Wind power prediction method considering wake effect and wind speed-power conversion mechanism

Abstract

A wind power prediction method considering a wake effect and a wind speed-power conversion mechanism comprises the following steps: firstly, collecting 2SCADA data and event data of each fan of a wind power plant, and then representing preprocessed information such as natural NWP wind speed and wind direction of the wind power plant, layout of the wind power plant, fan characteristics and the like as graphs; the wind power plant graph is input into the PGNN model, the PGN layer sequentially carries out edge updating, edge weight calculation, node updating and global updating to generate an updated graph, the updated graph is input into the full connection layer, a wind speed graph is output, node values represent estimated wind speeds of all wind turbines in the wind power plant, and prediction of the wind speeds at the wind turbine hubs in the future time period is achieved. And finally, inputting the predicted wind speed of the fan into an embedded neural network constructed based on a wind speed-power conversion mechanism to obtain the predicted power output of the single fan and the whole wind power plant, thereby effectively improving the interpretability and accuracy of the ultra-short-term wind power prediction model.

Description

Technical Field

[0001] This invention relates to the field of power grid dispatching technology, and in particular to renewable energy generation and integrated consumption technology. Specifically, it relates to an ultra-short-term wind power prediction method that considers wake effect and wind speed-power conversion mechanism, and a method for processing wind farm maps and obtaining predicted wind speed. Background Technology

[0002] As wind energy plays an increasingly important role in the global energy mix, accurate wind power forecasting is crucial for stable grid operation and optimized energy management. In particular, ultra-short-term wind power forecasting (0-4 hours in advance) directly impacts real-time grid dispatch and supply-demand balance. However, the randomness, volatility, and non-stationarity of wind speed present significant challenges. Traditional wind power forecasting methods typically focus only on the total power of the wind farm, neglecting the interactions between individual turbines. In reality, wake effects lead to uneven spatial distribution of wind speed within the wind farm, reducing the power generation efficiency of downstream turbines. By forecasting the wind speed at the hub of each turbine from within the wind farm, and then converting the forecasted wind speed into corresponding power output, a more detailed reflection of the dynamic changes in turbine output can be achieved, thereby improving the accuracy and reliability of wind power forecasting.

[0003] Most wind farms in my country consist of dozens to hundreds of wind turbines, which are geographically concentrated, resulting in strong correlations in wind speed and power among them. Therefore, the wind speed and power of each turbine within a wind farm are not only autocorrelated in time but also spatially correlated with adjacent turbines, and are influenced by the overall geographical location of the wind farm and external environmental conditions (such as air pressure and temperature). In particular, the wake effect of upstream turbines reduces the wind speed of downstream turbines, increasing the uncertainty of wind speed prediction and thus affecting the accuracy of power prediction. Therefore, in multi-turbine power prediction, considering the interaction between turbines and mining the complex spatiotemporal correlations in massive wind power data is of great significance for improving the accuracy of short-term wind power prediction. Existing technologies have proposed several methods, such as using Principal Component Analysis (PCA) to identify and process the vertical strata wind characteristics simulated by Numerical Weather Prediction (NWP) to improve the quality of wind power prediction under complex terrain. However, PCA can only capture linear relationships, while the relationships between features in wind power forecasting are usually non-linear. With the rapid development of artificial intelligence technology, deep learning technology has been widely applied in the field of wind power forecasting due to its powerful feature extraction capabilities and ability to process massive amounts of data. For example, convolutional neural networks (CNNs) are used to extract spatial features for the spatial correlation between adjacent sites of a wind farm. However, the convolutional kernel parameters remain fixed after training and do not change over time, which may lead to insufficient adaptability of CNNs to the dynamic changes in spatial correlation caused by wind direction changes, and this method ignores temporal correlation. In reality, however, the spatial and temporal correlations of wind speed and power of each wind turbine in a wind farm are coupled. In recent years, some new deep learning models, such as path-based graph neural networks (PGNNs), have significantly improved the ability to learn complex relational features under wake effects by combining graph structure modeling, path information extraction, and attention mechanisms. Such technologies provide strong technical support for accurately extracting the dynamic spatial features of each wind turbine within a wind farm under the complex flow characteristics of wind farms caused by high-noise wind speeds and wake effects.

[0004] After obtaining the wind speed prediction, the wind power prediction needs to be further obtained. The relationship between the wind turbine's output power and the wind speed at the hub is usually described by the wind power characteristic curve. However, due to environmental factors such as air density, turbulence intensity, and wind shear, there is a difference between the actual wind turbine output and the calculation results based on the standard characteristic curve provided by the manufacturer. In order to accurately reflect the actual wind speed-power conversion performance of the wind turbine, it is necessary to further combine actual operating data and external environmental factors, and use data-driven methods to fit the mapping relationship between wind speed and power. In this research, parametric and nonparametric wind speed-power curve modeling methods based on statistical laws are often used, such as modeling based on Bean's method and cubic spline interpolation techniques. Although these are intuitive and easy to implement, they are not effective when faced with complex nonlinear relationships and variable environmental factors. To improve the accuracy and adaptability of modeling, researchers have begun to turn to data-driven methods based on machine learning, such as building wind turbine power characteristic curve models based on fuzzy modeling techniques and artificial neural networks. Although these methods can better fit the power curve model after being trained by neural networks, the predictive ability of the power curve model may be affected when faced with abnormal fluctuations that have not appeared in the training dataset. Therefore, it is possible to embed the physical mechanism of the wind speed-power curve into the neural network. This way, even when faced with diverse data types, more flexible and accurate predictions can be achieved through the self-learning of the neural network and the embedded physical mechanism.

[0005] Based on the above analysis, this invention focuses on two key issues in the process of refined wind farm power prediction: the evolution law of dynamic spatiotemporal characteristics of wind speed and the adaptive embedding of wind speed-power conversion mechanism. It proposes an ultra-short-term wind power prediction technology that considers wake effect and wind speed-power conversion mechanism. Summary of the Invention

[0006] The purpose of this invention is to address the technical problem that traditional wind power forecasting methods often neglect the influence of the spatiotemporal characteristics of wind speed in wind farms and fail to establish a refined dynamic mapping relationship between wind speed and power output, resulting in poor performance and accuracy of wind power forecasting, especially ultra-short-term wind power forecasting.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for ultra-short-term wind power prediction based on wake effect and wind speed-power conversion mechanism includes the following steps: Step 1: Collect data from the wind farm m Typhoon machine currently t Looking back H A historical moment r Features (such as historical data on wind farm wind speed and pitch angle) and future performance of each wind turbine kThe forecast time includes NWP data (such as wind speed, wind direction, and temperature), as well as information on wind farm layout and turbine characteristics. Step 2: Calculate the physical model using the data obtained in Step 1 to generate high-precision physical simulation results. These results will serve as node information for the graph. Then, normalize the calculation results of the physical model with the input data to construct the graph structure. Obtain the dataset and divide it into three subsets according to a certain ratio for training, validation, and testing, respectively. Step 3: Combine the data and information obtained from Steps 1 and 2 to form node features, edge features, and global features, and construct a directed graph. ; directed graph As input to the PGNN model, wind speed prediction is performed on the wind turbines, and the output is a map showing the predicted wind speed values ​​at each wind turbine hub. The process of generating wind speed predictions using the PGNN model is explained. Step 4: Correct the predicted wind speed values ​​for each fan obtained in Step 3 by considering the influence of ambient temperature; then use the corrected wind speed as input data to perform wind speed-power conversion, and output the corrected predicted fan power value. Repeat this operation until each wind turbine and each prediction time step has been traversed, and output the predicted power value for each wind turbine. ; Step 5: Analyze the predicted power results for each wind turbine obtained in Step 4. The total predicted power of the wind farm is then superimposed and output.

