Heterogeneous wind power cluster-oriented adaptive graph space-time power prediction method
By using the adaptive graph spatiotemporal network method, the spatiotemporal coupling relationship between wind farms is dynamically modeled, which solves the problem of static modeling of adaptability and spatiotemporal correlation in heterogeneous wind farms in traditional models, and realizes high-precision wind power cluster power prediction.
Patent Information
- Application Number
- CN202511775768.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-28
- Publication Date
- 2026-02-24
AI Technical Summary
Traditional wind power cluster power prediction models struggle to capture the dynamic spatiotemporal relationships between heterogeneous wind farms, cannot adapt to structural differences in equipment configuration and terrain conditions, and have static spatiotemporal relationship modeling and simple feature fusion logic, resulting in insufficient prediction accuracy.
An adaptive graph spatiotemporal network approach is adopted, which dynamically models the spatiotemporal coupling relationship between wind farms through heterogeneous feature encoding, adaptive graph structure generation, and prediction modeling based on graph spatiotemporal networks. It adapts to different equipment types and terrain conditions of wind farms, and uses the Gumbel-Softmax mechanism to construct sparse graphs for graph convolution and GRU temporal modeling.
It improves the cross-regional generalization ability and accuracy of wind power cluster power prediction, and can effectively adapt to the dynamic spatiotemporal relationship of heterogeneous wind farms, thereby improving prediction accuracy and adaptability.
Smart Images
Figure QLYQS_1 
Figure QLYQS_4 
Figure QLYQS_8
Abstract
Description
Technical Field
[0001] This invention relates to an adaptive graph spatiotemporal power prediction method for heterogeneous wind power clusters, applicable to regional wind power prediction and scheduling optimization scenarios, and belongs to the field of new energy power generation technology. Background Technology
[0002] With the large-scale grid connection of wind power, regional wind power forecasting has become a key technology for ensuring the stable operation of the power system. Currently, the construction model of wind farms is gradually evolving from early decentralized, small-scale, and low-penetration approaches to centralized, large-scale, and high-penetration clustered layouts. The output characteristics of these large-scale wind power clusters are not only significantly affected by macro-weather systems but also face challenges from the inherent biases in the spatiotemporal accuracy and prediction amplitude of numerical weather prediction (NWP). Unlike isolated wind farms, wind farms within a cluster exhibit more complex spatiotemporal correlations. Traditional prediction models based on single wind farm information or simple aggregation are insufficient to effectively capture the dynamic interactions within the cluster and the complex mechanisms by which global meteorological conditions affect overall power output, thus failing to meet increasingly sophisticated operational needs.
[0003] Chinese patent application CN2021114217235, entitled "A Method and Device for Predicting Wind Power Clusters Based on Heterogeneous Data and Deep Learning," constructs a wind power cluster power prediction network model. It extracts features from heterogeneous data using a feature extraction network, predicts key information based on the extracted features using an attention mechanism, and then uses a multimodal fusion strategy to generate multimodal fused features. Wind power cluster power is then predicted based on these generated multimodal fused features. However, this technical solution has the following drawbacks: 1. Poor adaptability to heterogeneous features: It extracts features from all heterogeneous data using only a unified CNN+BiLSTM network, without differentiating the encoding of static attributes (machine type, terrain) and dynamic temporal features (wind speed, weather) of wind farms, making it difficult to adapt to the structural differences of different wind farms. 2. Static spatiotemporal correlation modeling: It does not consider the dynamic evolution characteristics of spatiotemporal correlations between wind farms, does not introduce a graph structure learning mechanism, and cannot dynamically construct adjacency relationships based on geographical location and physical attributes, thus failing to characterize the dynamic correlation changes brought about by wind propagation paths. 3. Simple feature fusion logic: It only uses a simple merging and splicing method for feature fusion, without fully exploring the spatiotemporal complementarity and correlation between heterogeneous features. The information density of the fused features is low, and the efficiency of extracting key information is limited.
