Floating wind wave power system control method based on knowledge distillation and partition strategy

Abstract

The present application belongs to the technical field of marine renewable energy power generation, and particularly relates to a floating wind and wave power generation system control method based on knowledge distillation and partition strategy, comprising: data acquisition and processing; multi-teacher model system construction; knowledge distillation and student model training; partition control strategy design; objective function design; weight adaptive mechanism design; constraint condition definition; upper layer optimization solution; lower layer control instruction execution; dynamic feedback correction mechanism. The method covers all operation and survival conditions of the combined power generation system, on the one hand, reduces the modeling difficulty and complexity of the control model, improves the real-time performance and reliability of the prediction and control, and on the other hand, from the perspective of the engineering application of the combined system, combined with the wind and wave conditions, a partition control strategy is proposed, thereby enhancing the energy utilization efficiency and operation stability of the floating wind and wave combined power generation system in complex sea conditions.

Description

Technical Field

[0001] This invention relates to the field of marine renewable energy power generation technology, specifically to a control method for a floating wind and wave power generation system based on knowledge distillation and zoning strategies. Background Technology

[0002] Floating wind and wave combined power generation systems operate in the complex marine environment of deep seas, where they are subjected to the coupling effects of multiple disturbances such as wind, waves, and currents. This results in strong nonlinear characteristics in their aerodynamic loads, platform motion, and structural response. Although the concept of floating wind and wave combined power generation systems has been proposed for many years, existing research has mostly focused on the individual control of floating wind turbines and wave energy generation devices, and a mature joint control framework has not yet been formed. This has become a major challenge restricting the engineering application of floating wind and wave combined power generation systems. For example, floating wind turbines typically rely on pitch control, speed control, yaw control, and structural control, while wave energy devices mainly rely on PTO (Power Take-Off) systems for wave energy capture and power generation.

[0003] In the control of floating wind turbines, existing research often relies on linearized modeling of the turbine's aerodynamics and structural dynamics. Simplified models are built near the rated wind speed, and proportional-integral (PI) controllers or optimal controllers are designed accordingly. PI control is widely used for pitch and speed regulation due to its simplicity and ease of implementation, but it suffers from response lag and insufficient control accuracy under severe wind fluctuations. Optimal control methods, such as linear quadratic regulation (LQR), H2, and H∞ control, can optimize load suppression and power balance to some extent, but their performance is highly dependent on the accuracy of the linearized model. The control effect significantly decreases when platform motion is significant or wind-wave coupling effects intensify. With the development of data-driven methods, model predictive control (MPC) has gradually become a research hotspot in floating wind turbine control. MPC can comprehensively consider system dynamic constraints and future disturbance trends in the prediction time domain and achieve real-time updates of control inputs through rolling optimization. Unlike traditional analytical model-based methods, data-driven MPC no longer relies on precise linearized models. Instead, it uses operational data or identification models to capture the nonlinear characteristics of the system, thereby maintaining high robustness and adaptability under complex sea conditions. In contrast, the PTO control strategies of wave energy devices mainly include passive control and reactive power control: the former achieves energy absorption by adjusting the forces related to PTO, with common methods including linear damping, lockout, and clutch control; the latter sacrifices some power output in exchange for power generation stability.

[0004] 1) Patent application number 202410615877.5 discloses a wind and wave combined power generation platform device and control method based on TMD vibration reduction. By using the wave energy power generation device as the mass component for TMD vibration reduction of the entire platform, vibration control of the platform under normal power generation conditions can be achieved without introducing external mass, thereby realizing vibration control of the overall structure. However, passive control technology based on TMD vibration reduction is difficult to respond to changes in the wind and wave environment in a timely manner.

[0005] 2) For example, patent application number 202510912109.0 discloses a model predictive control method for a floating wind-wave combined power generation system. This method reduces tower load and suppresses platform pitch motion through coordinated control of the floating wind turbine pitch control and the wave energy generator's PTO (Power Tolerance) system. Traditional control methods based on analytical models are difficult to model. When a floating wind-wave combined power generation system operates under complex sea conditions, it requires complex formula derivations to establish an accurate analytical model. However, due to the highly nonlinear nature of the system dynamics, modeling is extremely difficult, requiring hundreds or even thousands of linearized models to cover the operating range under different wind speeds and wave conditions.

[0006] 3) For example, patent application number 202411395046.8 discloses a data-driven model-free control method and system for floating wind turbines. It utilizes deep reinforcement learning to train the controller, achieving dynamic correction and real-time control, effectively reducing modeling difficulty and improving power generation efficiency and system stability. When existing data-driven model predictive control is applied to complex floating wind-wave combined power generation systems, traditional data-driven model predictive control often requires a large number of variables and parameters for training, encountering problems such as high computational load and cost in the MPC solution process, making real-time control difficult to achieve.