[0008] In step 3, when obtaining the predicted wind speed values ​​at each wind turbine hub, the specific steps include: Step 3-1: Collect the previous... H Historical data such as wind speed and pitch angle at a specific historical moment in a wind farm. and each fan k The data for the predicted time period are mostly meteorological forecasts. (Such as wind speed, wind direction, temperature, etc.) are preprocessed to generate a directed graph. ; Step 3-2: Directed graph The input is fed into a PGNN network for ultra-short-term wind turbine wind speed prediction, and the output is a map of the predicted wind speed values ​​at each turbine hub. The node values ​​represent the estimated wind speeds of each wind turbine in the wind farm, so... It can also be expressed as .

[0009] Step 3-1 specifically includes the following steps: 3-1-1) Collect parameters such as wind turbine location, height, blade length, rated wind speed, real-time wind speed, power, wake intensity, and NWP wind speed of the wind farm; 3-1-2) Each wind turbine corresponds to a node; these are used as features of the directed graph to construct the graph structure; node features This includes attributes such as the wind turbine's location, height, blade length, and rated wind speed. These attributes are crucial for describing the wind turbine's basic characteristics and operating status; side features Indicates from node j (Sender) to Node i The properties of the edges of the (receiver) include parameters such as real-time wind speed, power, and wake intensity, which mainly reflect the wind turbine. j To the fan i Wake characteristics; edge features are represented as ,in, d ij For two wind turbines i and j The downstream wake distance between them r ij Radial wake distance, h ij Vertical wake distance; global feature g Let be a scalar or vector, describing the overall properties of the wind farm, where g The NWP wind speed is used directly to characterize the intensity of environmental energy input.

[0010] In step 3-2, the PGNN network consists of stacked Physics-Guided Networks (PGNs) and fully connected layers (which are constructed from a single MLP). Specifically, the structure is: directed graph... As input to the stacked PGN layers, the final output of the stacked PGN layers is connected to the input of the fully connected layer.

[0011] The process of calculating and updating wind farm maps using the PGN layer includes the following steps: (1) The input graph is processed by a Path-based Graph Neural Network (PGNN) based on physical mechanisms, and the wind speed of all wind turbines is estimated as the output graph. PGNN is a message-passing graph neural network, which consists of stacked Physically Induced Graph Networks (PGNs) and fully connected layers. Each PGN layer updates the edges, nodes, and graph features of the wind farm graph by sending and updating information between nodes in the wind farm graph, so as to estimate the power generation of the wind turbines. In the message-passing process, the PGN layer utilizes four differentiable functions: edge update function Edge weight function Node update function and global update function The detailed calculation process of PGN is as follows: Edge update function Update all existing edges in the graph. The edge update function utilizes the characteristics of the sender node. Characteristics of receiver nodes and the edge features between two nodes Generate updated edge features The PGN layer utilizes a weight calculation function. Adjust dynamically based on the relative position of the wind turbine and wind conditions. The weighting; here, the downstream wake distance is selected. Radial wake distance and vertical wake distance As The input is used to calculate the adjustment. weights Using these weights, PGN adjusts the influence strength between turbines to ensure that the calculated weights are physically reasonable; the entire edge update process is shown below: ; Node update function Utilizing the features of the current input node Updated edge features The node features are updated using the global feature g; PGN blocks use an edge aggregation function. Simplify the edge feature set into a single aggregated edge feature. Then, the node update function. Processing the aggregated edge features and current node features Generate updated node features The edge aggregation and node update process is as follows: ; global update function To update global features, PGN uses the updated edge and node features. To do this, PGN utilizes... and All edge features E and all node features Aggregate into two vectors , Global features g The update process is as follows: ; in, , and It is a multilayer perceptron (MLP); in addition to other encoders, it uses physically induced basis functions to construct... Edge weight function The “strength” of the interaction between turbines is adjusted in a manner similar to that used in physical-induced engineering models to describe the relationship between two turbines. Furthermore, the GCH three-dimensional wake model is adopted as the physical mechanism part, and it is used as... Basis functions, applying physics-based weighting functions to the proposed PGN: ; In the formula, As an inducing factor, The rotor radius is... This refers to the surface roughness coefficient. (2) By sequentially applying these update functions to the graph input (i.e., the output graph of the previous PGN layer or the directed graph of the wind farm), G The PGN (Programme Generation Network) generates an updated graph. After a series of PGN layer calculations, the PGNN generates an updated wind farm graph. Subsequently, the fully connected layer uses this graph as input and outputs a wind speed graph for the wind turbine. The node values ​​represent the estimated wind speed of each wind turbine in the wind farm. The fully connected layer is constructed from a single MLP, which is typically used to estimate the wind speed of each wind turbine in the wind farm. The design of PGNN allows the computation process in the PGN layer to mimic the computation process of engineering-based wake models: 1) The edge weight calculation in the PGN layer directly follows the calculation of the tail flow loss factor. The calculation process is as follows: ; 2) The edge update step in the PGN layer corresponds to the computation of the wind turbine. i and j Between loss factors As shown in the following formula: ; 3) The edge aggregation step in the PGN layer corresponds to the wake defect factor aggregation step, which is used to calculate the wind turbine. i loss factors As shown in the following formula: ; 4) The node update step in the PGN layer corresponds to calculating the average wind speed using the aggregated defect factor. As shown in the following formula: ; Through the above steps, each PGN layer can send and update messages between nodes in the wind farm graph, thereby updating the edges, nodes, and graph features of the wind farm graph. The updated wind farm graph is then... Input into a fully connected layer to estimate the wind turbine wind speed map. This enables a complete process from data input to wind speed prediction.

