Wind power plant short-term wind power prediction method based on space-time attention mechanism

Abstract

The invention belongs to the technical field of wind power plant power prediction and intelligent operation and maintenance, and particularly relates to a wind power plant short-term wind power prediction method based on a space-time attention mechanism. According to the method, a dynamic graph evolved along with time is constructed through multi-source heterogeneous data fusion, and a time-varying adjacency matrix is output; obtaining space-time coupling characteristics through disturbance-sensitive space attention coding and wind speed-power coupling sensitivity driven time attention decoding; multi-scale features are output through dual-scale adaptive fusion; and finally, forming a continuously evolved prediction chain through edge end rolling learning and topological closed-loop correction by combining a loss back propagation optimization model of physical consistency constraint. According to the method, the problems of lack of physical constraints, insufficient dynamic adaptation and low efficiency of edge deployment in the prior art are solved, the prediction physical reasonability and dynamic adaptation are improved, efficient edge deployment is realized, and reliable support is provided for wind power plant scheduling.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wind farm power prediction and intelligent operation and maintenance, and particularly relates to a wind farm short-term wind power prediction method based on a space-time attention mechanism. BACKGROUND

[0002] With the increasing proportion of wind power in the energy structure, wind power prediction technology has become a key to ensuring the stability of the power grid and improving the wind power consumption capacity. At present, machine learning (such as recurrent neural networks and attention mechanisms) is widely used in wind power prediction. By mining the space-time correlation of historical meteorological and power data, the prediction accuracy is improved to some extent. For example, the long short-term memory network (LSTM) is used to capture the time series features, and the spatial interpolation algorithm is used to integrate the wind turbine distribution information to build a wind farm power prediction model, which has been preliminarily implemented in practical projects.

[0003] Defects of the prior art:

[0004] The lack of physical constraints leads to insufficient prediction rationality. Existing models rely on data-driven fitting and do not impose constraints on the physical segmentation characteristics (power logic in the cut-in wind speed, rated wind speed and cut-out wind speed interval) of the wind turbine power curve. When facing extreme working conditions (such as sudden changes in wind speed and equipment failure), it is easy to produce results that violate the physical law, such as "still predicting power output above the cut-out wind speed", which cannot meet the strict requirements of grid dispatching for the physical rationality of the prediction results.

[0005] Dynamic scene adaptation lags behind. The meteorological environment (real-time changes in wind direction and wind speed) and the device state (individual failure and performance degradation) of the wind farm are in dynamic evolution. Existing models cannot update the topology relationship (spatial correlation) between wind turbines and model weight parameters in real time. Due to the inability to adapt to dynamic changes, the prediction accuracy continues to decline over a long period of time, and the complex and variable operation scenarios of the wind farm cannot be effectively dealt with.

[0006] The efficiency of edge deployment is low. The storage and transmission of high-precision models after training have high overhead, and the edge computing unit of the wind turbine is limited by computing power and storage resources, which cannot support real-time prediction and dynamic updating of the model, resulting in a disconnection between the "cloud training-edge deployment" process and the formation of an efficient closed-loop prediction link.

[0007] Therefore, we propose a wind farm short-term wind power prediction method based on a space-time attention mechanism. SUMMARY

[0008] In view of the above defects in the prior art, the purpose of the present application is to provide a wind farm short-term wind power prediction method based on a space-time attention mechanism, which solves the problems of lack of physical constraints, insufficient dynamic adaptation and low efficiency of edge deployment, realizes accurate prediction, adapts to dynamic changes of the wind farm, and provides support for dispatching and the like.

[0009] To achieve the above object, the technical scheme adopted by the present application is as follows:

[0010] A wind farm short-term wind power prediction method based on a space-time attention mechanism, comprising the following steps:

[0011] S1. Multi-source heterogeneous data fusion and dynamic graph construction, fusing the power generation of distributed photovoltaics, the power generation of wind power, the SOC of energy storage, and the electricity consumption data of flexible load, constructing a dynamic graph evolving over time, and outputting a time-varying adjacency matrix;

[0012] S2. Disturbance-sensitive spatial attention encoding, marking the disturbance region based on the time-varying adjacency matrix on the node feature difference, and aggregating the neighborhood disturbance information of the dynamic edge weight, and outputting the disturbance-enhanced spatial feature;