[0007] Research on joint control methods for floating wind-wave combined power generation systems is still in its early stages. Most existing control methods for wind-wave combined power generation systems only consider wave energy control under single or specific operating conditions, and do not systematically and comprehensively consider the operating and survival conditions of wind-wave combined power generation systems, lacking a systematic consideration of the coupling effect between wind energy and wave energy. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art by providing a control method for floating wind and wave power generation systems based on knowledge distillation and a zoning strategy, covering all operating and survival conditions of the combined power generation system. This method reduces the modeling difficulty and complexity of the control model, improving the real-time performance and reliability of prediction and control. Furthermore, from the perspective of combined system engineering applications, it proposes a zoning control strategy based on wind and wave conditions, thereby enhancing the energy utilization efficiency and operational stability of the floating wind and wave combined power generation system under complex sea conditions.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] A control method for a floating wind and wave power generation system based on knowledge distillation and zoning strategies includes the following steps:

[0011] (1) Data acquisition and processing

[0012] Full-condition operation data of the floating wind and wave combined power generation system were obtained through sensors and high-fidelity models; the raw data were preprocessed by normalization, noise reduction and feature extraction; training set and test set were divided to provide a high-quality data foundation for subsequent modeling;

[0013] (2) Construction of a multi-teacher model system

[0014] A multi-teacher model system consisting of Transformer, LSTM, and PINN is constructed; each teacher model focuses on different feature dimensions such as trend, cycle, and physical constraint; high-precision prediction results and diverse knowledge representations are obtained through independent training.

[0015] (3) Knowledge distillation and student model training

[0016] We designed a composite distillation loss function to fuse the output of the multi-teacher model with the intermediate layer features; trained a lightweight student model to mimic the collective decision-making and physical consistency of the teacher model; and finally obtained a compact model that combines high accuracy, fast inference speed and strong generalization ability.

[0017] (4) Design of zone control strategy

[0018] Based on wind speed and wave height, the system operation is divided into six characteristic operating condition regions. Differentiated control objectives and weight configurations are designed for each region to achieve a full-condition control strategy covering everything from maximum power generation to survival protection.

[0019] (5) Objective function design

[0020] A multi-objective optimization function is constructed that comprehensively considers power generation efficiency, structural safety and control energy consumption. The optimal performance of the system is achieved by balancing state deviation penalty and control input penalty. A terminal cost term is introduced to ensure the stability and safety of the prediction time domain end.

[0021] (6) Design of weight adaptive mechanism

[0022] The objective function weights are dynamically adjusted based on real-time environmental parameters and operating condition zones. The automatic trade-off between power generation efficiency and structural safety is achieved through a zone lookup table method, enabling the control system to intelligently adapt to changes in different operating conditions.

[0023] (7) Definition of constraints

[0024] Hard constraints are set to control the amplitude and rate of change of the input to ensure the safety of the actuator, and the operation boundaries of key state variables are established to prevent the system from running beyond its limits. Soft constraints are used to handle tolerable temporary violations and maintain the feasibility of optimization.

[0025] (8) Upper-level optimization solution

[0026] A rolling time-domain optimization framework is adopted to solve the optimal control sequence that satisfies the constraints in real time. Efficient numerical algorithms are used to obtain suboptimal solutions in a finite time. Hot start and parallel computing are combined to improve the efficiency of online optimization.

[0027] (9) Execution of lower-level control commands

[0028] The optimization results are transformed into specific control commands that can be executed by each subsystem. The wind turbine control actuator and the PTO system actuator of the wave energy power generation device track the control commands, and the actuators are ensured to operate safely and reliably through command smoothing and dynamic limiting.

[0029] (10) Dynamic feedback correction mechanism

[0030] Based on real-time measurement data, online compensation and correction of prediction errors are performed, and a complete closed-loop correction system is established to improve the long-term robustness of the system.

[0031] Compared with the prior art, the advantages of the present invention are as follows:

[0032] 1) Reduce modeling difficulty and complexity: The model predictive control method based on the distillation model directly uses the operating data to capture system characteristics without the need for precise model analysis, thereby effectively reducing modeling difficulty and complexity.

[0033] 2) Reduce computational load and cost.