[0012] In step 4, the corrected wind turbine power prediction value is obtained. Includes the following steps: Step 4-1: Divide the wind turbine into four zones according to its operating status: the start-up zone with no output power, the wind speed-power conversion zone where power increases linearly to the rated power, the constant power zone where the rated power is maintained after exceeding the rated wind speed, and the safe zone where the rated power is maintained after exceeding the rated wind speed. Step 4-2: Considering the influence of ambient temperature, the predicted wind speed values ​​of each fan output in Step 3-2 are... Each predicted wind speed and the corresponding temperature prediction value The predicted wind speed is then corrected by using the input of the embedded network for wind speed-to-power conversion, and a Dense layer is applied to obtain the corrected wind speed. Then according to Cut-in wind speed Rated wind speed Cut-off wind speed Relationship, calculate wind speed Power weights falling into the corresponding interval ,in, For wind speed at The weight of the region; Step 4-3: Use the formula and formula Using the standardized wind speed and its polynomial combination as input, a three-layer fully connected neural network model is employed, based on the formula... Predicted to be located in the wind speed-power conversion zone Internal power output ; in, This is the normalized wind speed value; It is a collection , , The vector is used as the input to the neural network; , , It is the weight matrix of each hidden layer in the neural network; , , This refers to the bias of each hidden layer in a neural network; It is input The weighted summation represents the input. With weight matrix The product, plus the bias In a multi-layer neural network, the output of the previous layer becomes the input of the next layer. This indicates that the output of the first hidden layer, after being processed by the ReLU activation function, is compared with the weights of the second layer. Multiply, and add the bias of the second layer. ; This indicates that the output of the second hidden layer, after being processed by the ReLU activation function, is compared with the weights of the third layer. Multiply, and add the bias of the third layer. ; Activated by the Sigmoid function Limit the output to the range [0,1], multiply by The predicted normalized power is then converted into the actual power output. Step 4-4: Weight the power according to the wind speed region to obtain the predicted value of the wind turbine output power at the predicted time. ; Steps 4-5: Power correction for various factors is achieved by constructing two fully connected layers to output the final power of the wind turbine. .

[0013] In step 4-2, the formula is used. and formula Calculate wind speed Power weights falling into the corresponding interval ;in, For wind speed at The weight of the region; This is a parameter of the LeakySigmoid function, used to control the slope of the function. It is set according to the width of the transition region of the statistical sample, reflecting... The strength of the relationship with the corresponding threshold.

[0014] In step 4-4, the formula is used. The predicted output power of the wind turbine at the predicted time is obtained. ;in, In order to be in The power output value of the region; In order to be in The power output value of the region; In order to be in The power output value of the region; In order to be in The power output value of the region.

[0015] In steps 4-5, power correction for various factors is achieved by constructing two fully connected layers, resulting in the final output power of the wind turbine. The process is as follows: ; ; ; The fully connected layer in the expression is used to handle non-deterministic factors such as blade icing thickness, bearing wear, and blade contamination coefficient. These factors have a certain impact on power, but usually do not cause the wind turbine power output to be completely zero. The fully connected layer in the expression handles decisive factors such as mechanical failures and grid power rationing, which directly affect the start-up and shutdown status of the wind turbine; among which... r and u Let represent the eigenvectors of non-deterministic and deterministic factors, respectively. f and g This is the corresponding activation function.

[0016] In step 5, the wind farm predictive power integration network: superimposes the predicted power outputs of all wind turbines, according to the formula... Total predicted power output of wind farm , This refers to the predicted power output of a single wind turbine.

[0017] The present invention also includes a method for processing a wind farm map and obtaining a predicted wind speed. It can obtain a directed graph related to the wind turbine and wind farm by acquiring several data related to the wind turbine and wind farm, and then process the directed graph using the PGNN model proposed in this invention to obtain a predicted wind speed value.

[0018] Compared to existing wind speed prediction technologies, this method can obtain wind speed predictions more accurately and efficiently, thus making a technological contribution to the field of wind speed prediction technology.

[0019] Specifically, the method includes the following steps: Step 1: Collect the previous H Historical data such as wind speed and pitch angle at a specific historical moment in a wind farm. and each fan k The data for the predicted time period are mostly meteorological forecasts. A directed graph is generated after preprocessing. ; Step 2: Directed graph The input is fed into a PGNN network for ultra-short-term wind turbine wind speed prediction, and the output is a map of the predicted wind speed values ​​at each turbine hub. The node values ​​represent the estimated wind speeds of each wind turbine in the wind farm, so... It can also be expressed as .

[0020] Step 1 specifically includes the following steps: 1-1) Collect parameters such as wind turbine location, height, blade length, rated wind speed, real-time wind speed, power, wake intensity, and NWP wind speed of the wind farm; 1-2) Each wind turbine corresponds to a node; these are used as features of the directed graph to construct the graph structure; node features This includes attributes such as the wind turbine's location, height, blade length, and rated wind speed. These attributes are crucial for describing the wind turbine's basic characteristics and operating status; side features Indicates from node j Indicates the sender to the node i The attributes of the edges representing the receiver include parameters such as real-time wind speed, power, and wake intensity, which mainly reflect the wind turbine. j To the fan i Wake characteristics; edge features are represented as ,in, d ij For two wind turbines i and j The downstream wake distance between them r ij Radial wake distance, h ij Vertical wake distance; global feature g Let be a scalar or vector, describing the overall properties of the wind farm, where g The NWP wind speed is used directly to characterize the intensity of environmental energy input.

[0021] In step 2, the PGNN network consists of stacked Physically Induced Graph Networks (PGNs) and fully connected layers, wherein the fully connected layers consist of a single MLP, specifically a directed graph. As input to the stacked PGN layers, the final output of the stacked PGN layers is connected to the input of the fully connected layer.

[0022] The process of calculating and updating wind farm maps using the PGN layer includes the following steps: (1) Each PGN layer in the Path Graph Neural Network (PGNN) updates the edges, nodes, and graph features of the wind farm graph by sending and updating information between nodes in the wind farm graph, so as to estimate the power generation of the wind turbine; during the message passing process, the PGN layer utilizes four differentiable functions: edge update function Edge weight function Node update function and global update function The detailed calculation process of PGN is as follows: Edge update function Update all existing edges in the graph; the edge update function utilizes the characteristics of the sender node. Characteristics of receiver nodes and the edge features between two nodes Generate updated edge features The PGN layer utilizes a weight calculation function. Adjust dynamically based on the relative position of the wind turbine and wind conditions. The weighting; here, the downstream wake distance is selected. Radial wake distance and vertical wake distance As The input is used to calculate the adjustment. weights Using these weights, PGN adjusts the influence strength between turbines to ensure that the calculated weights are physically reasonable; the entire edge update process is shown below: ; Node update function Utilizing the features of the current input node Updated edge features The node features are updated using the global feature g; PGN blocks use an edge aggregation function. Simplify the edge feature set into a single aggregated edge feature. Then, the node update function. Processing the aggregated edge features and current node features Generate updated node features The edge aggregation and node update process is as follows: ; global update function To update global features, PGN uses the updated edge and node features. To do this, PGN utilizes... and All edge features E and all node features Aggregate into two vectors , Global features g The update process is as follows: ; in, , and It is a multilayer perceptron (MLP); in addition to other encoders, it uses physically induced basis functions to construct... Edge weight function The “strength” of the interaction between turbines is adjusted in a manner similar to that used in physical-induced engineering models to describe the relationship between two turbines. Furthermore, the GCH three-dimensional wake model is adopted as the physical mechanism part, and it is used as... Basis functions, applying physics-based weighting functions to the proposed PGN: ; In the formula, As an inducing factor, The rotor radius is... This refers to the surface roughness coefficient. (2) By sequentially applying these update functions to the graph input, i.e. the output graph of the previous PGN layer or the directed graph of the wind farm, the update functions are applied. G The PGN generates an updated graph. After a series of PGN layer calculations, the PGNN generates an updated wind farm graph. Subsequently, the fully connected layer uses this graph as input and outputs a wind speed graph for the wind turbine. The node values ​​represent the estimated wind speed of each wind turbine in the wind farm. The fully connected layer is constructed from a single MLP, which is typically used to estimate the wind speed of each wind turbine in the wind farm. The design of PGNN allows the computation process in the PGN layer to mimic the computation process of engineering-based wake models: 1) The edge weight calculation in the PGN layer directly follows the calculation of the tail flow loss factor. The calculation process is as follows: ; 2) The edge update step in the PGN layer corresponds to the computation of the wind turbine. i and j Between loss factors As shown in the following formula: ; 3) The edge aggregation step in the PGN layer corresponds to the wake defect factor aggregation step, which is used to calculate the wind turbine. i loss factors As shown in the following formula: ; 4) The node update step in the PGN layer corresponds to calculating the average wind speed using the aggregated defect factor. As shown in the following formula: ; Through the above steps, each PGN layer can send and update messages between nodes in the wind farm graph, thereby updating the edges, nodes, and graph features of the wind farm graph. The updated wind farm graph is then... Input into a fully connected layer to estimate the wind turbine wind speed map. This enables a complete process from data input to wind speed prediction.

