Wind power cluster wake optimization method and device based on model predictive control

Abstract

This application relates to the field of wind power predictive control technology, and discloses a method and apparatus for optimizing the wake of a wind power cluster based on model predictive control. The method includes: collecting the operating state variables and meteorological environment information of each wind turbine in the entire wind farm; inputting the operating state variables and meteorological environment information into a comprehensive prediction model; and outputting the future wake of each wind turbine in the entire wind farm. N p The power generation sequence at each time point; construct the objective function and constraints, and combine them with future... N p The power generation sequence at each moment determines the future power output of each wind turbine. N p The sequence of yaw angles at each moment; based on the future N p The value of the first moment in the yaw angle sequence controls the wind turbine to perform yaw actions, and forms a rolling closed-loop control according to the rolling time-domain optimization mechanism. In this way, the accuracy of the physical mechanism is preserved, while also possessing the nonlinear fitting capability of deep learning, which can improve the accuracy of overall power prediction and take into account both high-precision physical field description and fast response of cluster control.

Description

Technical Field

[0001] This application relates to the field of wind power predictive control technology, and for example to a method and apparatus for optimizing the wake of a wind power cluster based on model predictive control. Background Technology

[0002] The core objective of wind farm cluster control is to optimize the flow field distribution by coordinating the control variables (such as yaw angle and pitch angle) of each turbine while ensuring the safe operation of the turbines, thereby maximizing the total power generation of the entire farm. As the installed capacity and layout density of wind farms increase, the wake interference effect between turbines becomes significant, making it difficult for traditional single-unit control strategies to achieve optimal overall efficiency. Currently, wind farm cluster optimization control technologies are mainly divided into the following four categories:

[0003] (1) Independent control method based on traditional MPPT: This is the most common control strategy adopted in commercial wind farms. Each wind turbine independently adjusts its yaw and speed to track its own maximum power point based solely on its own sensor data (wind speed, wind direction). However, since each unit operates independently, the upstream unit is always facing the wind in order to capture its maximum wind energy. The strong wake generated will seriously interfere with the downstream unit. This not only leads to a decrease in the power generation of the downstream unit, but also increases its fatigue load, resulting in the overall power generation efficiency of the entire farm being lower than the theoretical optimal value.

[0004] (2) Co-optimization based on engineering wake model: Simplified analytical formulas (such as Jensen model or Gaussian model) are used to describe the decay and diffusion of wake velocity, and optimization algorithms are combined to find the optimal yaw angle for the entire field. Analytical models are usually based on simplified assumptions of static and linearity, which makes it difficult to accurately describe the complex nonlinear turbulent mixing and dynamic deflection process in the actual flow field. Under actual working conditions with frequent wind direction fluctuations, the prediction deviation is large, which can easily lead to the failure of control strategies;

[0005] (3) Dynamic control based on high-fidelity CFD: Large eddy simulation (LES) or Reynolds-averaged equations (RANS) are used to solve the fluid dynamics equations in real time, and control decisions are made based on detailed flow field information. This method theoretically has the highest control accuracy and can capture complex turbulent details, but the computation time is long. MPC control relies on rolling time-domain optimization, which requires repeated simulations of the flow field state over a period of time within seconds or milliseconds. The computational cost of CFD solving the Navier-Stokes equations is extremely high, and a single calculation often takes several hours, far exceeding the real-time response cycle of the control system, making it difficult to implement in real time in engineering.

[0006] (4) Data-driven control based on traditional deep learning: Using Long Short-Term Memory (LSTM) networks or standard Transformer models, the time-series mapping relationship between wind speed and power is learned through historical SCADA data, and this relationship is embedded as a prediction model into the Model Predictive Control (MPC) framework for rolling optimization. However, this method relies excessively on data quality, has poor generalization ability under extreme conditions where data is sparse or the training set is not covered, and lacks physical consistency constraints, making it easy to predict values ​​that violate physical common sense (such as violating the power curve). Moreover, existing time-series models often treat each unit as an independent node, ignoring the influence of adjacent units within the wind farm (i.e., how the upstream wake is directionally transmitted to the downstream), resulting in a significant decrease in model prediction accuracy when changes in wind direction cause changes in the wake. At the same time, the computational complexity of the standard neural network architecture is high. When MPC requires long-term rolling prediction, the computational resource consumption increases quadratically with the increase of the prediction step size, resulting in slow online inference speed.

[0007] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0008] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or describe the scope of protection of these embodiments, but rather as a prelude to the detailed description that follows.

[0009] This disclosure provides a method and apparatus for optimizing the wake of a wind power cluster based on model predictive control, which balances high-precision physical field description and rapid response of cluster control.

[0010] In some embodiments, the wind farm wake optimization method based on model predictive control includes: S10, data acquisition: collecting the operating state variables and meteorological environment information of each wind turbine in the entire farm; S20, comprehensive prediction: inputting the operating state variables and meteorological environment information into a comprehensive prediction model; the comprehensive prediction model includes: a PI-DANN model with embedded NS equations and vorticity equation physical constraints, a GAT model, and a PI-Informer model with physical constraints based on the standard power curve introduced in the loss function; the PI-DANN model generates the adjacency matrix of the wind farm; the GAT model determines the effective capture wind speed of each wind turbine based on the adjacency matrix; and the PI-Informer model outputs the future wind speed of each wind turbine in the entire farm based on the effective capture wind speed. N pThe power generation sequence at each time point; S30, optimization decision: construct the objective function and constraints, and combine them with the future... N p The power generation sequence at each moment determines the future power output of each wind turbine. N p The yaw angle sequence at each moment; S40, Execute feedback: based on the future... N p The value of the first moment in the yaw angle sequence controls the wind turbine to perform yaw action, and forms a rolling closed-loop control according to the rolling time-domain optimization mechanism.