[0013] S3. Wind speed-power coupling sensitivity driven time attention decoding, inputting the disturbance-enhanced spatial feature and the multi-step wind speed sequence in the window into the gated recurrent unit, generating a sensitivity matrix dynamically weighting the time step through the parameterized power curve, and outputting the space-time coupled feature;

[0014] S4. Double-scale adaptive feature distillation and fusion, splitting the space-time coupled feature into two branches of local wind turbine scale and global field scale, and outputting the multi-scale fusion feature;

[0015] S5. Loss backpropagation with physical consistency constraint, taking the predicted power obtained from the multi-scale fusion feature through the fully connected network as the benchmark, constructing the sum of the power curve segment deviation penalty and the disturbance sparsity regularization as the loss function, and updating all weight matrices of S2-S4 through loss function backpropagation;

[0016] S6. Edge rolling learning and topology closed loop correction, deploying the model on the edge, recalculating the time-varying adjacency matrix with the on-site weather every fixed time and incrementally training S2-S4, dynamically quantifying and replacing S1 in the closed loop, and continuously evolving the prediction chain.

[0017] Further, the power generation of distributed photovoltaics, the power generation of wind power, the SOC of energy storage, and the electricity consumption data of flexible load need to be preprocessed before fusion; preprocessing mainly includes data cleaning and standardization processing; when cleaning the data, the missing values in the power generation of distributed photovoltaics, the power generation of wind power, the SOC of energy storage, and the electricity consumption data of flexible load are filled by linear interpolation, and the abnormal values obviously exceeding the rated range of the equipment are directly removed and replaced by the mean value of the adjacent period; when standardizing, all data are mapped to the [0, 1] interval.

[0018] Further, the dynamic graph takes the wind turbine as the node, and the edge weight is calculated by the dynamic weather similarity and geographical adjacency.

[0019] Furthermore, the step of constructing the dynamic graph includes:

[0020] S111. Dynamic graph node definition: Each wind turbine in the wind farm is used as the basic node of the dynamic graph, and each node corresponds to a unique wind turbine number.

[0021] Node feature vector construction: The features of each node consist of two parts. The first is the real-time operating status of the wind turbine, which specifically includes indicators such as real-time active power, speed, pitch angle, generator temperature, and gearbox temperature. The second is the surrounding meteorological elements, including indicators such as wind speed, wind direction, atmospheric pressure, relative humidity, and ambient temperature. By integrating these two parts of information, a node feature vector with unified dimensions is finally formed.

[0022] S112. Edge weight calculation: The edge weight is calculated jointly by dynamic meteorological similarity and geographical adjacency.

[0023] Dynamic meteorological similarity Based on the wind direction-power response function, the formula is:

[0024] ;

[0025] in, For wind turbine and The real-time wind direction angle; The power response coefficient;

[0026] Geographic adjacency Based on the Euclidean distance calculation, the formula is as follows:

[0027] ;

[0028] ;

[0029] Fan and The geographical straight-line distance; , This represents the difference between latitude and longitude. This represents the maximum distance between wind turbines within the wind farm.

[0030] The formula for edge weight fusion is:

[0031] ;

[0032] For dynamic weights, the core is the "quantification value of the association strength"; The adjustment coefficient is dynamic; when the wind speed is ≥10m / s, When the wind speed is <10m / s, The edge weight ranges from [0,1], and the larger the value, the stronger the connection between the two nodes.

[0033] S113. Dynamic diagram representation, dynamic diagram As it evolves over time, it can be represented as:

[0034] ;

[0035] Represents a set of nodes; Denotes the set of edges. express The relationships between all wind turbine pairs at any given time are represented by edges, indicating "which wind turbines are related" and "how strong the relationship is". The elements are , Time Node and The edge between;

[0036] S114. Generation and output of time-varying adjacency matrix, time-varying adjacency matrix for A square matrix, the elements of which are defined as:

[0037] ;

[0038] Based on the dynamic changes in meteorological data, the latest meteorological data is used at fixed intervals. , combined with fixed Generate a new time-varying adjacency matrix and recalculate. Generate a new adjacency matrix;

[0039] Sparse matrix storage is used, only preserving The elements are stored in the form of Triples; output the time-varying adjacency matrix at the current time step, which is quantized to represent... The dynamic coupling strength between all wind turbines in the wind farm at any given time.