[0023] Compared with the prior art, the present invention has the following technical effects: 1) This invention starts from the inside of the wind farm, and through a more detailed analysis of the dynamic changes of each wind turbine, first predicts the wind speed at the hub, and then converts it into power output. Compared with GRU and BiLSTM, which directly predict power, the progressively refined prediction method of this invention is more targeted and can better consider the dynamic relationships inside the wind farm; 2) This invention considers the wake effect in wind farms, constructs a dynamic graph structure, and adaptively captures the spatial evolution of the wake path. It effectively solves the one-sidedness of traditional spatiotemporal feature separation extraction methods when dealing with the complex flow characteristics of wind farms, more accurately assesses the mutual influence between wind turbines, and enhances the reliability of prediction. 3) By combining physical mechanisms with data-driven learning methods, this invention embeds traditional power curves into a neural network model for wind speed-power conversion, which can capture the intrinsic relationship between wind speed and power more deeply. Thus, it can exhibit good adaptability and flexibility in the ever-changing wind farm environment, even when faced with new or abnormal wind speed data. 4) The spatiotemporal wind speed prediction and wind speed-power conversion model proposed in this invention aims to deeply analyze the impact of dynamic changes within the wind farm on the power output of each wind turbine, and further predict the total power of the entire wind farm, thus improving the interpretability and accuracy of short-term wind power prediction models. This research provides important basic data support for optimizing power generation strategies and improving grid dispatch efficiency, enhancing system operational performance. Simultaneously, it provides scientific guidance for future optimization of wind power resource utilization, wind farm planning, design, and operation. Attached Figure Description

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 A schematic diagram of the structure of an ultra-short-term wind power prediction method that takes into account wake effect and wind speed-power conversion mechanism; Figure 2 The flowchart shows the ultra-short-term wind speed prediction process for wind turbines based on the PGNN network. Figure 3 This is a schematic diagram of the PGNN network update process; Figure 4 Here is a flowchart of a PGNN network based on the wake effect; Figure 5 A flowchart for ultra-short-term wind power prediction based on the wind speed-power conversion mechanism; Figure 6 This is a map showing the regional divisions based on wind speed conditions. Figure 7 The graph of the Leaky Sigmoid function; Figure 8 This is a diagram of the embedded network structure for wind speed-to-power conversion in a wind turbine. Figure 9 A layout diagram of all wind turbines in a wind farm; Figure 10 for , , Evaluation indicator bar chart; Figure 11 To compare the wind power prediction curves of the model; Figure 12 This is a statistical chart of wind power prediction errors for the Kelmarsh wind farm. Detailed Implementation

[0025] like Figure 1 As shown, the ultra-short-term wind power prediction method based on wake effect and wind speed-power conversion mechanism includes the following steps: Step 1: Collect data from the wind farm m Typhoon machine currently t Looking back H A historical moment r Features (such as historical data on wind farm wind speed and pitch angle) and future performance of each wind turbine k The forecast time includes NWP data (such as wind speed, wind direction, and temperature), as well as information on wind farm layout and turbine characteristics. Step 2: Calculate the physical model using the data obtained in Step 1 to generate high-precision physical simulation results. These results will serve as node information for the graph. Then, normalize the calculation results of the physical model with the input data to construct the graph structure. Obtain the dataset and divide it into three subsets according to a certain ratio for training, validation, and testing, respectively. Step 3: Combine the data and information obtained from Steps 1 and 2 to form node features, edge features, and global features, and construct a directed graph. ; directed graph As input to the PGNN model, wind speed prediction is performed on the wind turbines, and the output is a map showing the predicted wind speed values ​​at each wind turbine hub. The process of generating wind speed predictions using the PGNN model is explained. Step 4: Correct the predicted wind speed values ​​for each fan obtained in Step 3 by considering the influence of ambient temperature; then use the corrected wind speed as input data to perform wind speed-power conversion, and output the corrected predicted fan power value. Repeat this operation until each wind turbine and each prediction time step has been traversed, and output the predicted power value for each wind turbine. ; Step 5: Analyze the predicted power results for each wind turbine obtained in Step 4. Superimposed, outputting the total predicted power of the wind farm .

[0026] In step 3, when obtaining the predicted wind speed values ​​at each wind turbine hub, the specific steps include: Step 3-1: Collect the previous... H Historical data such as wind speed and pitch angle at a specific historical moment in a wind farm. and each fan k The data for the predicted time period are mostly meteorological forecasts. (Such as wind speed, wind direction, temperature, etc.) are preprocessed to generate a directed graph. ; Step 3-2: Directed graph The input is fed into a PGNN network for ultra-short-term wind turbine wind speed prediction, and the output is a map of the predicted wind speed values ​​at each turbine hub. The node values ​​represent the estimated wind speeds of each wind turbine in the wind farm, so... It can also be expressed as .

[0027] Step 3-1 specifically includes the following steps: 3-1-1) Collect parameters such as wind turbine location, height, blade length, rated wind speed, real-time wind speed, power, wake intensity, and NWP wind speed of the wind farm; 3-1-2) Each wind turbine corresponds to a node; these are used as features of the directed graph to construct the graph structure; node features This includes attributes such as the wind turbine's location, height, blade length, and rated wind speed. These attributes are crucial for describing the wind turbine's basic characteristics and operating status; side features Indicates from node j (Sender) to Node i The properties of the edges of the (receiver) include parameters such as real-time wind speed, power, and wake intensity, which mainly reflect the wind turbine. j To the fan i Wake characteristics; edge features are represented as ,in, d ij For two wind turbines i and jThe downstream wake distance between them r ij Radial wake distance, h ij Vertical wake distance; global feature g Let be a scalar or vector, describing the overall properties of the wind farm, where g The NWP wind speed is used directly to characterize the intensity of environmental energy input.

[0028] In step 3-2, the PGNN network consists of stacked Physics-Guided Networks (PGNs) and fully connected layers (which are constructed from a single MLP). Specifically, the structure is: directed graph... As input to the stacked PGN layers, the final output of the stacked PGN layers is connected to the input of the fully connected layer.