[0034] 3) Improve prediction accuracy and real-time performance: Model predictive control based on distillation model trains student model through multiple teacher models, effectively reducing computational overhead and integrating the advantages of different teacher models to improve prediction performance from the algorithm's internal perspective, thereby achieving real-time control. Attached Figure Description

[0035] Figure 1 A schematic diagram of the floating wind and wave combined power generation system provided by the present invention;

[0036] Figure 2 A flowchart illustrating the overall technical solution provided by this invention;

[0037] Figure 3 The structural diagram of the prediction model for a floating wind-wave combined wind power system based on knowledge distillation provided by this invention;

[0038] Figure 4A flowchart illustrating the modeling process of the floating wind and wave combined wind power system prediction model provided by this invention;

[0039] Figure 5 A schematic diagram of the zoning strategy for the floating wind and wave combined wind power system provided by the present invention;

[0040] Figure 6 A comparison chart of the prediction performance of the prediction model provided by this invention;

[0041] Among them, 1. wind, 2. waves, 3. seabed, 4. blades, 5. hub, 6. nacelle, 7. tower, 8. hinge point connecting the wave power generation device and the platform, 9. rocker arm of the wave power generation device, 10. float of the wave power generation device, 11. floating platform, 12. anchor chain. Detailed Implementation

[0042] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0043] The structures, proportions, and sizes illustrated in the accompanying drawings are merely for illustrative purposes and to aid those skilled in the art in understanding and reading the invention. They are not intended to limit the scope of the invention and therefore have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to size, provided they do not affect the effectiveness or purpose of the invention, should still fall within the scope of the technical content disclosed herein. Furthermore, the terms "upper," "lower," "left," "right," "middle," and "one" used in this specification are merely for clarity and not intended to limit the scope of the invention. Changes or adjustments to their relative relationships, without substantially altering the technical content, should also be considered within the scope of the invention's implementation.

[0044] like Figure 1 As shown, the environmental conditions include wind (1), waves (2), and seabed (3). The floating wind and wave combined power generation system includes three blades (4), a hub (5), a nacelle (6), a tower (7), three hinge points (8) connecting the wave power generation devices to the platform, three rocker arms (9) for the wave power generation devices, three floats (10) for the wave power generation devices, a floating platform (11), and three anchor chains (12). The hinge point, rocker arm, and float connecting the third wave power generation device to the platform are not shown in the image.

[0045] like Figures 2-3As shown, the overall technical solution process includes: knowledge distillation model, partitioning strategy, upper-level optimization solution, lower-level execution, system state update, and feedback correction. The knowledge distillation model establishment process mainly includes: a state monitoring system or high-fidelity physical model, data acquisition and processing, and a student model based on distillation. The partitioning strategy design process includes: designing the partitioning strategy, designing adaptive weights, and defining the objective function and constraints. The upper-level optimization solution process includes: predicting the model, determining partitions, solving with a global optimization solver, and obtaining control commands. Lower-level execution refers to the distribution and execution of control commands.

[0046] like Figure 4 As shown, the modeling process of the prediction model for the floating wind and wave combined wind power system includes: state monitoring system or high-fidelity physical model, acquisition and processing of dataset, training of teacher model and student model, evaluation of loss, and comparison of prediction results.

[0047] like Figure 5 As shown, the zoning strategy of the floating wind and wave combined wind power system covers all operating and survival conditions of the system. The operating conditions include Zone 1, Zone 2 and Zone 3, and the survival conditions include Zone 4, Zone 5 and Zone 6.

[0048] Zone 1: Wind speed is between 0 and the cut-in wind speed, and wave height is between 0 and the cut-out wave height.

[0049] Zone 2: The wind speed is between the cut-in wind speed and the rated wind speed, and the wave height is between 0 wave height and the cut-out wave height.

[0050] Zone 3: The wind speed is between the rated wind speed and the cut-out wind speed, and the wave height is between 0 wave height and the cut-out wave height.

[0051] Zone 4: Wind speed is greater than cut-out wind speed, and wave height is between 0 wave height and cut-out wave height.

[0052] Zone 5: Wind speed is between 0 and cut-out wind speed, and wave height is greater than cut-out wave height.

[0053] Zone 6: Wind speed is greater than cut-out wind speed, and wave height is greater than cut-out wave height.

[0054] like Figure 6The figure shows a comparison of the prediction performance of the traditional prediction model and the prediction model of this invention in terms of power and turbine speed prediction. The vertical axis represents the corresponding normalized root mean square error, which serves as a quantitative indicator for evaluating prediction accuracy. Compared to the traditional prediction model using state-space equations, the prediction model of this invention can improve the prediction accuracy of floating wind-wave combined power generation systems. Under the condition of only wind speed variation, the root mean square errors of the predicted power and turbine speed are reduced by approximately 56% and 29%, respectively. This indicates that the method of this invention has great potential in improving the prediction accuracy of floating wind-wave combined power generation systems, which is a key factor affecting the optimization of control commands.