[0011] In some embodiments, the wind farm wake optimization device based on model predictive control includes: a data acquisition module configured to collect operating state variables and meteorological environment information of each wind turbine in the entire farm; and a comprehensive prediction module configured to input the operating state variables and meteorological environment information into a comprehensive prediction model. The comprehensive prediction model includes: a PI-DANN model with embedded Navier-Stokes equations and vorticity equations as physical constraints, a GAT model, and a PI-Informer model with physical constraints based on standard power curves introduced into the loss function. The PI-DANN model generates an adjacency matrix for the wind farm, the GAT model determines the effective capture wind speed of each wind turbine based on the adjacency matrix, and the PI-Informer model outputs the future wind speed of each wind turbine in the entire farm based on the effective capture wind speed. N p The power generation sequence at each time point; the optimization decision module is configured to construct the objective function and constraints, and combine them with the future... N p The power generation sequence at each moment determines the future power output of each wind turbine. N p The sequence of yaw angles at each moment; the execution feedback module is configured to, based on the future... N p The value of the first moment in the yaw angle sequence controls the wind turbine to perform yaw action, and forms a rolling closed-loop control according to the rolling time-domain optimization mechanism.

[0012] The wind power cluster wake optimization method and apparatus based on model predictive control provided in this disclosure can achieve the following technical effects:

[0013] A dynamic adjacency matrix conforming to hydrodynamics is generated using the PI-DANN model. Combined with the graph attention mechanism of the GAT model, the influence relationships between wind turbines are accurately described. Long-term prediction is then achieved through the PI-Informer model. By constructing a spatiotemporal integrated model that fuses the graph attention network (GAT) and the physical information informer (PI-Informer), the system can quickly calculate wind farm power changes over a future period and accurately describe the wake interference relationships between wind turbines. Based on this, a model predictive control (MPC) framework is used to achieve real-time rolling optimization of turbine yaw angles, thereby improving the total power generation and economic benefits of the wind farm cluster while ensuring operational safety. This architecture retains the accuracy of the physical mechanisms while possessing the nonlinear fitting capabilities of deep learning, improving the accuracy of overall power prediction and balancing high-precision physical field description with rapid cluster control response.

[0014] The above general description and the description below are exemplary and illustrative only and are not intended to limit this application. Attached Figure Description

[0015] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations and drawings do not constitute a limitation on the embodiments. Elements having the same reference numerals in the drawings are shown as similar elements. The drawings are not to be scaled. And wherein:

[0016] Figure 1 This is a schematic diagram of the wake optimization method for wind power clusters based on model predictive control provided in the embodiments of this disclosure;

[0017] Figure 2 This is a schematic diagram of the structure of the MPC closed-loop control system of the integrated prediction model provided in the embodiments of this disclosure;

[0018] Figure 3 This is a schematic diagram of the integrated power generation prediction framework and error correction mechanism provided in the embodiments of this disclosure;

[0019] Figure 4 This is a schematic diagram of a single-machine wake prediction model based on PI-DANN provided in an embodiment of this disclosure;

[0020] Figure 5 This is a schematic diagram of a multi-engine wake field prediction model based on GAT provided in an embodiment of this disclosure;

[0021] Figure 6 This is a schematic diagram of the power generation prediction model based on PI-Informer provided in the embodiments of this disclosure;

[0022] Figure 7 This is a schematic diagram of the online update and error correction mechanism of the model provided in the embodiments of this disclosure;

[0023] Figure 8 This is a schematic diagram comparing the predicted values ​​of a typical heavy wake turbine unit under different models according to embodiments of this disclosure;

[0024] Figure 9 This is a schematic diagram comparing the power generation prediction performance of different models provided in the embodiments of this disclosure in a typical range;

[0025] Figure 10 This is a schematic diagram of the power enhancement rate (wind speed 11m / s) of the MPC strategy provided in this embodiment under different wind directions;

[0026] Figure 11 This is a schematic diagram illustrating the power increment and boost rate statistics of the MPC cluster control strategy under year-round operating conditions provided in this embodiment of the disclosure;

[0027] Figure 12 This is a schematic diagram of a wind power cluster wake optimization device based on model predictive control provided in an embodiment of this disclosure.

[0028] Figure label:

[0029] 120. Wind power cluster wake optimization device based on model predictive control; 121. Data acquisition module; 122. Comprehensive prediction module; 123. Optimization decision module; 124. Execution feedback module. Detailed Implementation

[0030] To provide a more detailed understanding of the features and technical content of the embodiments of this disclosure, the implementation of the embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for illustrative purposes only and are not intended to limit the embodiments of this disclosure. In the following technical description, for ease of explanation, several details are used to provide a full understanding of the disclosed embodiments. However, one or more embodiments may still be implemented without these details. In other cases, well-known structures and devices may be simplified in their depiction to simplify the drawings.

[0031] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0032] Unless otherwise stated, the term "multiple" means two or more.

[0033] In this embodiment of the disclosure, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B.

[0034] The term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0035] The term "correspondence" can refer to an association or binding relationship. The correspondence between A and B means that there is an association or binding relationship between A and B.