[0029] The process of calculating and updating wind farm maps using the PGN layer includes the following steps: (1) The input graph is processed by a Path-based Graph Neural Network (PGNN) based on physical mechanisms, and the wind speed of all wind turbines is estimated as the output graph. PGNN is a message-passing graph neural network, which consists of stacked Physically Induced Graph Networks (PGNs) and fully connected layers. Each PGN layer updates the edges, nodes, and graph features of the wind farm graph by sending and updating information between nodes in the wind farm graph, so as to estimate the power generation of the wind turbines. In the message-passing process, the PGN layer utilizes four differentiable functions: edge update function Edge weight function Node update function and global update function The detailed calculation process of PGN is as follows: Edge update function Update all existing edges in the graph. The edge update function utilizes the characteristics of the sender node. Characteristics of receiver nodes and the edge features between two nodes Generate updated edge features The PGN layer utilizes a weight calculation function. Adjust dynamically based on the relative position of the wind turbine and wind conditions. The weighting; here, the downstream wake distance is selected. Radial wake distance and vertical wake distance As The input is used to calculate the adjustment. weights Using these weights, PGN adjusts the influence strength between turbines to ensure that the calculated weights are physically reasonable; the entire edge update process is shown below: ; Node update function Utilizing the features of the current input node Updated edge features The node features are updated using the global feature g; PGN blocks use an edge aggregation function. Simplify the edge feature set into a single aggregated edge feature. Then, the node update function. Processing the aggregated edge features and current node features Generate updated node features The edge aggregation and node update process is as follows: ; global update function To update global features, PGN uses the updated edge and node features. To do this, PGN utilizes... and All edge features E and all node features Aggregate into two vectors , Global features g The update process is as follows: ; in, , and It is a multilayer perceptron (MLP); in addition to other encoders, it uses physically induced basis functions to construct... Edge weight function The “strength” of the interaction between turbines is adjusted in a manner similar to that used in physical-induced engineering models to describe the relationship between two turbines. Furthermore, the GCH three-dimensional wake model is adopted as the physical mechanism part, and it is used as... Basis functions, applying physics-based weighting functions to the proposed PGN: ; In the formula, As an inducing factor, The rotor radius is... This refers to the surface roughness coefficient. (2) By sequentially applying these update functions to the graph input (i.e., the output graph of the previous PGN layer or the directed graph of the wind farm), GThe PGN (Programme Generation Network) generates an updated graph. After a series of PGN layer calculations, the PGNN generates an updated wind farm graph. Subsequently, the fully connected layer uses this graph as input and outputs a wind speed graph for the wind turbine. The node values ​​represent the estimated wind speed of each wind turbine in the wind farm. The fully connected layer is constructed from a single MLP, which is typically used to estimate the wind speed of each wind turbine in the wind farm. The design of PGNN allows the computation process in the PGN layer to mimic the computation process of engineering-based wake models: 1) The edge weight calculation in the PGN layer directly follows the calculation of the tail flow loss factor. The calculation process is as follows: ; 2) The edge update step in the PGN layer corresponds to the computation of the wind turbine. i and j Between loss factors As shown in the following formula: ; 3) The edge aggregation step in the PGN layer corresponds to the wake defect factor aggregation step, which is used to calculate the wind turbine. i loss factors As shown in the following formula: ; 4) The node update step in the PGN layer corresponds to calculating the average wind speed using the aggregated defect factor. As shown in the following formula: ; Through the above steps, each PGN layer can send and update messages between nodes in the wind farm graph, thereby updating the edges, nodes, and graph features of the wind farm graph. The updated wind farm graph is then... Input into a fully connected layer to estimate the wind turbine wind speed map. This enables a complete process from data input to wind speed prediction.

[0030] In step 4, the corrected wind turbine power prediction value is obtained. Includes the following steps: Step 4-1: Divide the wind turbine into four zones according to its operating status: the start-up zone with no output power, the wind speed-power conversion zone where power increases linearly to the rated power, the constant power zone where the rated power is maintained after exceeding the rated wind speed, and the safe zone where the rated power is maintained after exceeding the rated wind speed. Step 4-2: Considering the influence of ambient temperature, the predicted wind speed values ​​of each fan output in Step 3-2 are... Each predicted wind speed and the corresponding temperature prediction value The predicted wind speed is then corrected by using the input of the embedded network for wind speed-to-power conversion, and a Dense layer is applied to obtain the corrected wind speed. Then according to Cut-in wind speed Rated wind speed Cut-off wind speed Relationship, calculate wind speed Power weights falling into the corresponding interval ,in, For wind speed at The weight of the region; Step 4-3: Use the formula and formula Using the standardized wind speed and its polynomial combination as input, a three-layer fully connected neural network model is employed, based on the formula... Predicted to be located in the wind speed-power conversion zone Internal power output ; in, This is the normalized wind speed value; It is a collection , , The vector is used as the input to the neural network; , , It is the weight matrix of each hidden layer in the neural network; , , This refers to the bias of each hidden layer in a neural network; It is input The weighted summation represents the input. With weight matrix The product, plus the bias In a multi-layer neural network, the output of the previous layer becomes the input of the next layer. This indicates that the output of the first hidden layer, after being processed by the ReLU activation function, is compared with the weights of the second layer. Multiply, and add the bias of the second layer. ; This indicates that the output of the second hidden layer, after being processed by the ReLU activation function, is compared with the weights of the third layer. Multiply, and add the bias of the third layer. ; Activated by the Sigmoid function Limit the output to the range [0,1], multiply by The predicted normalized power is then converted into the actual power output. Step 4-4: Weight the power according to the wind speed region to obtain the predicted value of the wind turbine output power at the predicted time. ; Steps 4-5: Power correction for various factors is achieved by constructing two fully connected layers to output the final power of the wind turbine. .

[0031] In step 4-2, the formula is used. and formula Calculate wind speed Power weights falling into the corresponding interval ;in, For wind speed at The weight of the region; This is a parameter of the LeakySigmoid function, used to control the slope of the function. It is set according to the width of the transition region of the statistical sample, reflecting... The strength of the relationship with the corresponding threshold.

[0032] In step 4-4, the formula is used. The predicted output power of the wind turbine at the predicted time is obtained. ;in, In order to be in The power output value of the region; In order to be in The power output value of the region; In order to be in The power output value of the region; In order to be in The power output value of the region.

[0033] In steps 4-5, power correction for various factors is achieved by constructing two fully connected layers, resulting in the final output power of the wind turbine. The process is as follows: ; ; ; The fully connected layer in the expression is used to handle non-deterministic factors such as blade icing thickness, bearing wear, and blade contamination coefficient. These factors have a certain impact on power, but usually do not cause the wind turbine power output to be completely zero. The fully connected layer in the expression handles decisive factors such as mechanical failures and grid power rationing, which directly affect the start-up and shutdown status of the wind turbine; among which... r and u Let represent the eigenvectors of non-deterministic and deterministic factors, respectively. f and g This is the corresponding activation function.