[0040] Further, S2 includes the following steps:

[0041] S211. Node Feature Difference and Perturbation Region Labeling:

[0042] Based on the time-varying adjacency matrix output by S1, a difference operation is performed between the real-time feature vector and the historical steady-state feature vector of each wind turbine node, as shown in the formula:

[0043] ;

[0044] in, for Time Node eigenvectors; This represents the historical steady-state eigenvectors.

[0045] S212. Mark the disturbance area and determine the set disturbance threshold based on the equipment rated parameters and historical disturbance data statistics. , At that time, mark the node If the disturbance source is identified, it is marked as 1; otherwise, if it is a non-disturbance region, it is marked as 0, thus forming a disturbance label matrix. , Number of extensions;

[0046] S213. Dynamic edge weight aggregation of neighborhood perturbation information: Using the edge weights of the time-varying adjacency matrix, the neighborhood perturbation information of each node is weighted and aggregated. The formula is as follows:

[0047] ;

[0048] Here is the spatial attention weight matrix. For feature dimension, For activation functions; To show the first Typhoon generators are a source of disturbance;

[0049] S214. Output perturbation-enhanced spatial features, aggregate the enhanced features of all nodes to form a perturbation-enhanced spatial feature matrix, which serves as the input for subsequent temporal attention decoding.

[0050] Furthermore, the step of wind speed-power coupling sensitivity-driven time attention decoding in S3 includes:

[0051] S311. Construction of multi-source input sequence: The perturbation enhancement spatial features output by S2 are concatenated with the multi-step wind speed sequence within the same time window as the number of time steps to form the input sequence. :

[0052] ;

[0053] ;

[0054] For perturbation-enhanced spatial feature matrix; This is a multi-step wind speed sequence; For time steps;

[0055] S312. Gated Recurrent Unit (GRU) Feature Extraction: Extracting features from the input sequence... The data is fed into the ingress control loop unit to capture temporal dynamic features. The core computation of the GRU is as follows:

[0056] ;

[0057] in, For a moment The input vector, It updates the value of the gate. It is the sigmoid activation function. It is the reset gate bias vector. It updates the value of the gate. It updates the gate bias vector. Bias for candidate hidden states; It is the weight matrix of the reset gate. It updates the weight matrix of the gate. The weight matrix used to compute the candidate hidden states. for The final hidden state of time. The hidden state at all times This is the candidate hidden state;

[0058] It represents the Hadamah accumulation. The Hadamard product (element-wise multiplication) is used to adjust the output of the reset gate with the element-wise weights of the previous hidden state, and then it is combined with the current input. After concatenation, a linear transformation and activation are performed to obtain candidate hidden states;

[0059] This represents the portion retained from the previous hidden state. This represents the portion that is incorporated from the candidate hidden state into the current hidden state. The sum of the two states yields the complete hidden state at the current moment, which takes into account both historical information and the current input.

[0060] S313. Generation of parameterized power curves and sensitivity matrices: Construct a parameterized power curve model and calculate the wind speed-power coupling sensitivity matrix;

[0061] Parameterized power curve model Represented as:

[0062] ;

[0063] Multi-step wind speed sequence Elements in; This is the weight matrix for the power curve model; This is the bias vector for the power curve model;

[0064] Wind speed-power coupling sensitivity matrix The element is defined as:

[0065] ;

[0066] Indicates the first Wind speed change on the first The degree of influence of step power; , Indicates the first Step and the first The wind speed value at each step;

[0067] S314. Time-step dynamic weighted and spatiotemporally coupled feature output:

[0068] Through the sensitivity matrix Temporal characteristics of GRU output Weighted, the formula is: ;

[0069] To pass through the sensitivity matrix Weighted time characteristics;

[0070] Weighted time features splicing with spatial features This forms a spatiotemporal coupling feature matrix:

[0071] ;

[0072] This is the spatiotemporal coupling feature matrix after splicing.

[0073] Furthermore, the local wind turbine scale extracts fine-grained features of a single wind turbine through 1D convolution, using the following formula:

[0074] ;

[0075] This refers to the dimensional characteristics of a local wind turbine. The weight matrix is ​​for 1D convolution; The convolution bias vector; Indicates the kernel size; This is a 1D convolution operation;

[0076] The global field-level scale aggregates the full-field features through global mean pooling. The formula is:

[0077] ;

[0078] Number of wind turbines; In the spliced ​​spatiotemporal coupling feature matrix, the first... The eigenvectors of the row.