[0055] A control method for floating wind and wave power generation systems based on knowledge distillation and zoning strategies includes the following steps:

[0056] 1. Data Acquisition and Processing

[0057] By using a condition monitoring system or a high-fidelity physical model installed on the floating wind and wave combined power generation system, operational data under different operating conditions are obtained, resulting in a dataset, such as:

[0058] (1)

[0059] X represents the dataset. Indicates wind speed. represents the wave height, Indicates the wind turbine speed. Indicates generator torque. B represents the blade pitch angle, and B represents the damping coefficient of the PTO (Power Toll Collection) system in a rocker-arm wave energy generator. The value represents the rotational angular velocity at the hinge point of the rocker-arm wave energy generator; P represents the total power of the floating wind-wave combined power generation system; BFM represents the blade root flapping torque; and TBM represents the tower base pitching moment. This represents the sway displacement of a floating platform. This indicates the pitch angle of the floating platform.

[0060] High-fidelity physical models can be created using software such as F2A or WEC-Sim.

[0061] The dataset is divided into training and test sets, and the training set is normalized, denoised, and feature extracted.

[0062] Different operating conditions need to cover all operating conditions of the floating wind and wave combined power generation system, with wind speed range from cut-in wind speed to cut-out wind speed, and wave height range from cut-in wave height to cut-out wave height.

[0063] 2. Construction of a Multi-Teacher Model System

[0064] The constructed multi-teacher model distillation system can fully exploit the trend, periodicity, and physical constraint features in the operational data of floating wind and wave combined power generation systems. Each teacher model takes the wavelet-decomposed operational data as input, focusing on the modeling task of the trend component or the periodic component, and obtains high-precision prediction results and intermediate feature representations through independent training. This system aims to improve the generalization ability and modeling efficiency of student models through multi-source knowledge fusion, providing more interpretable and robust support for prediction models in control strategy design. Specifically, Transformer networks, LSTM networks, and PINN networks are used as independent teacher models.

[0065] The inputs and outputs of the independent teacher model are as follows:

[0066] (2)

[0067] in, This represents the input to the teacher model. f This indicates the role of the teacher model. y(k+1) This represents the output of the teacher model, i.e. k+1 The status of the floating wind and wave combined power generation system at any given time. y(k) express k The status of the floating wind and wave combined power generation system at any given time. v wind (k) express k Wind speed at any moment h wave (k) express k Wave height at any moment This represents the wind turbine rotation speed at time k. This represents the generator torque at time k. This represents the blade pitch angle at time k. This represents the damping coefficient of the PTO (Platelet-Roller) wave generator system at time k. Let k represent the rotational angular velocity at the hinge point of the rocker-arm wave energy generator at time k. This represents the total power of the floating wind and wave combined power generation system at time k. This represents the leaf root flapping torque at time k. This represents the pitching moment of the tower base at time k. Let k represent the sway displacement of the floating platform at time k. This represents the pitch angle of the floating platform at time k.

[0068] 3. Knowledge Distillation and Student Model Training

[0069] To systematically transfer the integrated knowledge of multi-teacher models, this invention designs a composite distillation loss function. This function not only focuses on the final output of the teacher model, but also emphasizes its internal decision-making process and physical consistency. The composite loss function is defined as follows:

[0070] (3),

[0071] in, L total Indicates the total loss. Indicates hard loss. This indicates the output layer distillation loss. This represents the distillation loss characteristic of attention. This indicates hidden layer hints at loss. Indicates the loss of physical consistency. The weighting coefficients representing the hard loss. The weighting coefficients represent the output layer distillation losses. The weighting coefficients representing the attention feature distillation loss. This represents the weighting coefficients of the hidden layer cue loss. The weighting coefficients represent the loss of physical consistency.

[0072] Hard loss measures the difference between the student model's predicted values ​​and the true labels, and is defined as follows:

[0073] (4),

[0074] Where N is the number of training samples, and i represents the sample index. For the true value, These are the predicted values ​​from the student model.

[0075] The output layer distillation loss measures the difference between the teacher's and student's output vectors, and is defined as follows:

[0076] (5)

[0077] Where M represents the number of teacher models, and m represents the teacher model index. Let be the predicted output vector of the m-th teacher model. This is the predicted output vector of the student model. It is the square of the L2 norm, which is the square of the Euclidean distance.

[0078] Attention feature distillation loss can transfer the temporal dependencies captured by the attention matrix in the Transformer teacher model, guiding the student model to focus on key time steps. The definition of attention feature distillation loss is as follows:

[0079] (6)

[0080] Where L is the total number of layers in the Transformer model. Here is the attention matrix of the l-th layer of the Transformer teacher model. Here is the attention matrix of the l-th layer of the student model. It is the square of the Frobenius norm, which is the sum of the squares of all elements of the matrix.