[0034] In step 5, the wind farm predictive power integration network: superimposes the predicted power outputs of all wind turbines, according to the formula... Total predicted power output of wind farm , This refers to the predicted power output of a single wind turbine.

[0035] The invention also includes a method for processing a wind farm map and obtaining predicted wind speed. It can obtain a directed graph related to the wind turbine and wind farm by acquiring several data related to the wind turbine and wind farm, processing the data, and then processing the directed graph using the PGNN model proposed in this invention to obtain the predicted wind speed value.

[0036] Compared to existing wind speed prediction technologies, this method can obtain wind speed predictions more accurately and efficiently, thus making a technological contribution to the field of wind speed prediction technology.

[0037] Specifically, the method includes the following steps: Step 1: Collect the previous H Historical data such as wind speed and pitch angle at a specific historical moment in a wind farm. and each fan k The data for the predicted time period are mostly meteorological forecasts. A directed graph is generated after preprocessing. ; Step 2: Directed graph The input is fed into a PGNN network for ultra-short-term wind turbine wind speed prediction, and the output is a map of the predicted wind speed values ​​at each turbine hub. The node values ​​represent the estimated wind speeds of each wind turbine in the wind farm, so... It can also be expressed as .

[0038] Step 1 specifically includes the following steps: 1-1) Collect parameters such as wind turbine location, height, blade length, rated wind speed, real-time wind speed, power, wake intensity, and NWP wind speed of the wind farm; 1-2) Each wind turbine corresponds to a node; these are used as features of the directed graph to construct the graph structure; node features This includes attributes such as the wind turbine's location, height, blade length, and rated wind speed. These attributes are crucial for describing the wind turbine's basic characteristics and operating status; side features Indicates from node j Indicates the sender to the node i The attributes of the edges representing the receiver include parameters such as real-time wind speed, power, and wake intensity, which mainly reflect the wind turbine. j To the fan iWake characteristics; edge features are represented as ,in, d ij For two wind turbines i and j The downstream wake distance between them r ij Radial wake distance, h ij Vertical wake distance; global feature g Let be a scalar or vector, describing the overall properties of the wind farm, where g The NWP wind speed is used directly to characterize the intensity of environmental energy input.

[0039] In step 2, the PGNN network consists of stacked Physically Induced Graph Networks (PGNs) and fully connected layers, wherein the fully connected layers consist of a single MLP, specifically a directed graph. As input to the stacked PGN layers, the final output of the stacked PGN layers is connected to the input of the fully connected layer.

[0040] The process of calculating and updating wind farm maps using the PGN layer includes the following steps: (1) Each PGN layer in the Path Graph Neural Network (PGNN) updates the edges, nodes, and graph features of the wind farm graph by sending and updating information between nodes in the wind farm graph, so as to estimate the power generation of the wind turbine; during the message passing process, the PGN layer utilizes four differentiable functions: edge update function Edge weight function Node update function and global update function The detailed calculation process of PGN is as follows: Edge update function Update all existing edges in the graph; the edge update function utilizes the characteristics of the sender node. Characteristics of receiver nodes and the edge features between two nodes Generate updated edge features The PGN layer utilizes a weight calculation function. Adjust dynamically based on the relative position of the wind turbine and wind conditions. The weighting; here, the downstream wake distance is selected. Radial wake distance and vertical wake distance As The input is used to calculate the adjustment. weights Using these weights, PGN adjusts the influence strength between turbines to ensure that the calculated weights are physically reasonable; the entire edge update process is shown below: ; Node update function Utilizing the features of the current input node Updated edge features The node features are updated using the global feature g; PGN blocks use an edge aggregation function. Simplify the edge feature set into a single aggregated edge feature. Then, the node update function. Processing the aggregated edge features and current node features Generate updated node features The edge aggregation and node update process is as follows: ; global update function To update global features, PGN uses the updated edge and node features. To do this, PGN utilizes... and All edge features E and all node features Aggregate into two vectors , Global features g The update process is as follows: ; in, , and It is a multilayer perceptron (MLP); in addition to other encoders, it uses physically induced basis functions to construct... Edge weight function The “strength” of the interaction between turbines is adjusted in a manner similar to that used in physical-induced engineering models to describe the relationship between two turbines. Furthermore, the GCH three-dimensional wake model is adopted as the physical mechanism part, and it is used as... Basis functions, applying physics-based weighting functions to the proposed PGN: ; In the formula, As an inducing factor, The rotor radius is... This refers to the surface roughness coefficient. (2) By sequentially applying these update functions to the graph input, i.e. the output graph of the previous PGN layer or the directed graph of the wind farm, the update functions are applied. G The PGN generates an updated graph. After a series of PGN layer calculations, the PGNN generates an updated wind farm graph. Subsequently, the fully connected layer uses this graph as input and outputs a wind speed graph for the wind turbine. The node values ​​represent the estimated wind speed of each wind turbine in the wind farm. The fully connected layer is constructed from a single MLP, which is typically used to estimate the wind speed of each wind turbine in the wind farm. The design of PGNN allows the computation process in the PGN layer to mimic the computation process of engineering-based wake models: 1) The edge weight calculation in the PGN layer directly follows the calculation of the tail flow loss factor. The calculation process is as follows: ; 2) The edge update step in the PGN layer corresponds to the computation of the wind turbine. i and j Between loss factors As shown in the following formula: ; 3) The edge aggregation step in the PGN layer corresponds to the wake defect factor aggregation step, which is used to calculate the wind turbine. i loss factors As shown in the following formula: ; 4) The node update step in the PGN layer corresponds to calculating the average wind speed using the aggregated defect factor. As shown in the following formula: ; Through the above steps, each PGN layer can send and update messages between nodes in the wind farm graph, thereby updating the edges, nodes, and graph features of the wind farm graph. The updated wind farm graph is then... Input into a fully connected layer to estimate the wind turbine wind speed map. This enables a complete process from data input to wind speed prediction.

[0041] Example: In this embodiment, six 2050kW wind turbines from the Kelmarsh wind farm in the UK were used. Wind power data from 10-minute intervals in 2020 were divided into training, validation, and test sets in an 8:1:1 ratio to perform ultra-short-term wind power predictions from 10 minutes to 4 hours in advance. The layout of each wind turbine in the wind farm is as follows. Figure 9 As shown. The model input includes natural wind speed, wind direction, and temperature in the predicted daily NWP, as well as wind speed, pitch angle, and power of each wind turbine in the wind farm during the historical lookback window (1h~12h). The output is the predicted power of the wind farm for the prediction period. This embodiment uses Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), and Root Mean Square Error (RMSE) as evaluation criteria for the prediction accuracy of each model. Their calculation formulas are as follows:

[0042]

[0043]

[0044] In the formula: The number of test samples; and They are the predicted time points. Actual and predicted wind power values ​​at each sampling point; This corresponds to the rated capacity of the wind farm.

[0045] To verify the application effect of the prediction model of this invention in actual engineering, three types of comparison models were selected: (1) Based on dynamic spatiotemporal feature separation extraction model: DBiLSTM-WSP; (2) Common time series models are used to predict wind turbine wind speed, combined with wind speed-power embedded module (WSP) models: including BiLSTM-S2S-WSP, BiLSTM-WSP and GRU-WSP; (3) Traditional time-series prediction models that directly predict wind farm power: BiLSTM and GRU; (4) Replace the wind speed-power conversion link in the model in this paper with the traditional wind speed-power conversion formula: DBiLSTM-TWSP.