[0004] Chinese patent application CN2025109186907, entitled "Method, Device, Equipment and Storage Medium for Short-Term Power Prediction of Wind Farm Clusters," aims to accurately capture the complex spatiotemporal relationships of wind farm clusters and improve prediction accuracy through clustering and joint modeling of homogeneous / heterogeneous dual graphs. However, this technical solution has the following drawbacks: 1. Lack of dynamic adaptability in the graph structure: Existing patents use clustering to divide fixed sub-clusters, constructing static homogeneous and heterogeneous graphs. The adjacency matrix is only updated daily, failing to capture the dynamic spatiotemporal relationships between wind farms as weather changes, thus limiting adaptability. 2. Insufficient characterization of heterogeneous features: Existing patents only use wind speed data as the core clustering basis, failing to integrate static physical attributes such as wind turbine model, terrain, and altitude, making it difficult to adapt to the heterogeneous output characteristics of wind farms due to different equipment configurations and underlying surface conditions. 3. Weak physical correlation and generalization ability: Existing patents rely on predefined clustering rules and fixed adjacency relationships, failing to deeply integrate physical laws such as wind propagation paths and distance attenuation, leading to a decrease in prediction accuracy when migrating across regions and scenarios.
[0005] In summary, the existing technologies have the following main shortcomings: (1) Static graph structure limitations: Traditional graph neural networks use fixed adjacency matrices, which cannot characterize the spatiotemporal relationships of dynamic evolution between wind farms. (2) Poor adaptability to heterogeneous features: CNN, LSTM and other models ignore the structural differences of wind farms in terms of equipment configuration and terrain conditions, making it difficult to capture non-stationary output characteristics. Summary of the Invention
[0006] The purpose of this invention is to provide an adaptive graph spatiotemporal power prediction method for heterogeneous wind power clusters, which can adapt to the dynamic spatiotemporal relationship modeling of heterogeneous wind power clusters, improve cross-regional generalization ability and prediction accuracy, and solve the above-mentioned problems existing in the background technology.
[0007] The technical solution of this invention is:
[0008] An adaptive graph spatiotemporal power prediction method for heterogeneous wind power clusters includes the following steps:
[0009] ① Heterogeneous feature encoding; ② Adaptive graph structure generation; ③ Predictive modeling based on graph spatiotemporal networks; This enables dynamic modeling of the spatiotemporal coupling relationship between wind farms in a wind power cluster, and adapts to the structural differences of different wind farms in terms of equipment type, terrain conditions and meteorological characteristics.
[0010] Step ① Heterogeneous feature encoding:
[0011] Suppose the region contains N wind farms, and each wind farm at time t is represented by node v in the adaptive graph. i ∈ ={v1, v2, …, v N}; ={v1,v2,…,vN} represents the set of wind farm nodes, containing the nodes corresponding to all N wind farms in the region; the corresponding input feature is represented as ∈R d Where d is the feature dimension, including static wind field attributes and meteorological time series information; construct a graph sequence with time evolution. (t) =( , (t) A (t) ),in (t) Let A be an edge set. (t) ∈ It is an adjacency matrix. ∈[0, 1] represents node v i With v j The correlation strength at time t; the model objective is to predict the output power sequence of the wind farm over the next H time steps. , …, };
[0012] N represents the total number of wind farms in the region, i.e., the total number of nodes in the graphical model; v i Represents the graph node corresponding to the i-th wind farm; R d d is a real number space used to define the dimension of the wind farm input features, where d is the total number of features;
[0013] Wind farm input characteristics include static characteristics s i ∈ and dynamic time series characteristics L is the length of the time window; the node input is concatenated as follows:
[0014] (1)
[0015] in, Let i represent the i-th wind farm in the region, t represent time t, and s represent the input characteristics of all wind farms in the region. i Let d represent the static properties of the i-th wind farm. i Let L represent the dynamic time-series features of the i-th wind farm, L represent the time window length, and Concat(•) represent feature-level concatenation; the static features and the dynamic features at each time step are respectively mapped through two layers of nonlinear embedding:
[0016] (2)
[0017] (3)
[0018] Among them, W s W dLet W represent the weight matrix, and W s ∈ W d ∈ b s b d b represents the bias term. s b d ∈ σ(•) is the ReLU activation function; the node embedding representation is obtained after concatenation. ∈ .