[0079] Furthermore, in S4, the step of outputting multi-scale fusion features through scale factor weighted fusion driven by meteorological abrupt change probability includes:

[0080] S411. Calculation of the probability of meteorological sudden changes: Based on the rate of change of wind speed and wind direction over a recent period, a logistic regression model is used to calculate the probability of meteorological sudden changes. The formula is as follows:

[0081] ;

[0082] in, For wind speed change rate, For wind direction change rate, For model parameters, The sigmoid function is used when wind speed changes drastically and wind direction shifts significantly within a short period of time. The value will approach 1, indicating a high probability of sudden weather changes;

[0083] S412. Scale factor weighted fusion, based on the probability of meteorological abrupt changes. As the global weight scaling factor, the corresponding local weight scaling factor is ,when At higher levels, the global weight scaling factor approaches 1, resulting in larger weights for global field-level features; when... At lower levels, When the value approaches 1, the weight of local wind turbine characteristics is greater;

[0084] S413. Dual-scale feature weighted fusion, which integrates local wind turbine scale features With global field-level scale features Weighted summation is used to output multi-scale fused features, as shown in the formula:

[0085] ;

[0086] in, To preserve both the overall meteorological trends of the wind farm and the fine-grained characteristics of individual wind turbines after integration.

[0087] Further, S5 includes:

[0088] S511. Predictive power calculation, fusing features across multiple scales. The input is a fully connected network, and the output is a predicted function. The fully connected network consists of an input layer, a hidden layer, and an output layer.

[0089] Input layer to hidden layer:

[0090] ;

[0091] This is the weight matrix from the input layer to the hidden layer; The bias vector of the input layer; It is the sigmoid function;

[0092] Hidden layer to output layer:

[0093] ;

[0094] For predicting the power matrix; This is the weight matrix from the hidden layer to the output layer. The bias vector of the hidden layer;

[0095] S512. Power Curve Segment Deviation Penalty: The wind turbine power curve exhibits different characteristics in different wind speed ranges, which are divided into three segments: below the cut-in wind speed, the rated wind speed range, and above the cut-out wind speed. A weighted penalty is applied to the deviation between the predicted power and the actual power in each segment, using the following formula:

[0096]

[0097] To predict the time step; Indicates the segmentation of the power curve; Weights for each segment;

[0098] S513. Perturbation Sparsity Regularization: For the perturbation feature matrix labeled S2, an L1 regularization term is introduced to constrain the sparsity of the perturbation features learned by the model, thus constraining the sparsity of the perturbation region.

[0099] ;

[0100] The perturbation characteristic matrix, The regularization coefficients are minimized. The model retains only the perturbation features that have a significant impact on power prediction, thus reducing noise interference;

[0101] By combining the piecewise deviation penalty of the power curve with the perturbation sparsity regularization, the total loss function is obtained:

[0102] ;

[0103] By minimizing It balances prediction accuracy with model generalization ability, and outputs power prediction results that conform to physical consistency.

[0104] Backpropagation and model optimization employ gradient descent based on the total loss function. The model parameters are updated via backpropagation; for all weight matrix updates involved in S2-S4, the update formula is:

[0105] ;

[0106] For learning rate, loss function For the updated weight matrix The gradient.

[0107] Furthermore, S6 includes:

[0108] S611. Model Deployment and Real-time Update Trigger: The trained wind power prediction model is deployed on the edge computing unit of the wind turbine, and real-time prediction is achieved by utilizing the local computing power of the edge device; an update cycle is set to trigger the model update process; each time an update is performed, the time-varying adjacency matrix in S1 is recalculated based on the latest collected meteorological data; due to the dynamic changes in the meteorological environment of the wind farm, the time-varying adjacency matrix can reflect the real-time changes in the spatial relationship between wind turbines in a timely manner, enabling the model to adapt to the new environment;

[0109] S612. Incremental training and dynamic quantization: The model weights of S2-S4 are fine-tuned using data collected over a period of time, and the parameters of the nearest neighbor nodes closely related to the new data are updated. Through incremental training, the model can quickly incorporate new information, adapt to changes in the operating status of the wind farm, and maintain overall stability and generalization ability.