[0036] Combination Figure 1 As shown, this disclosure provides a method for optimizing the wake of a wind power cluster based on model predictive control, including:

[0037] S10, Data Acquisition: Collect operational status variables and meteorological environment information for each wind turbine unit in the entire site;

[0038] S20, Integrated Prediction: Operating state variables and meteorological environmental information are input into the integrated prediction model. The integrated prediction model includes: a PI-DANN model with embedded Navier-Stokes equations and vorticity equations as physical constraints, a GAT model, and a PI-Informer model with physical constraints based on the standard power curve incorporated into the loss function. The PI-DANN model generates the adjacency matrix of the wind farm, the GAT model determines the effective wind speed captured by each wind turbine based on the adjacency matrix, and the PI-Informer model outputs the future wind speed of each wind turbine in the entire farm based on the effective wind speed. N p The power generation sequence at each moment;

[0039] S30, Optimization Decision: Constructing the objective function and constraints, and combining them with future... N p The power generation sequence at each moment determines the future power output of each wind turbine. N p Yaw angle sequence at each moment;

[0040] S40, Execution Feedback: Based on future... N p The value of the first moment in the yaw angle sequence controls the wind turbine to perform yaw action, and forms a rolling closed-loop control according to the rolling time-domain optimization mechanism.

[0041] Combination Figure 2 As shown in the embodiments of this disclosure, an MPC closed-loop control system based on a comprehensive prediction model is provided. This closed-loop control system consists of four stages: data acquisition, comprehensive prediction, optimization decision-making, and execution feedback.

[0042] (1) State perception and input

[0043] The control system collects real-time data from the entire field through the wind turbine (hereinafter referred to as "the turbine") data acquisition and monitoring control system (SCADA). N wt The operating status variables of the generator unit and meteorological environmental information. Among them, the operating status variables include: generator torque. T g,i Generator speed w g,i Wind turbine speed w r,i Pitch angle β i Current yaw angle γ i Meteorological information includes: incoming wind speed at the cabin. U 0. Incoming wind direction turbulence intensity I 0. These data are the inputs for the initial state.

[0044] (2) Predicting power generation

[0045] The collected information is input into the "Integrated Prediction Model". This model is composed of a single-unit wake model (PI-DANN), a multi-unit wake field model (GAT), and a power generation model (PI-Informer) connected in series (details of each sub-model are provided below). The PI-DANN model incorporates physical constraints based on the Navier-Stokes equations and vorticity equations, while the PI-Informer model introduces physical constraints based on the standard power curve into its loss function. This model is responsible for processing candidate control inputs generated by the optimizer (i.e., the yaw angle sequence to be optimized). γ Predicting the future N p The power output sequence of each unit in the entire field at any given moment. P i Specifically, the adjacency matrix of the wind farm is dynamically generated using the wake envelope generated by the PI-DANN model. The GAT model aggregates features in a graph structure based on the adjacency matrix and determines the effective wind speed captured by each wind turbine. The PI-Informer model determines the future wind speed captured by each wind turbine across the entire field based on the effective wind speed captured by each wind turbine. N p The power generation sequence at each time point.

[0046] (3) MPC optimization decision

[0047] 1) Constructing the objective function: to predict the time domain N p The goal is to maximize the total power generation of the entire site.

[0048] (1)

[0049] 2) Constructing constraints: Three constraints must be satisfied simultaneously:

[0050] Yaw angle amplitude constraint: The yaw angle of the unit Limited to a safe range To prevent mechanical interference between the nacelle and the tower or excessive load due to excessive yaw;

[0051] Yaw rate constraint: Limit the change in yaw angle between adjacent time points to no more than To reduce wear and fatigue load on the yaw actuator; among which, This represents the maximum permissible yaw rate per step.

[0052] Power output constraints: ensuring predicted power P i It is within a reasonable range that is non-negative and does not exceed the rated power.

[0053] (2)

[0054] 3) Optimized Solver: The Differential Evolution (DE) algorithm is employed. To address the nonlinear characteristics of the prediction model, the optimal yaw angle sequence is searched within the rolling time domain through population mutation, crossover, and selection calculations. .

[0055] 4) Execution and Feedback: The controller selects only the first value in the optimal yaw angle sequence and sends it to the wind farm's main control system to drive the turbine to perform the yaw action. At the next moment, the system collects new state feedback and repeats the above "prediction-optimization-execution" process to form a rolling closed-loop control.

[0056] This disclosure presents a wind farm cluster wake optimization method based on model predictive control. It utilizes a PI-DANN model to dynamically generate an adjacency matrix conforming to hydrodynamics, combines a GAT model's graph attention mechanism to accurately describe the influence relationships between turbines, and then uses a PI-Informer model to achieve long-sequence prediction. By constructing a spatiotemporal integrated model that fuses the graph attention network (GAT) and physical information informer (PI-Informer), it can quickly calculate wind farm power changes over a future period and accurately describe the wake interference relationships between turbines. Based on this, combined with a model predictive control (MPC) framework, it achieves real-time rolling optimization of turbine yaw angles, thereby improving the total power generation and economic benefits of the wind farm cluster while ensuring operational safety. This architecture retains the accuracy of the physical mechanism while possessing the nonlinear fitting capabilities of deep learning, improving the accuracy of overall power prediction and balancing high-precision physical field description with rapid cluster control response.