[0019] Step ② Adaptive graph structure generation:
[0020] Based on the geographical location of the wind farm, the initial adjacency matrix A is constructed. (0) :
[0021] (4)
[0022] Where d ij The distance between stations is represented by σ, where i represents the x-coordinate, j represents the y-coordinate, r0 represents the neighborhood radius, and σ represents the distance between stations. d The distance decay scale is represented by 1(•); the indicator function is represented by 1(•); subsequently, a candidate edge mask M is used in the attention edge weight calculation. ij =1( >0), that is, only in E0={(i, j): M ij When learning on =1}, the adjacency A is variable. (t) and with A (0) As A (t) The initialization (warm-start) introduces physical constraints of spatial proximity from the early stages of training; M ij Denotes the candidate edge mask; E0 represents the initial candidate edge set; A(t) represents the dynamic adjacency matrix at time t; A (0) Represents the initial adjacency matrix;
[0023] Based on node embedding Constructing the adjacency matrix A through an attention mechanism (t) ; any node (v i , v j The edge weights are calculated as follows:
[0024] (5)
[0025] In the formula, Represents the i-th wind farm node v i The node embedding vector at time t; a ij (t) Represents any node (v) at time t i , vj The edge weights of ) are given by φ(•), which represents the scoring function.
[0026] The scoring function is as follows:
[0027] (6)
[0028] Among them W q W k Let b represent the trainable weights, b represent the bias, and W represent the weights. q W k ∈ b∈ ; a represents the attention vector, a∈ .
[0029] Step ③ is based on prediction modeling using graph spatiotemporal networks:
[0030] Introducing the Gumbel-Softmax mechanism to discretize edge weights and construct a sparse graph:
[0031] (7)
[0032] Among them, ij (t) Represents any node (v) at time t i , v j The edge weights of ) are given by τ, which represents the temperature coefficient of the Gumbel-Softmax mechanism. Let vj represent the original edge weights between wind farm node vi (the i-th wind farm) and node vj (the j-th wind farm) at time t.
[0033] Finally, a dynamic adjacency matrix is generated. ∈ ;
[0034] Represents a dynamic adjacency matrix;
[0035] At each time step t, based on the normalized adjacency matrix = Perform graph convolution operations:
[0036] (8)
[0037] in, D represents the normalized adjacency matrix; D represents the node degree matrix, D∈ ; This represents the embedding representation of all nodes. ∈ ; Represented as a graph convolution weight matrix, ∈ ; Indicates spatial feature output, ∈ ;
[0038] Convolution result of time series graph Input time modeling module gated loop unit (GRU):
[0039] (9)
[0040] in, Represents node v i Graph characteristics at time tk ∈ ; Output This represents the fused spacetime representation. ∈ ;
[0041] The final prediction of the wind power sequence for the next H time steps:
[0042] (10)
[0043] in, i (t+1:t+H) This represents the wind power sequence over the next H time steps. , Indicates the output mapping parameters. ∈ , ∈ ;
[0044] The loss function used in the prediction process consists of the prediction error and a graph structure regularization term:
[0045] (11)
[0046] Where L is the loss function; N represents the total number of wind farms, defining the spatial range for loss calculation; and H represents the prediction time step, defining the time range for loss calculation. Indicates the predicted value. The actual value; denoted by Frobenius norm; λ is the regularization coefficient.
[0047] In step ①, the static features include wind turbine model, terrain type, altitude, historical wind speed level, etc.; the dynamic features include wind speed, wind direction, temperature, etc.
[0048] In step ②, the scoring function, static features (aircraft type, terrain, altitude, etc.) and dynamic features (wind speed, wind direction, temperature, etc.) are jointly embedded into the scoring function of formula (6) through node embedding, so that the attention weight remains sensitive to the wind propagation path and the difference of the underlying surface, thereby explicitly injecting physical environment information into the graph structure learning process.
[0049] The beneficial effects of this invention are as follows: Addressing the problems of complex models and long training times in current traditional photovoltaic power prediction methods, this application proposes a lazy learning method for photovoltaic power prediction based on fast retrieval. This method can directly calculate predicted values using statistical methods based on the structural characteristics of historical datasets and flexibly adjust the contribution of each feature quantity to the result based on local features of the historical datasets. This solves the problems of complex models and long training times in traditional photovoltaic power prediction methods. The method of this application effectively improves the speed of photovoltaic power prediction and the efficiency of establishing photovoltaic power prediction models, which is of great significance for improving the power quality of photovoltaic power generation. Detailed Implementation
[0050] The present invention will be further illustrated by the following examples.
[0051] An adaptive graph spatiotemporal power prediction method for heterogeneous wind power clusters includes the following steps:
[0052] ① Heterogeneous feature encoding; ② Adaptive graph structure generation; ③ Predictive modeling based on graph spatiotemporal networks; This enables dynamic modeling of the spatiotemporal coupling relationship between wind farms in a wind power cluster, and adapts to the structural differences of different wind farms in terms of equipment type, terrain conditions and meteorological characteristics.