[0110] The updated weight matrix is ​​quantized from 32-bit floating-point to 8-bit integer using the following formula:

[0111] ;

[0112] in, This is the updated weight matrix; This is the quantized weight matrix; The minimum value of the updated weight matrix; The maximum value of the updated weight matrix; Rounding function;

[0113] S613. Topology closed-loop correction: The quantized new weight matrix and the updated time-varying adjacency matrix are fed back to S1 to replace the original static topology information. This process forms a closed-loop feedback mechanism of "data acquisition-model update-prediction". When the wind turbine layout in the wind farm is adjusted, new wind turbines are added, or long-term changes in the meteorological environment cause changes in the spatial relationship between wind turbines, the model continuously updates its own structure and parameters according to the latest operating data and environmental changes through topology closed-loop correction. This optimizes the dynamic adaptability of the prediction chain, ensures the accuracy and reliability of wind power prediction, and meets the actual operation needs of the wind farm.

[0114] In summary, due to the adoption of the above technical solution, the beneficial technical effects of the invention are as follows:

[0115] To improve the physical rationality of predictions and address the lack of physical constraints, the model employs multi-source heterogeneous data preprocessing to ensure the reliability of input data, combines dynamic graph construction to capture wind turbine correlations, processes features through a spatiotemporal attention mechanism, and finally uses a loss function with physical consistency constraints (power curve segment deviation penalty and disturbance sparsity regularization) to ensure that the model predictions strictly conform to the physical characteristics of wind turbines below the cut-in wind speed, within the rated wind speed range, and above the cut-out wind speed, thus avoiding the situation where existing technology predictions deviate from the actual operating logic.

[0116] Enhanced dynamic adaptability addresses the issue of lag in adapting to dynamic scenarios. The dynamic graph evolves over time, and the edge weights between wind turbines are updated in real time. Dual-scale feature fusion dynamically adjusts feature weights based on the probability of meteorological mutations. The time-varying adjacency matrix is ​​recalculated and incrementally trained at the edge, enabling the model to quickly adapt to changes in meteorological environment and equipment status, overcoming the shortcomings of existing technologies where long-term prediction accuracy decays rapidly.

[0117] It achieves efficient edge deployment and closed loop, solving the problem of inefficient edge deployment. The model is deployed on the edge computing unit, and the weight matrix is ​​compressed to low precision through dynamic quantization, reducing storage and transmission overhead. At the same time, the topology closed loop correction sends the updated matrix back, forming a closed loop link of "data acquisition-model update-prediction", which solves the problem of difficult edge deployment in existing technologies and improves the real-time performance and practicality of prediction. Attached Figure Description

[0118] Figure 1 A flowchart of a short-term wind power prediction method for wind farms based on a spatiotemporal attention mechanism;

[0119] Figure 2 A flowchart for the method of constructing a dynamic graph;

[0120] Figure 3 A flowchart for a perturbation-sensitive spatial attention encoding method;

[0121] Figure 4 A flowchart of a time-attention decoding method driven by wind speed-power coupling sensitivity;

[0122] Figure 5 A flowchart for a multi-scale fusion feature method that outputs features by using scale factor weighted fusion driven by meteorological mutation probability;

[0123] Figure 6 A flowchart of the loss backpropagation method for physical consistency constraints;

[0124] Figure 7 The flowchart shows the edge-end rolling learning and topology closure correction method. Detailed Implementation

[0125] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0126] like Figure 1 As shown, a short-term wind power prediction method for wind farms based on a spatiotemporal attention mechanism is proposed:

[0127] S1. Multi-source heterogeneous data fusion and dynamic graph construction: Data from distributed optical harvesting is collected and dynamic graphs are constructed by fusing power generation from distributed photovoltaics, wind power, energy storage SOC, and power consumption data from flexible loads. In the dynamic graph, wind turbines are used as nodes, and node features include real-time operating status and meteorological elements. The edge weights of the dynamic graph are updated in real time based on dynamic meteorological similarity (such as wind direction-power response function) and geographical adjacency, forming a time-varying adjacency matrix that characterizes the meteorological coupling strength between wind turbines, thus quantifying the dynamic coupling strength between wind turbines.

[0128] S2. Perturbation-sensitive spatial attention encoding, based on a time-varying adjacency matrix, performs perturbation-sensitive spatial convolution on each wind turbine node; first, it calculates the difference between the current feature and the historical steady-state feature to mark the perturbation region; then, it aggregates the neighborhood perturbation information through dynamic edge weights and outputs the perturbation-enhanced feature.