[0057] To accurately predict the power response of a wind farm under yaw control, this embodiment provides, as follows: Figure 3 The illustrated series-based integrated prediction model consists of three stages:

[0058] (1) Physical topology generation stage: The single-machine wake prediction model (PI-DANN) is used, based on the real-time wind direction. Set the current coordinate system and update the coordinates of all units, taking into account the position of each unit. x i , y i , z i Calculate the wake region of each turbine. If a downstream turbine is located within the wake envelope of an upstream turbine, a connection is determined between them. This dynamically generates the adjacency matrix of the wind farm. This matrix defines the influence relationships between units under the current operating conditions at a physical level, providing an accurate topology for subsequent graph networks;

[0059] (2) Spatial feature aggregation stage: Using a graph attention neural network (GAT), each unit is regarded as a node in a graph, and an adjacency matrix that can express the influence relationship between units is generated. The model is considered as an edge of a graph. It calculates the physical equivalent wind speed for each turbine by aggregating the wake influence characteristics of upstream turbines within the graph structure. To eliminate biases in the physical model and integrate measured information, a Kalman filter algorithm is further introduced to weight and fuse the physical wind speed calculated by GAT with the measured nacelle wind speed collected by the SCADA system, yielding the final effective wind speed. U i；

[0060] (3) Time-series power prediction stage: Using the physical information-informer neural network (PI-Informer), the effective captured wind speed output by the fusion of GAT and Kalman filter is received. U i And the historical state sequence of the generating units. Utilizing Informer's long-sequence inference capabilities, the future power generation of each generating unit can be quickly predicted. P i .

[0061] In addition, such as Figure 3 As shown below, considering the potential for cumulative errors in the cascaded model, the system incorporates an online correction module. Specifically, in the flow field correction loop, the prediction error of the PI-DANN model is monitored in real time using measured data from the unit's LiDAR (Light Detection and Ranging) system. Wind speed prediction error compared to GAT model In the power correction loop, the power prediction error of the PI-Informer is monitored in real time using the measured power from the SCADA system. If the error in any step continues to exceed the preset threshold... (For example MAPE When the error rate exceeds 10%, the framework will trigger an online update mechanism to fine-tune and update the corresponding modules using the latest measured data, thereby dynamically correcting the differences in the distribution of historical accumulated errors and new data, and maintaining the high-precision operation of the model throughout its entire lifecycle.

[0062] Combination Figure 4 As shown, the Physical Information-Domain Adversarial Neural Network (PI-DANN) is used as the topology generator, specifically including:

[0063] (1) Input

[0064] Unit coordinates ( x i , y i , z i ), environmental parameters ( U 0, I 0) and control variables (yaw angle γ) i These data are obtained from real-time data collection by the SCADA system, wind farm weather forecasts, and the coordinates of each turbine in the wind farm.

[0065] (2) Model structure

[0066] The main body of the model consists of a PINN-based wake predictor, which incorporates physical constraints from the NS equation and vorticity equation, enabling it to accurately calculate the deflection trajectory and width of a single-unit wake.

[0067] (3) Output

[0068] Instead of directly using the flow velocity field output by the PI-DANN model, the wake envelope calculated by it is used as a geometric criterion to dynamically determine whether there is a wake obstruction relationship between any two units, thereby generating an adjacency matrix. The adjacency matrix is ​​passed to the next sub-model (GAT) as edge information in the graph neural network, defining the information transmission paths between nodes. It is dynamically updated as the wind direction changes, ensuring the topology always conforms to the laws of hydrodynamics.

[0069] It should be noted that the specific structure of the PI-DANN model can be found in the patent application number CN202511339876.3.

[0070] Combination Figure 5 As shown, to address the complex coupling between multiple units, this embodiment constructs a physically guided graph attention network (GAT), specifically including:

[0071] (1) Input:

[0072] The input to a graph includes two parts: the node information and the edge information.

[0073] Node information: Unit coordinates ( x i , y i , z i ), incoming air velocity U 0. Incoming wind direction I 0 and control variables (yaw angle γ) i );

[0074] Edge information: Adjacency matrix generated by the previous PI-DANN model ;

[0075] The adjacency matrix comes from the PI-DANN model, and the remaining data comes from the unit's SCADA system.

[0076] (2) Model structure

[0077] Graph structure definition: Each unit is defined as a node in the graph, and the wake influence relationship between units is defined as a directed edge in the graph.

[0078] Attention Mechanism: The model introduces a multi-head attention mechanism to automatically calculate the upstream node. j For downstream nodes i Influence weight α ij :

[0079] (3)

[0080] in, This represents the node feature vector. || indicates a vector concatenation operation. This represents the weight vector of a single-layer feedforward neural network, with LeakyReLU as the non-linear activation function and Softmax as the normalized exponential function. This mechanism can adaptively allocate wake influence weights based on distance and yaw angle.

[0081] (3) Output

[0082] The physical equivalent wind speed for each unit. While the unit's SCADA anemometer can provide real-time feedback, its measurements suffer from significant high-frequency noise and systematic bias due to the impeller wake and nacelle obstruction effect. Although the GAT model can accurately calculate the wake deficit trend based on physical mechanisms, it cannot detect instantaneous random disturbances in the flow field. Therefore, to further improve data quality, a Kalman filter algorithm is introduced to fuse the physical wind speed calculated by GAT with the nacelle wind speed measured by SCADA, resulting in an effective captured wind speed that combines physical and measured data. U i .