[0053] Step ① Heterogeneous feature encoding:
[0054] Suppose the region contains N wind farms, and each wind farm at time t is represented by node v in the adaptive graph. i ∈ ={v1, v2, …, v N}; ={v1,v2,…,v N} represents the set of wind farm nodes, containing the nodes corresponding to all N wind farms within the region. The corresponding input feature is represented as: ∈R d Where d is the feature dimension, including static wind field attributes and meteorological time series information; construct a graph sequence with time evolution. (t) =( , (t) A (t) ),in (t) Let A be an edge set. (t) ∈ It is an adjacency matrix. ∈[0, 1] represents node v i With v j The correlation strength at time t; the model objective is to predict the output power sequence of the wind farm over the next H time steps. , …, };
[0055] N represents the total number of wind farms in the region, i.e., the total number of nodes in the graphical model; v i Represents the graph node corresponding to the i-th wind farm; R d d is a real number space used to define the dimension of the wind farm input features, where d is the total number of features;
[0056] Wind farm input characteristics include static characteristics s i ∈ and dynamic time series characteristics L is the length of the time window; the node input is concatenated as follows:
[0057] (1)
[0058] in, Let i represent the i-th wind farm in the region, t represent time t, and s represent the input characteristics of all wind farms in the region. i Let d represent the static properties of the i-th wind farm. i Let L represent the dynamic time-series features of the i-th wind farm, L represent the time window length, and Concat(•) represent feature-level concatenation; the static features and the dynamic features at each time step are respectively mapped through two layers of nonlinear embedding:
[0059] (2)
[0060] (3)
[0061] Among them, W s W d Let W represent the weight matrix, and W s ∈ W d ∈ b s , b d b represents the bias term. s , b d ∈ σ(•) is the ReLU activation function. The concatenated representation yields the node embedding. ∈ .
[0062] Step ② Adaptive graph structure generation:
[0063] Based on the geographical location of the wind farm, the initial adjacency matrix A is constructed. (0) :
[0064] (4)
[0065] Where d ij The distance between stations is represented by σ, where i represents the x-coordinate, j represents the y-coordinate, r0 represents the neighborhood radius, and σ represents the distance between stations. d The distance decay scale is represented by 1(•); the indicator function is represented by 1(•); subsequently, a candidate edge mask M is used in the attention edge weight calculation. ij =1( >0), that is, only in E0={(i, j): M ij When learning on =1}, the adjacency A is variable. (t) and with A (0) As A (t) The initialization (warm-start) introduces physical constraints of spatial proximity from the early stages of training; M ij Denotes the candidate edge mask; E0 represents the initial candidate edge set; A(t) represents the dynamic adjacency matrix at time t; A (0) Represents the initial adjacency matrix;
[0066] Based on node embedding Constructing the adjacency matrix A through an attention mechanism (t) Any node (v) i , v j The edge weights are calculated as follows:
[0067] (5)
[0068] In the formula, Represents the i-th wind farm node v i The node embedding vector at time t; a ij (t) Represents any node (v) at time t i , v j The edge weights of ) are given by φ(•), which represents the scoring function.
[0069] The scoring function is as follows:
[0070] (6)
[0071] Among them W q W k Let b represent the trainable weights, b represent the bias, and W represent the weights. q W k ∈ b∈ ; a represents the attention vector, a∈ .
[0072] Step ③ is based on prediction modeling using graph spatiotemporal networks:
[0073] Introducing the Gumbel-Softmax mechanism to discretize edge weights and construct a sparse graph:
[0074] (7)
[0075] Among them, ij (t) Represents any node (v) at time t i , v j The edge weights of ) are given by τ, which represents the temperature coefficient of the Gumbel-Softmax mechanism. Let vj represent the original edge weights between wind farm node vi (the i-th wind farm) and node vj (the j-th wind farm) at time t.