[0129] S3. Wind speed-power coupling sensitivity-driven time attention decoding sends the disturbance enhancement spatial features and the multi-step wind speed sequence within the same window to the inlet control loop unit. The sensitivity matrix is ​​generated by parameterized power curve and the time steps are dynamically weighted to output the spatiotemporal coupling features.

[0130] S4. Dual-scale adaptive feature distillation and fusion: The spatiotemporal coupled features are split into two branches: local wind turbine scale and global field scale. The multi-scale fused features are output by weighted fusion through scale factors driven by meteorological change probability (dynamically generated by meteorological change probability).

[0131] S5. Backpropagation of loss for physical consistency constraint: Based on the predicted power obtained by multi-scale fusion features through a fully connected network, construct the sum of power curve segment deviation penalty and perturbation sparsity regularization as the loss function, and backpropagate the loss function to update all weight matrices of S2-S4.

[0132] S6. Edge-end rolling learning and topology closed-loop correction: The trained model is deployed on the edge computing unit of the wind turbine. The edge nodes correct the time-varying adjacency matrix with real-time meteorological data at fixed intervals and send it back to S1, forming a continuously evolving closed-loop prediction link.

[0133] The preprocessing mainly includes data cleaning and standardization. Data cleaning involves filling missing values ​​in four types of data (distributed photovoltaic power generation, wind power generation, energy storage SOC, and electricity consumption data of flexible loads) using linear interpolation (calculated based on two valid data points before and after the data). Outliers that significantly exceed the rated range of the equipment are directly removed and replaced with the average of adjacent time periods. In the standardization process, all data are mapped to the [0,1] interval.

[0134] like Figure 2 As shown, the steps for constructing the dynamic graph include:

[0135] S111. Dynamic graph node definition: Each wind turbine in the wind farm is used as the basic node of the dynamic graph, and each node corresponds to a unique wind turbine number.

[0136] Node feature vector construction: The features of each node consist of two parts. The first is the real-time operating status of the wind turbine, which specifically includes indicators such as real-time active power, speed, pitch angle, generator temperature, and gearbox temperature. The second is the surrounding meteorological elements, including indicators such as wind speed, wind direction, atmospheric pressure, relative humidity, and ambient temperature. By integrating these two parts of information, a node feature vector with unified dimensions is finally formed.

[0137] S112. Edge weight calculation: The edge weight is calculated jointly by dynamic meteorological similarity and geographical adjacency.

[0138] Among them, dynamic meteorological similarity Based on the wind direction-power response function, the formula is:

[0139] ;

[0140] in, For wind turbine and The real-time wind direction angle is calculated by the absolute value of the difference between the two wind directions, ranging from 0 to 180°. The power response coefficient is obtained by fitting historical data over a period of time, with a value ranging from 0.8 to 1.2, reflecting the degree of influence of wind direction differences on power correlation.

[0141] Geographic adjacency Based on the Euclidean distance calculation, the formula is as follows:

[0142] ;

[0143] ;

[0144] Fan and The geographical straight-line distance; , This represents the difference between latitude and longitude. This represents the maximum distance between wind turbines within the wind farm.

[0145] The formula for edge weight fusion is:

[0146] ;

[0147] For dynamic weights, the core is the "quantification value of the association strength"; The adjustment coefficient is dynamic; when the wind speed is ≥10m / s, (Meteorological factors dominate the coupling relationship); when the wind speed is <10m / s, (Geographical factors dominate the coupling relationship); the edge weight ranges from [0,1], and the larger the value, the stronger the association between the two nodes;

[0148] S113. Dynamic diagram representation, dynamic diagram As it evolves over time, it can be represented as:

[0149] ;

[0150] Represents a set of nodes; Denotes the set of edges. express The relationships between all wind turbine pairs at any given time are represented by edges, indicating "which wind turbines are related" and "how strong the relationship is". The elements are , Time Node (fan) )and (fan) The edges between );

[0151] S114. Generation and output of time-varying adjacency matrix, time-varying adjacency matrix for A square matrix, the elements of which are defined as:

[0152] ;

[0153] Based on the dynamic changes in meteorological data, the latest meteorological data is used at fixed intervals. , combined with fixed Generate a new time-varying adjacency matrix and recalculate. Generate a new adjacency matrix;

[0154] Sparse matrix storage is used, only preserving Elements with a weight less than 0.05 are considered to have no significant correlation and are stored in the following format: Triples; output the time-varying adjacency matrix at the current time step, which is quantized to represent... The dynamic coupling strength between all wind turbines in the wind farm at any given time.