[0083] First, a one-dimensional discrete stochastic state model of wind speed is established. k Real-time, effective wind speed capture is defined as a state variable. The calculated values ​​of the GAT model As a priori prediction of the system, the nacelle wind speed recorded by the unit's SCADA system is defined as a measurement variable. The state equation and the equation for the measured variable are defined as follows:

[0084] (4)

[0085] In the formula, N (0, Q )and N (0, R ) represent a mean of 0 and a variance of , respectively. Q and R The Gaussian distribution. For process noise, representing the instantaneous flow field disturbances in the part not represented by the GAT model; To measure noise, this represents the measurement error of the SCADA anemometer including wake interference.

[0086] The Kalman filter algorithm consists of two processes: prediction and update. In the prediction phase, the algorithm directly uses the dynamic model (in this case, the GAT output) to provide a prior estimate of the current state. In the update phase, the algorithm first calculates the Kalman gain. K kThis parameter determines the degree of confidence in the anemometer measurement residuals, and then the prior estimate is weighted and corrected using the current measurement information. By constructing a minimum mean square error estimate between the physical prediction value and the real-time measurement value, the optimal posterior estimate value is obtained, and its update formula is shown in Equation (5).

[0087] (5)

[0088] In the formula, for k The effective capture wind speed estimated by Kalman filtering at each moment. Here, it is used as a priori state estimate. K k This represents the Kalman gain. For a one-dimensional variable like wind speed, it can be expressed as:

[0089] (6)

[0090] In the formula, The covariance of the state estimation error at the previous time step. Let be the prior covariance at the current moment. K k The magnitude of reflects the estimation uncertainty during the model's prediction phase.

[0091] Effectively capturing wind speed as an intermediate variable is passed to the next-level power prediction model (PI-Informer) as its core input feature. Simultaneously, this value can also be output to the monitoring interface, allowing maintenance personnel to intuitively assess the flow field status.

[0092] Thus, the GAT model dynamically determines the unit connection relationship to generate an adjacency matrix using prior knowledge of fluid dynamics (wake centerline and width), and adaptively calculates the influence weight of upstream units on downstream unit wakes using a multi-head attention mechanism. Furthermore, Kalman filtering is used to fuse GAT physical calculations with SCADA measured values ​​to reconstruct a high-precision "effective wind speed," thereby eliminating sensor noise and correcting physical model biases. Simultaneously, a parameter fine-tuning strategy based on transfer learning enables the model to automatically correct for data distribution differences caused by seasonal variations or equipment aging over time, ensuring the robustness of the control system throughout its entire lifecycle.

[0093] Combination Figure 6 As shown, in order to meet the requirements of MPC control for real-time performance (low latency) and long sequence prediction, this embodiment constructs an Informer neural network based on physical information, specifically including:

[0094] (1) Input:

[0095] Environmental characteristics: Effective wind speed output from the previous GAT model and corrected by Kalman filtering.U i；

[0096] Unit status sequence: The sequence of generator speed, torque, rotor speed, pitch angle, and yaw angle over a past period;

[0097] The effective wind speed is obtained from the GAT model, and the historical status of the unit is obtained from the SCADA database.

[0098] (2) Model structure

[0099] 1) Model architecture: The encoder-decoder structure is adopted.

[0100] 2) Innovation of the Informer model (low-complexity inference): To address the high computational complexity of traditional Transformer models, Informer introduces a probabilistic sparse self-attention mechanism. This mechanism uses KL divergence to select the query vectors that contribute most to the prediction, and only calculates the dot product of these dominant queries with the key. This significantly reduces time complexity, enabling the model to complete parallel predictions of the power of all 46 generator units for the next hour (prediction step size of 10 minutes) within milliseconds, thus supporting the MPC algorithm to complete multiple iterations of optimization within a limited time.

[0101] 3) Generative decoding: A one-step generative decoder is used to directly output the power sequence for all future moments, avoiding the error accumulation caused by traditional stepwise regression.

[0102] 4) Innovation of the Physical Information-Informer Model (PI-Informer): Introducing physical constraints based on the standard power curve into the loss function. The construction of the loss function includes: calculating the mean square error between the predicted power generation value and the actual monitored power value of the wind turbine at a single time step, and determining the average mean square error of all training samples as the data-driven loss. Based on the pre-constructed standard power curve function of the turbine, the theoretical power value corresponding to the current input wind speed is calculated; the mean square error between the predicted power generation value and the theoretical power value is calculated, and the average mean square error of all training samples is determined as the physical prior loss. The backpropagation gradient norm of the data-driven loss and the physical prior loss is calculated in real time using the gradient normalization algorithm (GradNorm), and the data-driven weights and physical prior weights are dynamically adjusted according to the gradient changes. The product of the data-driven loss and the data-driven weights, plus the product of the physical prior loss and the physical prior weights, is determined as the total loss.

[0103] The calculation process is shown in equation (7). This ensures that the model prediction results conform to the physical characteristics of energy conversion of wind turbines, preventing outliers that violate the unit's design power curve.

[0104] (7)

[0105] In the formula, 1 and 2 represents the weights for the power curve and data loss, respectively. The first calculation for the model i Predicted power of each unit , The actual power recorded by the unit. The rated power of the unit is obtained by fitting discrete data points of this unit model.

[0106] The fitting method adopts the piecewise cubic Hermite interpolation polynomial (PCHIP) method, which includes: obtaining discrete design parameter points for each wind turbine, and constructing a nonlinear fitting function for the partial load area using the piecewise cubic Hermite interpolation polynomial. When the input wind speed is less than the cut-in wind speed or greater than the cut-out wind speed, the theoretical power value is set to zero. When the input wind speed is greater than or equal to the cut-in wind speed and less than or equal to the rated wind speed, the value corresponding to the current input wind speed is calculated using the nonlinear fitting function, and this value is determined as the theoretical power value. When the input wind speed is greater than the rated wind speed and less than or equal to the cut-out wind speed, the rated power of the wind turbine is determined as the theoretical power value. Specifically, as shown in equation (8).