[0076] Finally, a dynamic adjacency matrix is generated. ∈ ;
[0077] Represents a dynamic adjacency matrix;
[0078] At each time step t, based on the normalized adjacency matrix = Perform graph convolution operations:
[0079] (8)
[0080] in, D represents the normalized adjacency matrix; D represents the node degree matrix, D∈ ; This represents the embedding representation of all nodes. ∈ ; Represented as a graph convolution weight matrix, ∈ ; Indicates spatial feature output, ∈ ;
[0081] Convolution result of time series graph Input time modeling module gated loop unit (GRU):
[0082] (9)
[0083] in, Represents node vi Graph characteristics at time tk ∈ ; Output This represents the fused spacetime representation. ∈ ;
[0084] The final prediction of the wind power sequence for the next H time steps:
[0085] (10)
[0086] in, i (t+1:t+H) This represents the wind power sequence over the next H time steps. , Indicates the output mapping parameters. ∈ , ∈ ;
[0087] The loss function used in the prediction process consists of the prediction error and a graph structure regularization term:
[0088] (11)
[0089] Where L is the loss function; N represents the total number of wind farms, defining the spatial range for loss calculation; and H represents the prediction time step, defining the time range for loss calculation. Indicates the predicted value. The actual value; denoted by Frobenius norm; λ is the regularization coefficient.
[0090] In step ①, the static features include wind turbine model, terrain type, altitude, historical wind speed level, etc.; the dynamic features include wind speed, wind direction, temperature, etc.
[0091] In step ②, the scoring function, static features (aircraft type, terrain, altitude, etc.) and dynamic features (wind speed, wind direction, temperature, etc.) are jointly embedded into the scoring function of formula (6) through node embedding, so that the attention weight remains sensitive to the wind propagation path and the difference of the underlying surface, thereby explicitly injecting physical environment information into the graph structure learning process. Specific Implementation
[0092] A wind power cluster in Northwest my country was selected as the application scenario. This cluster comprises N=8 wind farms (numbered V1-V8), distributed across an area spanning approximately 80km east-west and 60km north-south. The specific parameters of each wind farm are shown in the table below:
[0093]
[0094] Time parameters: The data collection period is from July 1st to July 7th (7 days in total), the sampling frequency is 15 minutes / time, 96 time points are collected per day, and the total collection time step T=7×96=672.
[0095] Dynamic time-series characteristics: The collected dynamic characteristics include wind speed (m / s), wind direction (°), and air temperature (°C), with a time window length of L=16 (i.e., predicting future power based on data from the previous 4 hours).
[0096] Prediction objective: Predict wind power output over the next H = 8 time steps (i.e., the next 2 hours), in MW.
[0097] Numerical weather forecast (NWP) data: High-precision regional NWP data is used, with a spatiotemporal resolution of 1km×1km and 15 minutes / time, to provide a reference for meteorological trends within the forecast period.
[0098] Set model hyperparameters
[0099]
[0100] Calculation steps based on the method of this invention
[0101] Step ①: Heterogeneous Feature Encoding
[0102] Feature integration: The input feature dimension d for each wind farm = static feature dimension + dynamic feature dimension = 4 (wind turbine model code + terrain type code + altitude + historical wind speed level) + 3 (wind speed + wind direction + temperature) = 7. The input features are represented as follows: (i=1-8, t=1-672).
[0103] Static feature coding: The wind turbine model and terrain type are numerically coded (golden wind = 1, bright sun = 2, distant view = 3; plain = 1, gentle slope = 2, mountain = 3), and combined with altitude and historical wind speed level to form static features. ∈ Nonlinear embedding is performed using equation (2):
[0104] ,in , σ is the ReLU activation function, and the output is... .
[0105] Dynamic feature encoding: Take the dynamic features of the first L=16 time steps of each time step t. (each) ∈ ), and perform nonlinear embedding through equation (3):
[0106] ,in Output .
[0107] Node embedding concatenation: The static embedding is concatenated with the dynamic embedding from 16 time steps using the Concat function to obtain the node embedding. .
[0108] Step 2: Adaptive graph structure generation
[0109] Initial adjacency matrix construction: Calculate the initial adjacency matrix A(0) according to equation (4). Taking V1 and V2 as an example, the Euclidean distance is... Therefore:
[0110] Similarly, calculate the edge weights for all pairs of nodes to obtain... .
[0111] Dynamic Adjacency Matrix Learning: Based on Node Embedding The attention edge weights are calculated using equations (5) and (6). The scoring function in equation (6) is... , where W_q, , Output .
[0112] Sparse graph construction: The edge weights are discretized and approximated using the Gumbel-Softmax mechanism in equation (7), retaining the top three edges by weight (i.e., each node retains only its three most relevant neighboring nodes), thus generating a dynamic adjacency matrix. .