[0081] Hidden cueing loss forces the intermediate layer features of the student model to be similar to the hidden layer states of teacher models such as LSTM, in order to learn their long-term dynamic characteristics. The definition of hidden cueing loss is as follows:

[0082] (7)

[0083] Where N is the number of training samples, and i represents the sample index. This represents the intermediate layer features of the input for the LSTM teacher model. This represents the intermediate layer features of the input for the student model. It is the square of the L2 norm, which is the square of the Euclidean distance.

[0084] To enhance the generalization ability and reliability of the student model in regions with scarce data, a physical constraint based on the PINN concept is introduced. The physical consistency loss is defined as follows:

[0085] (8)

[0086] Where N is the number of training samples, i represents the sample index, is the air density, A is the swept area of ​​the wind turbine, is the power coefficient, is the tip speed ratio of the i-th sample, is the blade pitch angle of the i-th sample, is the relative wind speed of the i-th sample, and is the turbine-side power of the floating wind-wave combined power generation system of the i-th sample. Where N is the number of training samples, i represents the sample index, Where is the air density, and A is the swept area of ​​the wind turbine. For power coefficient, Let be the tip velocity ratio of the i-th sample. Let be the blade pitch angle of the i-th sample. Let be the relative wind speed of the i-th sample. Let be the wind turbine-side power in the floating wind-wave combined power generation system of the i-th sample.

[0087] It is worth noting that this only demonstrates the physical formula constraints for calculating the wind turbine power in a floating wind-wave combined power generation system. More physical formulas can be used to better constrain the student model of the floating wind-wave combined power generation system, which will not be shown here.

[0088] During the training phase, the parameters of all teacher models are fixed. Training data is simultaneously input into both the multi-teacher model and the student model. By minimizing the aforementioned composite loss function and using gradient descent backpropagation, only the parameters of the student model are updated.

[0089] Through this process, the student model not only learns the real mapping relationships in the data, but also internalizes the trend and periodic features and physical prior knowledge integrated by the teacher model, ultimately forming a lightweight prediction model that combines high accuracy, fast reasoning speed and strong physical interpretability, laying the core foundation for the design of the subsequent real-time optimal controller.

[0090] 4. Partition control strategy design

[0091] Based on the two-dimensional operating space of wind speed and wave height, this invention proposes a six-zone control strategy to achieve refined intelligent control of floating wind-wave combined power generation systems. The zoning function is as follows:

[0092] (9)

[0093] in, This represents the partitioning function for a floating wind-wave combined power generation system based on wind speed and wave height. Indicates wind speed. represents the wave height, Indicates region 1, Indicates region 2, Indicates region 3. Indicates region 4, Indicates region 5, Indicates region 6, Indicates the cut-in wind speed. Indicates the rated wind speed. Indicates the cut-out wind speed. Indicates the cut-in wave height, This indicates the cut-out wave height.

[0094] 5. Objective Function Design

[0095] To achieve multi-objective optimization control of a floating wind and wave combined power generation system in a complex marine environment, this invention designs a multi-objective function based on a model predictive control (MPC) framework. This function optimizes system performance in the prediction time domain, and its discrete-time expression is as follows: (10)

[0096] in, This represents the objective function value, and k represents the prediction step size index. Indicates the length of the prediction time domain. Represents the tracking error vector. This represents the transpose of the tracking error vector. Represents the state error weight matrix. Represents the control input vector. This represents the transpose of the control input vector. This represents the control input weight matrix. Represents the terminal state error vector. This represents the transpose of the terminal state error vector. This represents the terminal cost weight matrix.

[0097] 6. Weight Adaptive Mechanism Design

[0098] To address the challenge of fixed weights failing to handle dynamic changes across all system operating conditions, this invention proposes a weight adaptive mechanism based on real-time operating condition partitioning. This mechanism dynamically adjusts the weight matrix in the objective function according to the current environmental state and system response. The weight adaptive strategy function is defined as follows:

[0099] (11)

[0100] Where k represents the prediction step size index, This represents the state error weight matrix at step k. This represents the control input weight matrix at step k. Indicates wind speed. represents the wave height, Represents the partitioning function;

[0101] (12)

[0102] in, Represents the state error weight matrix. Represents a diagonal matrix , indicating the power generation tracking weight, Indicates the structural load suppression weight. Indicates the platform motion suppression weights;

[0103] (13)

[0104] in, This represents the control input weight matrix. Represents a diagonal matrix. Indicates the wind turbine speed. Indicates generator torque. Indicates the blade pitch angle. This represents the damping coefficient of the first rocker arm wave energy generation device PTO system. This represents the rotational angular velocity at the hinge point of the first rocker-arm wave energy generation device. This represents the damping coefficient of the second rocker-arm wave energy generation device PTO system. This represents the rotational angular velocity at the hinge point of the second rocker-arm wave energy generation device. This represents the damping coefficient of the third rocker-arm wave energy generation device PTO system. This indicates the rotational angular velocity at the hinge point of the third rocker-arm wave energy generator.