[0046] Compared with the DBiLSTM-WSP model based on dynamic spatiotemporal feature separation extraction, from Figure 10 It can be seen that the prediction errors of MAPE, RMSE, and MAE of the PGNN-WSP model in this chapter are improved by 0.41%, 85.17kW, and 22.38kW, respectively, compared with the DBiLSTM-WSP model. Figure 11 In the data analysis, the DBiLSTM-WSP model (solid green line) performs well in tracking power change trends, generally reflecting the overall direction of power changes. However, it still has shortcomings in accurately tracking actual power. For example, in the time period from 23 to 35, although the model output and the actual power trend are basically consistent, there is always a certain deviation between the two, failing to completely and accurately match the actual power value.

[0047] Comparing the BiLSTM-S2S-WSP, BiLSTM-WSP, and GRU-WSP models that only consider the temporal evolution of wind speed, from... Figure 10 It can be seen that the model of this invention , and The average prediction error improved by 2.26%, 405.37kW and 278.34kW respectively compared to the three time-series comparison models. Figure 11 Among the three models, the BiLSTM-S2S-WSP model (blue solid line) performed best, effectively tracking the actual trend of wind power changes. However, when wind speed and power fluctuate drastically (e.g., between 35 and 50 mph), this model cannot effectively capture these changes. In contrast, the model of this invention (red solid line) considers the impact of the wake effect within the wind farm on the dynamic relationship between wind turbines, thus enabling real-time tracking of power changes.

[0048] Comparing the BiLSTM and GRU models that directly predict wind farm power, from Figure 10 It can be seen that the prediction error of the model in this invention... , , Compared to the two comparative models, the average improvement was 1.75%, 4.82kW, and 119.15kW. Figure 11 In the data, the BiLSTM (orange solid line) and GRU models (yellow solid line) performed reasonably well in tracking power change trends, but their predictions exhibited a lag effect. Especially at the 20-30 time point, the actual power rose rapidly, forming a significant peak, but the prediction curves of both models failed to keep up with this sudden increase, showing a clear lag.

[0049] Compared to the DBiLSTM-TWSP model based on the traditional wind speed-power conversion formula, Figure 10 The prediction error of the model of this invention , , Compared to this model, the power outputs were improved by 6.10%, 962.46 kW, and 750.22 kW, respectively. Figure 11 The data shows a significant deviation between the predicted and actual values ​​of the DBiLSTM-TWSP model (solid purple line) (e.g., at time 20-30). This is mainly because the model relies solely on a preset wind speed-power curve and cannot accurately reflect the dynamic relationship between wind speed and actual power output. In contrast, the model of this invention combines physical mechanisms and data-driven techniques to construct and train a neural network, enabling it to more effectively analyze and simulate the impact of wind speed changes on power output, thereby significantly enhancing the accuracy of predictions.

[0050] Figure 12The chart shows the frequency statistics of the test sample errors for each model's prediction results 1 hour in advance. It can be seen that the errors of this model are mainly concentrated in the per-unit prediction error region [-0.10, 0.10], accounting for over 90% of the total test samples, indicating that the prediction errors of this model are mainly distributed within the small error range. From the perspective of the positive and negative error distribution, the prediction errors of this method are more evenly distributed near the 0 value, resulting in more stable prediction performance. In contrast, the error distribution of the comparison models is more dispersed and fluctuates more significantly, making it difficult to guarantee the stability of the prediction performance.

Claims

1. A wind power prediction method considering wake effect and wind speed-power conversion mechanism, characterized in that, Includes the following steps: Step 1: Collect data from the wind farm m Typhoon machine currently t Looking back H A historical moment r Features and Future of Each Wind Turbine k The NWP data at each predicted time point, as well as information on wind farm layout and wind turbine characteristics; Step 2: Calculate the physical model using the data obtained in Step 1 to generate high-precision physical simulation results. These results will be used as node information in the graph. The calculation results of the physical model are then normalized in combination with the input data to construct a graph structure; the dataset is then obtained and divided into three subsets according to a certain ratio for training, validation and testing, respectively. Step 3: Combine the data and information obtained from Steps 1 and 2 to form node features, edge features, and global features, and construct a directed graph. ; directed graph As input to the PGNN model, wind speed prediction is performed on the wind turbines, and the output is a map showing the predicted wind speed values ​​at each wind turbine hub. ; Step 4: Correct the predicted wind speed values ​​for each fan obtained in Step 3 by considering the influence of ambient temperature; then use the corrected wind speed as input data to perform wind speed-power conversion, and output the corrected predicted fan power value. Repeat this operation until each wind turbine and each prediction time step has been traversed, and output the predicted power value for each wind turbine. ; Step 5: Analyze the predicted power results for each wind turbine obtained in Step 4. The total predicted power of the wind farm is then superimposed and output.

2. The method according to claim 1, characterized in that, In step 3, when obtaining the predicted wind speed values ​​at each wind turbine hub, the specific steps include: Step 3-1: Collect the previous... H Historical data such as wind speed and pitch angle at a specific historical moment in a wind farm. and each fan k The data for the predicted time period are mostly meteorological forecasts. A directed graph is generated after preprocessing. ; Step 3-2: Transform the directed graph The input is fed into a PGNN network for ultra-short-term wind turbine wind speed prediction, and the output is a map of the predicted wind speed values ​​at each turbine hub. The node values ​​represent the estimated wind speeds of each wind turbine in the wind farm, so... It can also be expressed as .

3. The method according to claim 2, characterized in that, Step 3-1 specifically includes the following steps: 3-1-1) Collect parameters such as wind turbine location, height, blade length, rated wind speed, real-time wind speed, power, wake intensity, and NWP wind speed of the wind farm; 3-1-2) Each wind turbine corresponds to a node; these are used as features of the directed graph to construct the graph structure; node features This includes attributes such as the wind turbine's location, height, blade length, and rated wind speed. These attributes are crucial for describing the wind turbine's basic characteristics and operating status; side features Indicates from node j Indicates the sender to the node i The attributes of the edges representing the receiver include parameters such as real-time wind speed, power, and wake intensity, which mainly reflect the wind turbine. j To the fan i Wake characteristics; edge features are represented as ,in, d ij For two wind turbines i and j The downstream wake distance between them r ij Radial wake distance, h ij Vertical wake distance; global feature g Let be a scalar or vector, describing the overall properties of the wind farm, where g The NWP wind speed is used directly to characterize the intensity of environmental energy input.

4. The method according to claim 2, characterized in that, In step 3-2, the PGNN network consists of stacked Physically Induced Graph Networks (PGNs) and fully connected layers, wherein the fully connected layers consist of a single MLP, specifically a directed graph. As input to the stacked PGN layers, the final output of the stacked PGN layers is connected to the input of the fully connected layer.