[0113] Step 3: Predictive Modeling Based on Graph Spatiotemporal Networks
[0114] Graph convolution operation: For each time step t, compute the normalized adjacency matrix. (D is the node degree matrix), graph convolution is performed using equation (8):
[0115] ,in Output .
[0116] Temporal feature modeling: Input the graph convolutional structure {Z(t-3), Z(t-2), Z(t-1), Z(t)} for K=4 consecutive time steps into the GRU module, fuse the spatiotemporal features through equation (9), and output the spatiotemporal representation. .
[0117] Power prediction: The power prediction values for the next H = 8 time steps are obtained by mapping using equation (10). .
[0118] Loss optimization: The loss function L is calculated using equation (11), and the Adam optimizer is used for iterative training 500 times to minimize the loss function.
[0119] Comparison of predicted results and actual power
[0120] The accuracy of predictions is evaluated using mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE).
[0121] After training, wind power was predicted for the period from 00:00 to 02:00 on July 8, 2024 (H=8 time steps). The overall performance indicators are shown in the table below:
[0122]
[0123] Comparison of prediction results for a single wind farm
[0124] Taking V1 (2.5MW gold wind, plains terrain) and V3 (2.0MW distant view, mountainous terrain) as examples, the specific predicted values are compared with the actual values in the table below (time step is 15 minutes):
[0125] Wind farm V1 power comparison (unit: MW)
[0126]
[0127] Wind farm V3 power comparison (unit: MW)
[0128]
[0129] Overall accuracy: The overall MAPE of the eight wind farms in this wind power cluster is only 2.18%, MAE is 0.32MW, and RMSE is 0.45MW, indicating that the predicted values are highly consistent with the actual power.
[0130] Heterogeneous adaptability: The average relative error of V1 wind farm in plain terrain is 1.05%, and the average relative error of V3 wind farm in mountainous terrain is 1.50%, both controlled within 3%, proving that the method of the present invention can effectively adapt to heterogeneous wind farms with different equipment types and terrain conditions.
[0131] This example verifies the effectiveness of the adaptive graph spatiotemporal power prediction method of this invention using actual data from a heterogeneous wind power cluster in Northwest China. In the power prediction for the next two hours, the overall average relative error is only 2.18%, and the maximum average relative error for a single wind farm does not exceed 1.5%, with the predicted value differing very little from the actual power. This result fully demonstrates the effectiveness of the method in dynamic spatiotemporal relationship modeling and heterogeneous feature adaptation, significantly improving the power prediction accuracy of wind power clusters and meeting the practical needs of regional power system dispatch optimization.
Claims
1. An adaptive graph spatiotemporal power prediction method for heterogeneous wind power clusters, characterized in that... It includes the following steps: ① Heterogeneous feature encoding; ② Adaptive graph structure generation; ③ Predictive modeling based on graph spatiotemporal networks; It enables dynamic modeling of the spatiotemporal coupling relationship between wind farms in a wind power cluster, and adapts to the structural differences of different wind farms in terms of equipment type, terrain conditions and meteorological characteristics.
2. The adaptive graph spatiotemporal power prediction method for heterogeneous wind power clusters according to claim 1, characterized in that... Step ① Heterogeneous feature encoding: set up The region contains N wind farms, and each wind farm at time t is represented by node v in the adaptive graph. i ∈ ={v1,v2, …, v N }; ={v1,v2,…,v N } represents the set of wind farm nodes, containing the nodes corresponding to all N wind farms in the region; the corresponding input feature is represented as ∈R d Where d is the feature dimension, including static wind field attributes and meteorological time series information; construct a graph sequence with time evolution. (t) =( , (t) A (t) ),in (t) Let A be an edge set. (t) ∈ It is an adjacency matrix. ∈[0, 1] represents node v i With v j The correlation strength at time t; the model objective is to predict the output power sequence of the wind farm over the next H time steps. , …, }; N represents the total number of wind farms in the region, i.e., the total number of nodes in the graphical model; v i Represents the graph node corresponding to the i-th wind farm; R d d is a real number space used to define the dimension of the wind farm input features, where d is the total number of features; Wind farm input characteristics include static characteristics s i ∈ and dynamic time series characteristics L is the length of the time window; the node input is concatenated as follows: (1) in, Let i represent the i-th wind farm in the region, t represent time t, and s represent the input characteristics of all wind farms in the region. i Let d represent the static properties of the i-th wind farm. i Let L represent the dynamic time-series features of the i-th wind farm, L represent the time window length, and Concat(•) represent feature-level concatenation; the static features and the dynamic features at each time step are respectively mapped through two layers of nonlinear embedding: (2) (3) Among them, W s W d Let W represent the weight matrix, and W s ∈ W d ∈ b s b d b represents the bias term. s b d ∈ σ(•) is the ReLU activation function; the node embedding representation is obtained after concatenation. ∈ .