[0105] 7. Constraint Definition

[0106] To ensure the physical and operational feasibility of the optimization solution, the present invention imposes the following constraints on each prediction step size.

[0107] Control input constraints: (14)

[0108] in, This represents the control input vector at step k. This represents the lower limit vector of the control input. This represents the upper limit vector of the control input;

[0109] Control input rate of change constraint: (15)

[0110] in, This represents the control input increment vector at step k. This represents the lower bound vector for controlling the input increment. This represents the upper bound vector controlling the input increment;

[0111] Power constraints in state variables: (16)

[0112] in, This represents the total power of the floating wind and wave combined power generation system at step k. This indicates the rated total power of the floating wind and wave combined power generation system;

[0113] Leaf root flapping moment constraints in state variables: (17)

[0114] in, This represents the leaf root flapping torque at step k. This represents the maximum value of the leaf root flapping torque;

[0115] Forward and backward pitch moment constraints of the tower base in the state variables: (18)

[0116] in, This represents the pitching moment of the tower base at step k. This represents the maximum value of the pitching moment at the front and rear of the tower base;

[0117] Floating platform oscillation displacement constraints in state variables: (19)

[0118] in, This represents the sway displacement of the floating platform at step k. This represents the maximum value of the sway displacement of the floating platform;

[0119] The pitch angle constraint of the floating platform in the state variables: (20)

[0120] in, This represents the pitch angle of the floating platform at step k. This indicates the maximum pitch angle of the floating platform.

[0121] 8. Upper-level optimization solution

[0122] The upper-level optimization solver employs a rolling time-domain optimization framework, solving for the optimal control sequence that satisfies multi-objective constraints in real time based on the predicted output of a lightweight student model during each control cycle. This module comprehensively considers system operating states, environmental disturbance predictions, and dynamic weight configuration, rapidly generating a globally optimal solution that balances power generation efficiency, structural safety, and control energy consumption through efficient numerical optimization algorithms. The optimization process fully integrates boundary constraints from the partitioning strategy, ensuring the adaptability and safety of control commands under different operating conditions.

[0123] 9. Execution of lower-level control commands

[0124] The lower-level control execution module transforms the optimized control sequence generated by the upper level into specific executable instructions for each subsystem, including actuator signals such as wind turbine pitch angle, generator torque, and wave energy device damping. This module employs instruction smoothing and dynamic limiting strategies to avoid frequent actuator movements and excessive loads, ensuring reliable execution of control actions within physical constraints. All subsystems maintain coordinated operation during execution, achieving efficient and comprehensive utilization of wind and wave energy and stable system operation.

[0125] 10. Dynamic feedback correction mechanism

[0126] To address model uncertainties and changes in the external environment, the system introduces a dynamic feedback correction mechanism across multiple time scales. Through error analysis of real-time sensor data and predicted outputs, the system performs online corrections to state estimates, model parameters, and control weights, effectively suppressing disturbance accumulation and performance drift. This mechanism continuously improves the prediction accuracy of the student model and the robustness of the control system, ensuring that the system maintains optimal operating conditions throughout its entire lifecycle.

