the model output fits the training data. This is a physical loss term used to force the model output to satisfy the physical equations of the turbine regulation system. and These are the corresponding data loss hyperparameters and physical loss hyperparameters. The performance of the Physical Information Neural Network (PINN) is optimized by comprehensively utilizing a multi-objective loss function, and the PINN model in this application is verified through over 800 iterations to demonstrate its error-free operation. In some embodiments, =1, =0.1.
[0035] It should be noted that the physical loss item It consists of the sum of squared residuals from the 5-state mechanism equations of the turbine regulating system, defined as follows: ; in, , , and The residuals of the dynamic equations for rotational speed, guide vane, flow rate, and head are represented in sequence. This represents the number of training data points.
[0036] Data loss items The mean squared error between the neural network prediction and the target value in the high-fidelity physical simulation dataset is defined as follows:
[0037] in, The predicted value output by the neural network. For the target value in the simulation dataset, This represents the number of training data points.
[0038] Then, a high-fidelity physical simulation dataset is used as training data to train the physical information neural network. The physical information multi-objective loss function is used to ensure training accuracy, so that the loss function gradually decreases and finally converges to obtain a well-trained PINN model with fixed weights, realizing a complete training loop, validation and early stopping.
[0039] The trained physical information neural network is deployed to the online inference system, and a PID controller is equipped as the main control loop of the turbine regulation system. The PID controller is an enhanced PID, which can achieve anti-integral saturation, differential filtering, output limiting, and rate of change limitation. By using the clamping anti-saturation method, the integral term can be prevented from being too large. By using a first-order low-pass filter with a time constant set to 0.2 seconds, high-frequency noise can be suppressed by differential filtering. The limiting and rate of change limitation can ensure the stability of the control output. Specifically, the output limiting ensures that the control output is within the range of [0.0, 1.0], and the rate of change limitation ensures that the guide vane change rate does not exceed 0.05 pu / s.
[0040] The PID controller outputs a real control signal and performs state evolution based on a multi-state mechanism model, and outputs a real-time state feedback signal. The physical model (real system) of the turbine regulating system performs state evolution according to the control signal u generated by the enhanced PID controller, and outputs a real-time state feedback signal.
[0041] The physical model of the turbine regulation system performs state evolution based on the control signal u generated by the enhanced PID controller. The RK45 algorithm, i.e., the fourth / fifth order embedded Runge-Kutta method, is used to solve the system dynamic equations in real time and output the real-time state feedback signal [n,y,q,h,Mg].
[0042] The real-time state data, i.e., the real-time state feedback signal, output above is input into the physical information neural network for state recursion prediction and the predicted state feedback signal is output; specifically, the current state variables [n_current, y_current, q_current, h_current, ... The current and control signal u are used as inputs to the PINN model to obtain the state prediction value [n_pinn, y_pinn, q_pinn, h_pinn, ...] at the next time step. _pinn].
[0043] The system continuously compares the real-time status feedback signal with the predicted status feedback signal to evaluate the accuracy and reliability of the physical information neural network under the current operating conditions and outputs the evaluation results. Based on the evaluation results, a bypass verification loop is constructed. Although this loop does not directly participate in control, it provides crucial information for system health monitoring, model inaccuracy warning, and controller optimization. Together with the PID main control loop in step 8, it forms a closed-loop intelligent control system. Furthermore, the bypass verification loop and the main control loop together constitute a closed-loop intelligent control system.
[0044] like Figure 3-4 In one embodiment provided in this application, the turbine regulating system is based on a five-state high-precision physical model that includes elastic water hammer, dynamic head, and nonlinear turbine characteristics, and a high-fidelity simulation dataset is generated using a fourth / fifth-order embedded Runge-Kutta algorithm. The system initially operates under rated steady-state conditions, with initial values for each state set as follows: rotational speed n=1.0, guide vane opening y, flow rate q, head h, and load torque. They are all at their respective equilibrium points.
[0045] To simulate a common load change scenario in a real power grid, a negative load step disturbance is applied to the system at simulation time t=50s. The core objective of the control system is achieved by an enhanced PID controller, which rapidly adjusts the guide vane opening to drive the rotational speed n to track its target reference value. Figure 4 Medium-speed n-target value curve to maintain grid frequency stability.
[0046] Under load step perturbation, the deployed trained PINN model is recursively predicted online, and its output is continuously compared with the output of a high-precision physical model representing the real system. Figure 3It clearly demonstrates the dynamic response process of total loss, physical loss, and data loss.
[0047] from Figure 4 As can be seen, the PINN prediction curve (dashed line) and the physical model curve (solid line) highly coincide throughout the entire transient response of each state variable. In particular, the PINN model can accurately predict the nonlinear dynamic behavior of the system during the dynamic process after the disturbance occurs from t=50s to t=100s and during the process of entering a new steady state.
[0048] For the core controlled variable, rotational speed *n*, the PINN prediction not only accurately reproduced the overshoot and settling time of the physical model, but also maintained the same tracking trend as the target value throughout the transition process. For the head *h* and flow rate *q*, which exhibit significant nonlinear characteristics, the PINN model also demonstrated excellent prediction accuracy, proving that it successfully learned the complex intrinsic dynamics of the system through physical constraints.
[0049] The results show that the PINN model trained by this invention successfully integrates the physical mechanism and data patterns of the turbine regulation system, enabling it to maintain physical consistency and prediction reliability even under disturbance conditions not fully covered by the training data, thus overcoming the weakness of pure data-driven models in terms of generalization ability.