5. The method according to claim 4, characterized in that, The process of calculating and updating wind farm maps using the PGN layer includes the following steps: (1) Each PGN layer in the Path Graph Neural Network (PGNN) updates the edges, nodes, and graph features of the wind farm graph by sending and updating information between nodes in the wind farm graph, so as to estimate the power generation of the wind turbine; during the message passing process, the PGN layer utilizes four differentiable functions: edge update function Edge weight function Node update function and global update function The detailed calculation process of PGN is as follows: Edge update function Update all existing edges in the graph; the edge update function utilizes the characteristics of the sender node. Characteristics of receiver nodes and the edge features between two nodes Generate updated edge features The PGN layer utilizes a weight calculation function. Adjust dynamically based on the relative position of the wind turbine and wind conditions. The weighting; here, the downstream wake distance is selected. Radial wake distance and vertical wake distance As The input is used to calculate the adjustment. weights Using these weights, PGN adjusts the influence strength between turbines to ensure that the calculated weights are physically reasonable; the entire edge update process is shown below: ； Node update function Utilizing the features of the current input node Updated edge features The node features are updated using the global feature g; PGN blocks use an edge aggregation function. Simplify the edge feature set into a single aggregated edge feature. Then, the node update function. Processing the aggregated edge features and current node features Generate updated node features The edge aggregation and node update process is as follows: ； global update function To update global features, PGN uses the updated edge and node features. To do this, PGN utilizes... and All edge features E and all node features Aggregate into two vectors , Global features g The update process is as follows: ； in, , and It is a multilayer perceptron (MLP); in addition to other encoders, it uses physically induced basis functions to construct... Edge weight function The "strength" of the interaction between turbines is adjusted in a manner similar to that used in physical-induced engineering models to describe the relationship between two turbines; Furthermore, the GCH three-dimensional wake model is adopted as the physical mechanism part, and it is used as... Basis functions, applying physics-based weighting functions to the proposed PGN: ； In the formula, As an inducing factor, The rotor radius is... This refers to the surface roughness coefficient. (2) By sequentially applying these update functions to the graph input, i.e. the output graph of the previous PGN layer or the directed graph of the wind farm, the update functions are applied. G The PGN generates an updated graph. After a series of PGN layer calculations, the PGNN generates an updated wind farm graph. Subsequently, the fully connected layer uses this graph as input and outputs a wind speed graph for the wind turbine. The node values ​​represent the estimated wind speed of each wind turbine in the wind farm. The fully connected layer is constructed from a single MLP, which is typically used to estimate the wind speed of each wind turbine in the wind farm. The design of PGNN allows the computation process in the PGN layer to mimic the computation process of engineering-based wake models: 1) The edge weight calculation in the PGN layer directly follows the calculation of the tail flow loss factor. The calculation process is as follows: ； 2) The edge update step in the PGN layer corresponds to the computation of the wind turbine. i and j Between loss factors As shown in the following formula: ； 3) The edge aggregation step in the PGN layer corresponds to the wake defect factor aggregation step, which is used to calculate the wind turbine. i loss factors As shown in the following formula: ； 4) The node update step in the PGN layer corresponds to calculating the average wind speed using the aggregated defect factor. As shown in the following formula: ； Through the above steps, each PGN layer can send and update messages between nodes in the wind farm graph, thereby updating the edges, nodes, and graph features of the wind farm graph. The updated wind farm graph is then... Input into a fully connected layer to estimate the wind turbine wind speed map. This enables a complete process from data input to wind speed prediction.

6. The method according to any one of claims 1 to 5, characterized in that, In step 4, the corrected wind turbine power prediction value is obtained. Includes the following steps: Step 4-1: Divide the wind turbine into four zones according to its operating status: the start-up zone with no output power, the wind speed-power conversion zone where power increases linearly to the rated power, the constant power zone where the rated power is maintained after exceeding the rated wind speed, and the safe zone where the rated power is maintained after exceeding the rated wind speed. Step 4-2: Considering the influence of ambient temperature, the predicted wind speed values ​​of each fan output in Step 3-2 are... Each predicted wind speed and the corresponding temperature prediction value The predicted wind speed is then corrected by using the input of the embedded network for wind speed-to-power conversion, and a Dense layer is applied to obtain the corrected wind speed. Then according to Cut-in wind speed Rated wind speed Cut-off wind speed Relationship, calculate wind speed Power weights falling into the corresponding interval ,in, For wind speed at The weight of the region; Step 4-3: Use the formula and formula Using the standardized wind speed and its polynomial combination as input, a three-layer fully connected neural network model is employed, based on the formula... Predicted to be located in the wind speed-power conversion zone Internal power output ; in, This is the normalized wind speed value; It is a collection , , The vector is used as the input to the neural network; , , It is the weight matrix of each hidden layer in the neural network; , , This refers to the bias of each hidden layer in a neural network; It is input The weighted summation represents the input. With weight matrix The product, plus the bias In a multi-layer neural network, the output of the previous layer becomes the input of the next layer. This indicates that the output of the first hidden layer, after being processed by the ReLU activation function, is compared with the weights of the second layer. Multiply, and add the bias of the second layer. ; This indicates that the output of the second hidden layer, after being processed by the ReLU activation function, is compared with the weights of the third layer. Multiply, and add the bias of the third layer. ; Activation via Sigmoid function Limit the output to the range [0,1], multiply by The predicted normalized power is then converted into the actual power output. Step 4-4: Weight the power according to the wind speed region to obtain the predicted value of the wind turbine output power at the predicted time. ; Steps 4-5: Power correction for various factors is achieved by constructing two fully connected layers to output the final power of the wind turbine. .

7. The method according to claim 6, characterized in that, In step 4-2, the formula is used. and formula Calculate wind speed Power weights falling into the corresponding interval ;in, For wind speed at The weight of the region; This is a parameter of the LeakySigmoid function, used to control the slope of the function. It is set according to the width of the transition region of the statistical sample, reflecting... The strength of the relationship with the corresponding threshold.

8. The method according to claim 6, characterized in that, In step 4-4, the formula is used. The predicted output power of the wind turbine at the predicted time is obtained. ;in, In order to be in The power output value of the region; In order to be in The power output value of the region; In order to be in The power output value of the region; In order to be in The power output value of the region.

9. The method according to claim 6, characterized in that, In steps 4-5, power correction for various factors is achieved by constructing two fully connected layers, resulting in the final output power of the wind turbine. The process is as follows: ； ； ； The fully connected layer in the expression is used to handle non-deterministic factors such as blade icing thickness, bearing wear, and blade contamination coefficient. These factors have a certain impact on power, but usually do not cause the wind turbine power output to be completely zero. The fully connected layer in the expression handles decisive factors such as mechanical failures and grid power rationing, which directly affect the start-up and shutdown status of the wind turbine; among which... r and u Let represent the eigenvectors of non-deterministic and deterministic factors, respectively. f and g This is the corresponding activation function.

10. The method according to claim 1, 2, 3, 4, 5, 7, 8, or 9, characterized in that, In step 5, the wind farm predictive power integration network: superimposes the predicted power outputs of all wind turbines, according to the formula... Total predicted power output of wind farm , This refers to the predicted power output of a single wind turbine.

Images