[0127] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A control method for a floating wind and wave power generation system based on knowledge distillation and partitioning strategies, characterized in that, Includes the following steps: (1) Data acquisition and processing: acquire full-condition operation data of floating wind and wave combined power generation system through sensors and high-fidelity models; normalize, denoise and extract features from raw data for preprocessing. Divide the data into training and testing sets to provide a high-quality data foundation for subsequent modeling. (2) Construction of a multi-teacher model system: A multi-teacher model system consisting of Transformer, LSTM and PINN is constructed. Each teacher model focuses on different feature dimensions of trend, cycle and physical constraint, and obtains high-precision prediction results and diversified knowledge representations through independent training. (3) Knowledge distillation and student model training: Design a composite distillation loss function to fuse the output and intermediate layer features of the multi-teacher model; The lightweight student model is trained to mimic the collective decision-making and physical consistency of the teacher model; ultimately, a compact model with high accuracy, fast inference speed and strong generalization ability is obtained. (4) Zonal control strategy design: Based on wind speed and wave height, the system operation is divided into six characteristic operating condition zones. Differentiated control objectives and weight configurations are designed for each zone to achieve a full-condition control strategy covering everything from maximum power generation to survival protection. The partitioning function of the partition control strategy is as follows: (9) in, This represents the partitioning function for a floating wind-wave combined power generation system based on wind speed and wave height. Indicates wind speed. h wave represents the wave height, Indicates region 1, Indicates region 2, Indicates region 3. Indicates region 4, Indicates region 5, Indicates region 6, Indicates the cut-in wind speed. Indicates the rated wind speed. Indicates the cut-out wind speed. Indicates the cut-in wave height, Indicates the cut-out wave height; (5) Objective function design: Construct a multi-objective optimization function that comprehensively considers power generation efficiency, structural safety and control energy consumption. Achieve optimal system performance by balancing state deviation penalty and control input penalty. Introduce terminal cost term to ensure stability and safety at the end of the prediction time domain. (6) The weight adaptive mechanism is designed to dynamically adjust the weight of the objective function according to the real-time environmental parameters and operating condition partitions. The automatic trade-off between power generation efficiency and structural safety is achieved through the partition lookup table method, so that the control system can intelligently adapt to changes in different operating conditions. The function definition of the weight adaptive mechanism is: (11) Where k represents the prediction step size index, This represents the state error weight matrix at step k. This represents the control input weight matrix at step k. Indicates wind speed. represents the wave height, Represents the partitioning function; (12) in, Represents the state error weight matrix. Represents a diagonal matrix , indicating the power generation tracking weight, Indicates the structural load suppression weight. Indicates the platform motion suppression weights; (13) in, This represents the control input weight matrix. Represents a diagonal matrix. Indicates the wind turbine speed. Indicates generator torque. Indicates the blade pitch angle. This represents the damping coefficient of the first rocker arm wave energy generation device PTO system. This represents the rotational angular velocity at the hinge point of the first rocker-arm wave energy generation device. This represents the damping coefficient of the second rocker-arm wave energy generation device PTO system. This represents the rotational angular velocity at the hinge point of the second rocker-arm wave energy generation device. This represents the damping coefficient of the third rocker-arm wave energy generation device PTO system. This indicates the rotational angular velocity at the hinge point of the third rocker-arm wave energy generator. (7) Define constraints, set hard constraints on the amplitude and rate of change of the control input to ensure the safety of the actuator, establish the operation boundary of key state variables to prevent the system from running beyond the limit, and handle tolerable temporary violations through soft constraints to maintain the feasibility of optimization; (8) Upper-level optimization solution: The rolling time-domain optimization framework is used to solve the optimal control sequence that satisfies the constraints in real time. The efficient numerical algorithm is used to obtain the suboptimal solution in a finite time. The efficiency of online optimization is improved by combining hot start and parallel computing. (9) The lower-level control command is executed, and the optimization results are converted into specific control commands that can be executed by each subsystem. The wind turbine control actuator and the PTO system actuator of the wave energy power generation device track the control commands, and ensure the safe and reliable operation of the actuator through command smoothing and dynamic limiting. (10) Dynamic feedback correction mechanism: Based on real-time measurement data, online compensation and correction of prediction error are carried out, and a complete closed-loop correction system is established to improve the long-term robustness of the system.

2. The control method for a floating wind and wave power generation system based on knowledge distillation and partitioning strategy as described in claim 1, characterized in that, In step (2), the input and output of the independent teacher model are as follows: (2) in, This represents the input to the teacher model. f This indicates the role of the teacher model. y(k+1) This represents the output of the teacher model, i.e. k+1 The status of the floating wind and wave combined power generation system at any given time. y (k) express k The status of the floating wind and wave combined power generation system at any given time. v wind (k) express k Wind speed at any moment h wave (k) express k Wave height at any moment This represents the wind turbine rotation speed at time k. This represents the generator torque at time k. This represents the blade pitch angle at time k. This represents the damping coefficient of the PTO (Platelet-Roller) wave generator system at time k. Let k represent the rotational angular velocity at the hinge point of the rocker-arm wave energy generator at time k. This represents the total power of the floating wind and wave combined power generation system at time k. This represents the leaf root flapping torque at time k. This represents the pitching moment of the tower base at time k. Let k represent the sway displacement of the floating platform at time k. This represents the pitch angle of the floating platform at time k.

3. The control method for a floating wind and wave power generation system based on knowledge distillation and partitioning strategy as described in claim 1, characterized in that, In step (3), the composite loss function is defined as follows: (3), in, L total Indicates the total loss. Indicates hard loss. This indicates the output layer distillation loss. This represents the distillation loss characteristic of attention. This indicates hidden layer hints at loss. Indicates the loss of physical consistency. The weighting coefficients representing the hard loss. The weighting coefficients represent the output layer distillation losses. The weighting coefficients representing the attention feature distillation loss. This represents the weighting coefficients of the hidden layer cue loss. The weighting coefficients represent the loss of physical consistency.