[0155] like Figure 3 As shown, step S2 includes the following steps:

[0156] S211. Node Feature Difference and Perturbation Region Labeling:

[0157] Based on the time-varying adjacency matrix output by S1, a difference operation is performed between the real-time feature vector and the historical steady-state feature vector of each wind turbine node, as shown in the formula:

[0158] ;

[0159] in, for Time Node The feature vector (including real-time operating status and meteorological elements); This represents the historical steady-state feature vector. Through differential calculation, the changes in the current wind turbine node state compared to its steady-state state can be clearly highlighted, providing crucial information for subsequent identification of disturbance areas. For example, if the active power of a wind turbine node deviates significantly from its historical steady-state value at the current moment, then... The corresponding active power dimension value will change significantly;

[0160] S212. Mark the disturbance area and determine the set disturbance threshold based on the equipment rated parameters and historical disturbance data statistics. , At that time, mark the node If the disturbance source is identified, it is marked as 1; otherwise, if it is a non-disturbance region, it is marked as 0, thus forming a disturbance label matrix. , This refers to the number of individual wind turbines. In actual wind farm environments, wind turbines may be disturbed by sudden strong winds, equipment failures, etc. By comparing with a threshold, these wind turbine nodes in abnormal states can be quickly and accurately marked.

[0161] S213. Dynamic edge weight aggregation of neighborhood perturbation information: Using the edge weights of the time-varying adjacency matrix, the neighborhood perturbation information of each node is weighted and aggregated. The formula is as follows:

[0162] ;

[0163] Here is the spatial attention weight matrix. For feature dimension, For activation functions; To show the first Typhoon generators are a source of disturbance;

[0164] Calculation process:

[0165] First, based on the time-varying adjacency matrix Middle node Other nodes The edge weights, relative to the original features of the nodes. and disturbance characteristics Perform a weighted summation; then, use the spatial attention weight matrix. By performing dimensional attention adjustments on the summation result, the model can automatically learn the importance of different feature dimensions in spatial correlation; finally, after... The activation function filters out negative values ​​and highlights effective features, thus obtaining the nodes. Enhanced features after considering neighborhood perturbation information and spatial attention ;

[0166] S214. Output the perturbation-enhanced spatial features, aggregate the enhanced features of all nodes to form a perturbation-enhanced spatial feature matrix, which serves as the input for subsequent temporal attention decoding. The perturbation-enhanced spatial feature matrix integrates the perturbation and correlation information of each wind turbine node in the entire wind farm in the spatial dimension, laying the foundation for further mining of spatiotemporal coupling features. For example, the perturbation-enhanced spatial feature matrix can intuitively show the differences in the degree of spatial impact of perturbations on different wind turbine nodes, as well as the strength of their interrelationships.

[0167] In S3, since wind speed has a direct and critical impact on wind power, combining wind speed sequence with spatial features allows the model to consider information in both spatial and temporal dimensions, comprehensively capturing the evolution of wind farm status over time and the interaction between different wind turbines.

[0168] like Figure 4 As shown, the steps of time-attention decoding driven by wind speed-power coupling sensitivity in S3 include:

[0169] S311. Multi-source input sequence construction: The perturbation enhancement spatial features output from S2 are concatenated with the multi-step wind speed sequence (time steps) within the same time window to form the input sequence. :

[0170] ;

[0171] ;

[0172] For perturbation-enhanced spatial feature matrix; This is a multi-step wind speed sequence; For time steps;

[0173] S312. Gated Recurrent Unit (GRU) Feature Extraction: Extracting features from the input sequence... The data is fed into the ingress control loop unit to capture temporal dynamic features. The core computation of the GRU is as follows:

[0174] ;

[0175] in, For a moment The input vector, It updates the value of the gate. It is the sigmoid activation function. It is the reset gate bias vector. It updates the value of the gate. It updates the gate bias vector. Bias for candidate hidden states; It is the weight matrix of the reset gate. It updates the weight matrix of the gate. The weight matrix used to compute the candidate hidden states. for The final hidden state of time. The hidden state at all times This is the candidate hidden state;

[0176] It represents the Hadamah accumulation. The Hadamard product (element-wise multiplication) is used to adjust the output of the reset gate with the element-wise weights of the previous hidden state, and then it is combined with the current input. After concatenation, a linear transformation and activation are performed to obtain candidate hidden states;

[0177] This represents the portion retained from the previous hidden state. This represents the portion that is incorporated from the candidate hidden state into the current hidden state. The sum of the two states yields the complete hidden state at the current moment, which takes into account both historical information and the current input.