[0107] (8)

[0108] In the formula, Input wind speed; , and These refer to the unit's input, output, and rated wind speed, respectively. Rated power; This represents a nonlinear fitting function constructed based on the PCHIP method in the partially loaded region.

[0109] (3) Output

[0110] Every unit in the field in the future N p Power generation sequence at each time point P i The prediction result is directly input into the MPC optimizer as the objective function for computation. J The basis for this.

[0111] Thus, the Informer architecture offers low computational complexity. The PI-Informer model employs a probabilistic sparse self-attention mechanism to reduce the complexity of long sequence predictions, achieving millisecond-level inference, and embeds standard power curve (PCHIP fitting) physical constraints into the loss function to ensure physical consistency of prediction results. Experimental results demonstrate that the average time for a single full-field state prediction is only 0.712 ms, enabling the differential evolution algorithm to complete thousands of iterations of optimization within a finite control cycle, ensuring rapid response in cluster control.

[0112] Combination Figure 7 As shown, to address the decline in model accuracy caused by equipment aging and seasonal wind changes, this invention proposes a transfer learning strategy based on parameter freezing and fine-tuning, specifically including:

[0113] (1) Pre-training (offline): Using historical unit and radar data, or full-condition datasets generated by CFD and FLORIS simulations, to train all parameters of the integrated model ( and Pre-training should be conducted to enable them to master the general physical laws of flow fields;

[0114] (2) Online fine-tuning (real-time): When the prediction error monitoring mechanism triggers an update signal (i.e., the wake prediction error, wind speed prediction error, and / or power prediction error exceed the preset threshold), the system initiates online transfer learning. At this time, the backbone network parameters are maintained ( (e.g., the Attention layer) is frozen and remains unchanged, using only the most recent measured data to adjust the output layer parameters ( Fine-tuning and updating can be performed on layers such as fully connected layers.

[0115] This strategy preserves the model's expression of complex physical laws while rapidly updating and correcting the errors of each model through low-cost computation, thus achieving robust operation of the control system throughout its entire lifecycle and solving the problem of data distribution drift caused by equipment aging and seasonal wind changes.

[0116] Performance verification of the method provided in this embodiment:

[0117] 1. The predictive advantages of the GAT model in handling strong wake interference.

[0118] Figure 8 The paper presents a comparison of wind speed predictions for a turbine located in a severe wake region under different models. It can be seen that under complex dynamic wake conditions, the traditional physical model (Jensen) fails, and the ordinary time series model (SVR / LSTM) fluctuates significantly. However, the GAT model (red line) provided in this embodiment closely follows the actual value, proving that it can fully express the mutual influence between multiple turbines.

[0119] 2. Tracking performance of the PI-Informer model under power fluctuations

[0120] Figure 9 This demonstrates the tracking performance of different prediction models on generator power output during periods of drastic wind speed fluctuations. Compared to Transformer and Informer models, which lack physical information, the PI-Informer model (orange line) used in this embodiment exhibits the smallest lag during periods of sharp power increases and decreases, and the highest goodness of fit during periods of stable performance, proving its stability in long-term predictions.

[0121] 3. Improvement effect of MPC cluster control under strong wake conditions

[0122] Figure 10 The results demonstrate the power gain rate of the MPC strategy compared to the traditional MPPT strategy under wind speeds near the rated wind speed of 11 m / s, as a function of wind direction. At strong wake directions such as 150° (southeast), through coordinated control, the overall power gain rate reached 5.38%, effectively mitigating energy loss caused by wake obstruction.

[0123] 4. Economic benefit assessment of long-term operation throughout the year

[0124] Figure 11 Statistical results based on year-round measured data are presented. The MPC strategy achieved significant power increases in both the cut-off wind speed range and the full-power transition zone (around 11 m / s). The cumulative annual power generation increased by 0.95%, validating the long-term value of this method in practical engineering applications.

[0125] Combination Figure 12 As shown in the figure, this disclosure provides a wind farm wake optimization device 120 based on model predictive control, including: a data acquisition module 121, a comprehensive prediction module 122, an optimization decision module 123, and an execution feedback module 124. The data acquisition module 121 is configured to collect the operating state variables and meteorological environment information of each wind turbine in the entire wind farm. The comprehensive prediction module 122 is configured to input the operating state variables and the meteorological environment information into a comprehensive prediction model; the comprehensive prediction model includes: a PI-DANN model with embedded NS equations and vorticity equations as physical constraints, a GAT model, and a PI-Informer model with physical constraints based on the standard power curve introduced into the loss function. The PI-DANN model generates an adjacency matrix of the wind farm, the GAT model determines the effective capture wind speed of each wind turbine based on the adjacency matrix, and the PI-Informer model outputs the future wind speed of each wind turbine in the entire wind farm based on the effective capture wind speed. N pThe power generation sequence at each time point. The optimization decision module 123 is configured to construct the objective function and constraints, and combine them with the future... N p The power generation sequence at each moment determines the future power output of each wind turbine. N p The sequence of yaw angles at each moment. The execution feedback module 124 is configured to, based on the future... N p The value of the first moment in the yaw angle sequence controls the wind turbine to perform yaw action, and forms a rolling closed-loop control according to the rolling time-domain optimization mechanism.