4. The control method for a floating wind and wave power generation system based on knowledge distillation and partitioning strategy as described in claim 3, characterized in that, Hard loss measures the difference between the student model's predicted values ​​and the true labels, and is defined as follows: (4), Where N is the number of training samples, and i represents the sample index. For the true value, These are the predicted values ​​from the student model; The output layer distillation loss measures the difference between the teacher's and student's output vectors, and is defined as follows: （5） Where M represents the number of teacher models, and m represents the teacher model index. Let be the predicted output vector of the m-th teacher model. This is the predicted output vector of the student model. It is the square of the L2 norm, which is the square of the Euclidean distance.

5. The control method for a floating wind and wave power generation system based on knowledge distillation and partitioning strategy as described in claim 3, characterized in that the attention feature distillation loss can transfer the temporal dependencies captured by the attention matrix in the Transformer teacher model, guiding the student model to focus on key time steps. The definition of attention feature distillation loss is as follows: （6） in, L is the total number of layers in the Transformer model. Here is the attention matrix of the l-th layer of the Transformer teacher model. Here is the attention matrix of the l-th layer of the student model. It is the square of the Frobenius norm, which is the sum of the squares of all elements of the matrix; Hidden cueing loss forces the intermediate layer features of the student model to be similar to the hidden layer states of the LSTM teacher model in order to learn its long-term dynamic characteristics. The definition of hidden cueing loss is as follows: (7) Where N is the number of training samples, and i represents the sample index. This represents the intermediate layer features of the input for the LSTM teacher model. This represents the intermediate layer features of the input for the student model. It is the square of the L2 norm, which is the square of the Euclidean distance.

6. The control method for a floating wind and wave power generation system based on knowledge distillation and partitioning strategy as described in claim 3, characterized in that, To enhance the generalization ability and reliability of the student model in regions with scarce data, a physical constraint based on the PINN concept is introduced. The physical consistency loss is defined as follows: (8) Where N is the number of training samples, and i represents the sample index. Where is the air density, and A is the swept area of ​​the wind turbine. For power coefficient, Let be the tip velocity ratio of the i-th sample. Let be the blade pitch angle of the i-th sample. Let be the relative wind speed of the i-th sample. Let be the wind turbine-side power in the floating wind-wave combined power generation system of the i-th sample.

7. The control method for a floating wind and wave power generation system based on knowledge distillation and partitioning strategy as described in claim 1, characterized in that, The discrete-time expression of the multi-objective optimization function in step (5) is as follows: (10) in, This represents the objective function value, and k represents the prediction step size index. Indicates the length of the prediction time domain. Represents the tracking error vector. This represents the transpose of the tracking error vector. Represents the state error weight matrix. Represents the control input vector. This represents the transpose of the control input vector. This represents the control input weight matrix. Represents the terminal state error vector. This represents the transpose of the terminal state error vector. This represents the terminal cost weight matrix.

8. The control method for a floating wind and wave power generation system based on knowledge distillation and partitioning strategy as described in claim 1, characterized in that, In step (7), the hard constraints include control input constraints and control input change rate constraints, and the soft constraints include power constraints, blade root flapping moment constraints, tower base pitching moment constraints, floating platform sway displacement constraints, and floating platform pitching angle constraints in the state variables. Control input constraints: (14) in, This represents the control input vector at step k. This represents the lower limit vector of the control input. This represents the upper limit vector of the control input; Control input rate of change constraint: (15) in, This represents the control input increment vector at step k. This represents the lower bound vector for controlling the input increment. This represents the upper bound vector controlling the input increment; Power constraints in state variables: (16) in, This represents the total power of the floating wind and wave combined power generation system at step k. This indicates the rated total power of the floating wind and wave combined power generation system; Leaf root flapping moment constraints in state variables: (17) in, This represents the leaf root flapping torque at step k. This represents the maximum value of the leaf root flapping torque; Forward and backward pitch moment constraints of the tower base in the state variables: (18) in, This represents the pitching moment of the tower base at step k. This represents the maximum value of the pitching moment at the front and rear of the tower base; Floating platform oscillation displacement constraints in state variables: (19) in, This represents the sway displacement of the floating platform at step k. This represents the maximum value of the sway displacement of the floating platform; The pitch angle constraint of the floating platform in the state variables: (20) in, This represents the pitch angle of the floating platform at step k. This indicates the maximum pitch angle of the floating platform.

Images