3. The adaptive graph spatiotemporal power prediction method for heterogeneous wind power clusters according to claim 2, characterized in that... Step ② Adaptive graph structure generation: Based on the geographical location of the wind farm, the initial adjacency matrix A is constructed. (0) : (4) Where d ij The distance between stations is represented by σ, where i represents the x-coordinate, j represents the y-coordinate, r0 represents the neighborhood radius, and σ represents the distance between stations. d Indicates the distance attenuation scale; 1(•) denotes the indicator function; subsequently, a candidate edge mask M is used in the attention edge weight calculation. ij =1( >0), that is, only in E0={(i, j): M ij When learning on =1}, the adjacency of A changes. (t) and with A (0) As A (t) The initialization (warm-start) introduces physical constraints of spatial proximity from the early stages of training; M ij Denotes the candidate edge mask; E0 represents the initial candidate edge set; A(t) represents the dynamic adjacency matrix at time t; A (0) Represents the initial adjacency matrix; Based on node embedding Constructing the adjacency matrix A through an attention mechanism (t) ; any node (v i , v j The edge weights are calculated as follows: (5) In the formula, Represents the i-th wind farm node v i The node embedding vector at time t; a ij (t) Represents any node (v) at time t i , v j The edge weights of ) are given by φ(•), which represents the scoring function. The scoring function is as follows: (6) Among them W q W k Let b represent the trainable weights, b represent the bias, and W represent the weights. q W k ∈ b∈ ; a represents the attention vector, a∈ .
4. The adaptive graph spatiotemporal power prediction method for heterogeneous wind power clusters according to claim 3, characterized in that... Step ③ is based on prediction modeling using graph spatiotemporal networks: Introducing the Gumbel-Softmax mechanism to discretize edge weights and construct a sparse graph: (7) Among them, ij (t) Represents any node (v) at time t i , v j The edge weights of ) are given by τ, which represents the temperature coefficient of the Gumbel-Softmax mechanism. This represents the original edge weights between wind farm node vi and node vj at time t; Finally, a dynamic adjacency matrix is generated. ∈ ; Represents a dynamic adjacency matrix; At each time step t, based on the normalized adjacency matrix = Perform graph convolution operations: (8) in, D represents the normalized adjacency matrix; D represents the node degree matrix, D∈ ; This represents the embedding representation of all nodes. ∈ ; Represented as a graph convolution weight matrix, ∈ ; Indicates spatial feature output, ∈ ; Convolution result of time series graph Input time modeling module gated recurrent unit (GRU): (9) in, Represents node v i Graph characteristics at time tk ∈ ; Output This represents the fused spacetime representation. ∈ ; The final prediction of the wind power sequence for the next H time steps: (10) in, i (t+1:t+H) This represents the wind power sequence over the next H time steps. , Indicates the output mapping parameters. ∈ , ∈ ; The loss function used in the prediction process consists of the prediction error and a graph structure regularization term: (11) Where L is the loss function; N represents the total number of wind farms, defining the spatial range for loss calculation; and H represents the prediction time step, defining the time range for loss calculation. Indicates the predicted value. The actual value; denoted by Frobenius norm; λ is the regularization coefficient.
5. The adaptive graph spatiotemporal power prediction method for heterogeneous wind power clusters according to claim 2, characterized in that: In step ①, the static features include wind turbine model, terrain type, altitude, and historical wind speed level; the dynamic features include wind speed, wind direction, and temperature.
6. The adaptive graph spatiotemporal power prediction method for heterogeneous wind power clusters according to claim 3, characterized in that: In step ②, the scoring function, static features and dynamic features are jointly embedded into the scoring function of equation (6) through node embedding, so that the attention weight remains sensitive to the wind propagation path and the difference of the underlying surface, thereby explicitly injecting physical environment information into the graph structure learning process.