[0126] The wind farm cluster wake optimization device based on model predictive control provided in this disclosure utilizes a PI-DANN model to dynamically generate an adjacency matrix conforming to hydrodynamics, combines a GAT model graph attention mechanism to accurately describe the influence relationships between turbines, and then achieves long-sequence prediction through a PI-Informer model. By constructing a spatiotemporal integrated model that fuses the graph attention network (GAT) and physical information informer (PI-Informer), it can quickly calculate the wind farm power changes over a future period and accurately describe the wake interference relationships between turbines. Based on this, combined with a model predictive control (MPC) framework, it achieves real-time rolling optimization of turbine yaw angles, thereby improving the total power generation and economic benefits of the wind farm cluster while ensuring operational safety. This architecture retains the accuracy of the physical mechanism and possesses the nonlinear fitting capability of deep learning, which can improve the accuracy of overall power prediction and balance high-precision physical field description with rapid response of cluster control.

[0127] The technical solutions of this disclosure can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in this disclosure. The aforementioned storage medium can be a non-transitory storage medium, such as a USB flash drive, external hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, etc., and other media capable of storing program code.

[0128] The foregoing description and accompanying drawings fully illustrate embodiments of this disclosure to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. Without further limitations, an element defined by the phrase "comprising a…" does not exclude the presence of additional identical elements in the process, method, or apparatus that includes said element. Throughout this document, each embodiment may focus on its differences from other embodiments, and similar or identical parts between embodiments may be referred to mutually. For methods, products, etc., disclosed in the embodiments, if they correspond to the method section disclosed in the embodiments, then the relevant parts may be referred to the description of the method section.

[0129] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this disclosure. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0130] The methods and products disclosed in the embodiments herein (including but not limited to devices and equipment) can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units may be merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection of devices or units may be electrical, mechanical, or other forms. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to implement this embodiment according to actual needs. In addition, the functional units in the embodiments of this disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0131] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. In the descriptions corresponding to the flowcharts and block diagrams in the accompanying drawings, the operations or steps corresponding to different blocks may also occur in a different order than disclosed in the description, and sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. Each block in a block diagram and / or flowchart, and combinations of blocks in a block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

Claims

1. A method for optimizing the wake of a wind power cluster based on model predictive control, characterized in that, include: S10, Data Acquisition: Collect operational status variables and meteorological environment information for each wind turbine unit in the entire site; S20, Comprehensive Prediction: The operating state variables and the meteorological environment information are input into the comprehensive prediction model; the comprehensive prediction model includes: a PI-DANN model with embedded NS equations and vorticity equations as physical constraints, a GAT model, and a PI-Informer model with physical constraints based on the standard power curve introduced into the loss function. The PI-DANN model generates the adjacency matrix of the wind farm, the GAT model determines the effective wind speed captured by each wind turbine based on the adjacency matrix, and the PI-Informer model outputs the future wind speed of each wind turbine in the entire farm based on the effective wind speed. N p The power generation sequence at each moment; S30, Optimization Decision: Construct the objective function and constraints, and combine them with the future... N p The power generation sequence at each moment determines the future power output of each wind turbine. N p The sequence of yaw angles at each moment; S40, Execution Feedback: Based on the stated future N p The value of the first moment in the yaw angle sequence at each moment controls the wind turbine to perform yaw action, and forms a rolling closed-loop control according to the rolling time-domain optimization mechanism; In S20, the GAT model determines the effective wind speed captured by each wind turbine based on the adjacency matrix, including: Each wind turbine is defined as a node in the graph, and the wake influence relationship between each wind turbine is defined as a directed edge in the graph. Introducing a multi-head attention mechanism to automatically calculate upstream nodes j For downstream nodes i Influence weight α ij : ， in, represents the node feature vector; || denotes the vector concatenation operation. Here, represents the weight vector of a single-layer feedforward neural network, LeakyReLU is a non-linear activation function, and Softmax is a normalized exponential function. Based on the distance and yaw angle between each wind turbine, the wake influence weight is adaptively allocated, and the physical wind speed is output. The physical wind speed and the measured wind speed in the nacelle are weighted and fused using the Kalman filter algorithm to obtain the effective capture wind speed for each wind turbine. The step of using the Kalman filter algorithm to weightedly fuse the physical wind speed and the measured wind speed in the nacelle to obtain the effective capture wind speed for each wind turbine includes: Will k Real-time, effective wind speed capture is defined as a state variable. The output value of the GAT model As a priori prediction, nacelle wind speed is defined as a measurement variable. The state equation and the equation for the measured variable are defined as follows: ， in, N (0, Q )and N (0, R ) represent a mean of 0 and a variance of , respectively. Q and R Gaussian distribution; The process noise represents the instantaneous flow field disturbances in the part not characterized by the GAT model. This represents the measurement error of the anemometer including wake interference, used to measure noise. ， In the formula, for k The effective capture wind speed estimated by Kalman filtering at each moment. Here, it is used as a priori state estimate. This is the Kalman gain.

2. The wind power cluster wake optimization method based on model predictive control according to claim 1, characterized in that, The construction of the loss function of the PI-Informer model includes: The mean square error between the predicted power generation of the wind turbine and the actual monitored power generation at a single time step is calculated, and the average mean square error of all training samples is determined as the data-driven loss. Based on the pre-constructed standard power curve function of the unit, the theoretical power value corresponding to the current input wind speed is calculated; the mean square error between the predicted power value and the theoretical power value is calculated, and the average mean square error of all training samples is determined as the physical prior loss. The backpropagation gradient norms of the data-driven loss and the physical prior loss are calculated in real time using the gradient normalization algorithm, and the data-driven weights and physical prior weights are dynamically adjusted according to the gradient changes. The total loss is determined by adding the product of the data-driven loss and the data-driven weights to the product of the physical prior loss and the physical prior weights.