[0050] This embodiment verifies the practical value of the dual-loop architecture for main control and verification proposed in this invention. In this simulation, the main control loop, namely the enhanced PID controller, independently and reliably completed the speed regulation task.
[0051] In the bypass verification loop, the PINN model, acting as a parallel digital twin, provides high-precision system state predictions in real time. For example... Figure 3 As shown, after more than 800 iterations, the total loss, data loss, and physical loss all converge, indicating that the model is running correctly under the current conditions.
[0052] This embodiment, through load step disturbance testing, fully demonstrates the effectiveness of the strategy combining Physical Information Neural Network (PINN) with PID control. This strategy not only ensures the stability and control quality of the control system under complex operating conditions, but also provides a solid data foundation and decision support for state monitoring, early warning, and intelligent optimization control of the turbine regulating system through the bypass verification loop constructed by PINN, significantly improving the overall intelligence level and operational reliability of the system.
[0053] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A PID control optimization method for a turbine regulating system based on PINN, characterized in that, include: Construct and train a physical information neural network until the physical information neural network converges; The trained physical information neural network is deployed to the online inference system, and a PID controller is equipped as the main control loop of the turbine regulation system. The PID controller outputs a real control signal and performs state evolution based on a multi-state mechanism model, and outputs a real-time state feedback signal. The real-time state feedback signal is input into the physical information neural network for state recursion prediction and the predicted state feedback signal is output. The real-time state feedback signal and the predicted state feedback signal are continuously compared to evaluate the accuracy and reliability of the physical information neural network under the current operating conditions, and the evaluation results are output. Based on the evaluation results, a bypass verification loop is constructed, which together with the main control loop forms a closed-loop intelligent control system.
2. The PID control optimization method for a turbine regulating system based on PINN according to claim 1, characterized in that, The construction and training of the physical information neural network includes: A multi-state mechanism model is constructed, and a numerical solution algorithm is used to solve the multi-state mechanism model to obtain a high-fidelity physical simulation dataset; A physical information neural network is constructed based on an artificial neural network framework. The state variables and control signals of the turbine regulation system are used as the input units of the physical information neural network, and the state prediction values of the turbine regulation system are used as the output units of the physical information neural network. The physical information neural network is jointly trained using the multi-state mechanism model as physical driving information and the high-fidelity physical simulation dataset as data driving information. Based on the physical driving information and the data driving information, a multi-objective loss function for physical information is constructed. The physical information neural network is trained using the high-fidelity physical simulation dataset as training data. The physical information multi-objective loss function is used to ensure training accuracy and optimize the physical information neural network until the model converges.
3. The PID control optimization method for a turbine regulating system based on PINN according to claim 2, characterized in that, The multi-state mechanism model includes: The rotational speed dynamic equation is as follows: ; The guide vane servo equation is as follows: ； The flow dynamics equation is as follows: ;as well as The head dynamics equation is as follows: ; in, As an external input, no physical constraints are imposed; The system time constant; and Indicates the turbine characteristic coefficient; The damping coefficient; Rated head; This is the head loss coefficient; For reference traffic; For speed, For guide vane opening, For traffic, For water head, This represents the load torque.
4. The PID control optimization method for a turbine regulating system based on PINN according to claim 2, characterized in that, The input unit is a 7-dimensional vector [dt,n0,y0,q0,h0, The output unit is a five-dimensional vector [n, y, q, h, 0, u], The neural network includes a feedforward neural network structure with multiple hidden layers, and the activation function of the hidden layers is the tanh function.
5. The PID control optimization method for a turbine regulating system based on PINN according to claim 3, characterized in that, The multi-objective loss function for the physical information is: ; in This is the data loss term, used to ensure that the model output fits the training data. This is a physical loss term, used to force the model output to satisfy the physical equations of the turbine regulating system. and These are the corresponding data loss hyperparameters and physical loss hyperparameters.
6. The PID control optimization method for a turbine regulating system based on PINN according to claim 5, characterized in that, The physical loss item Including the sum of squared residuals from the multi-state mechanism equations of the turbine regulating system, specifically: ; Where F1, F2, F3, and F4 are the residuals of the dynamic equations for rotational speed, guide vane, flow rate, and head, respectively; N is the number of training data.
7. The PID control optimization method for a turbine regulating system based on PINN according to claim 5, characterized in that, The data loss item This includes the mean square error between the neural network prediction and the target value in the high-fidelity physical simulation dataset, specifically: in, The predicted value output by the neural network. For the target value in the simulation dataset, This represents the number of training data points.
8. The PID control optimization method for a turbine regulating system based on PINN according to claim 1, characterized in that, The PID controller prevents the integral term from becoming too large by using anti-integral saturation, suppresses high-frequency noise by using derivative filtering, and ensures output stability by using amplitude limiting and rate of change limiting.
9. The PID control optimization method for a turbine regulating system based on PINN according to claim 1, characterized in that, When performing state recursive prediction, the physical information neural network takes the current real state and control signal as input and outputs a prediction of the state of the turbine regulation system at the next moment, and executes it cyclically within a preset control cycle.
10. The PID control optimization method for a turbine regulating system based on PINN according to claim 1, characterized in that, The bypass verification loop is used to provide early warning of model inaccuracies and health status information to trigger parameter self-tuning of the PID controller or online optimization of the control strategy.