3. The wind power cluster wake optimization method based on model predictive control according to claim 1, characterized in that, S20 also includes: The output layer parameters of the comprehensive prediction model are updated using the latest measured data, while the backbone network parameters are kept frozen.

4. The wind power cluster wake optimization method based on model predictive control according to claim 3, characterized in that, The step of updating the output layer parameters of the comprehensive prediction model using the most recent measured data includes: In the flow field correction loop, the wake prediction error of the PI-DANN model is monitored in real time. The wind speed prediction error of the GAT model ; In the power correction loop, the power prediction error of the PI-Informer model is monitored in real time. ; When any one of the wake prediction error, the wind speed prediction error, or the power prediction error continuously exceeds a preset threshold, the output layer parameters of the corresponding model are updated using the latest measured data.

5. The wind power cluster wake optimization method based on model predictive control according to claim 2, characterized in that, The theoretical power value of each wind turbine is determined in the following way: The discrete design parameter points of each wind turbine are obtained, and a nonlinear fitting function for the partial load region is constructed using a piecewise cubic Hermite interpolation polynomial. When the input wind speed is less than the cut-in wind speed or greater than the cut-out wind speed, the theoretical power value is set to zero. When the input wind speed is greater than or equal to the cut-in wind speed and less than or equal to the rated wind speed, the nonlinear fitting function is used to calculate the value corresponding to the current input wind speed, and this value is determined as the theoretical power value. When the input wind speed is greater than the rated wind speed and less than or equal to the cut-out wind speed, the rated power of the wind turbine is determined as the theoretical power value.

6. The wind power cluster wake optimization method based on model predictive control according to claim 1, characterized in that, In S20, generating the adjacency matrix of the wind farm using the PI-DANN model includes: The wake envelope of each wind turbine was calculated using the PI-DANN model. The existence of a wake-shading relationship between any two wind turbines is determined based on the wake envelope, and then the adjacency matrix is ​​generated.

7. The wind power cluster wake optimization method based on model predictive control according to any one of claims 1 to 6, characterized in that, In S30, the construction of the objective function and constraints includes: To predict the time domain N p The goal is to maximize the total power generation of the entire site. The constraints include: (1) The yaw angle of each wind turbine is limited to a safe range; (2) The change in yaw angle between adjacent time points is less than or equal to the maximum permissible yaw rate in a single step; (3) The predicted power is within a reasonable range that is non-negative and does not exceed the rated power.

8. A wind power cluster wake optimization device based on model predictive control, characterized in that, include: The data acquisition module is configured to collect the operating status variables and meteorological environment information of each wind turbine in the entire field; The integrated forecasting module is configured to input the operating state variables and the meteorological environment information into the integrated forecasting model; The integrated prediction model includes: a PI-DANN model with embedded Navier-Stokes equations and vorticity equations as physical constraints, a GAT model, and a PI-Informer model with physical constraints based on the standard power curve incorporated into the loss function. The PI-DANN model generates the adjacency matrix of the wind farm; the GAT model determines the effective wind speed captured by each wind turbine based on the adjacency matrix; and the PI-Informer model outputs the future wind speed of each wind turbine in the entire farm based on the effective wind speed. N p The power generation sequence at each moment; The optimization decision module is configured to construct the objective function and constraints, and combine them with the future... N p The power generation sequence at each moment determines the future power output of each wind turbine. N p The sequence of yaw angles at each moment; The execution feedback module is configured to, based on the future N p The value of the first moment in the yaw angle sequence at each moment controls the wind turbine to perform yaw action, and forms a rolling closed-loop control according to the rolling time-domain optimization mechanism; The GAT model determines the effective wind speed captured by each wind turbine based on the adjacency matrix, including: Each wind turbine is defined as a node in the graph, and the wake influence relationship between each wind turbine is defined as a directed edge in the graph. Introducing a multi-head attention mechanism to automatically calculate upstream nodes j For downstream nodes i Influence weight α ij : ， in, represents the node feature vector; || denotes the vector concatenation operation. Here, represents the weight vector of a single-layer feedforward neural network, LeakyReLU is a non-linear activation function, and Softmax is a normalized exponential function. Based on the distance and yaw angle between each wind turbine, the wake influence weight is adaptively allocated, and the physical wind speed is output. The physical wind speed and the measured wind speed in the nacelle are weighted and fused using the Kalman filter algorithm to obtain the effective capture wind speed for each wind turbine. The step of using the Kalman filter algorithm to weightedly fuse the physical wind speed and the measured wind speed in the nacelle to obtain the effective capture wind speed for each wind turbine includes: Will k Real-time, effective wind speed capture is defined as a state variable. The output value of the GAT model As a priori prediction, nacelle wind speed is defined as a measurement variable. The state equation and the equation for the measured variable are defined as follows: ， in, N (0, Q )and N (0, R ) represent a mean of 0 and a variance of , respectively. Q and R Gaussian distribution; The process noise represents the instantaneous flow field disturbances in the part not characterized by the GAT model. This represents the measurement error of the anemometer including wake interference, used to measure noise. ， In the formula, for k The effective capture wind speed estimated by Kalman filtering at each moment. Here, it is used as a priori state estimate. This is the Kalman gain.

Images