[0178] GRU can effectively handle long-term dependencies in time series data through the control of reset gates and update gates. In the process, GRU performs in-depth mining of the time information in the input sequence to extract key features of wind speed and turbine status changes over time, such as capturing the periodic changes in wind speed and the response patterns of turbine power at different times.

[0179] S313. Generation of parameterized power curves and sensitivity matrices: Construct a parameterized power curve model and calculate the wind speed-power coupling sensitivity matrix;

[0180] Parameterized power curve model Represented as:

[0181] ;

[0182] Multi-step wind speed sequence Elements in; This is the weight matrix for the power curve model; This is the bias vector for the power curve model;

[0183] Wind speed-power coupling sensitivity matrix The element is defined as:

[0184] ;

[0185] Indicates the first Wind speed change on the first The degree of influence of step power; , Indicates the first Step and the first The wind speed value at each time step; by constructing a parameterized power curve model, the complex nonlinear relationship between wind speed and power can be described more accurately; while the sensitivity matrix quantifies the difference in the impact of wind speed changes on power at different time steps. For example, in some periods, a small change in wind speed may cause a large fluctuation in power, which will be reflected in the larger corresponding element value in the sensitivity matrix.

[0186] S314. Time-step dynamic weighted and spatiotemporally coupled feature output:

[0187] Through the sensitivity matrix Temporal characteristics of GRU output Weighted, the formula is:

[0188] ;

[0189] To pass through the sensitivity matrix Weighted time characteristics;

[0190] Weighted time features splicing with spatial features This forms a spatiotemporal coupling feature matrix:

[0191] ;

[0192] This is the spatiotemporal coupling feature matrix after concatenation;

[0193] Dynamic weighting allows the model to emphasize or weaken features at different time steps based on wind speed-power coupling sensitivity, highlighting features at time steps that have a greater impact on power prediction. Finally, spatiotemporal feature concatenation yields coupled features containing rich spatiotemporal information, providing a more comprehensive data foundation for subsequent analysis and prediction.

[0194] The local wind turbine scale extracts fine-grained features of a single wind turbine through 1D convolution, using the following formula:

[0195] ;

[0196] This refers to the dimensional characteristics of a local wind turbine. The weight matrix is ​​for 1D convolution; The convolution bias vector; Indicates the kernel size; This is a 1D convolution operation; 1D convolution extracts fine-grained features of a single wind turbine. Without changing the feature dimension, it locally extracts and integrates the features of each wind turbine node, preserving the detailed features of a single node (such as individual fault signals). For example, when a wind turbine experiences an abnormal gearbox temperature... The temperature characteristics of this node will be given high weight, making The features of the corresponding dimension are amplified, accurately capturing individual abnormal signals.

[0197] Global field-level scale, aggregating full-field features through global mean pooling. The formula is:

[0198] ;

[0199] Number of wind turbines; In the spliced ​​spatiotemporal coupling feature matrix, the first... The eigenvectors of the row;

[0200] Global mean pooling averages the features of all wind turbine nodes to obtain a feature representation that reflects the macroscopic state of the entire wind farm. For example, it can obtain global information such as the average wind speed and average power of the entire wind farm, which can be used to grasp the overall operating trend of the wind farm.

[0201] like Figure 5 As shown, in S4, the step of outputting multi-scale fusion features through scale factor weighted fusion driven by meteorological abrupt change probability includes:

[0202] S411. Calculation of the probability of meteorological sudden changes: Based on the rate of change of wind speed and wind direction over a recent period, a logistic regression model is used to calculate the probability of meteorological sudden changes. The formula is as follows:

[0203] ;

[0204] in, For wind speed change rate, For wind direction change rate, For model parameters, The sigmoid function is used when wind speed changes drastically and wind direction shifts significantly within a short period of time. The value will approach 1, indicating a high probability of sudden weather changes;

[0205] S412. Scale factor weighted fusion, based on the probability of meteorological abrupt changes. As the global weight scaling factor, the corresponding local weight scaling factor is ,when At higher levels (due to sudden weather changes), the global weight scaling factor approaches 1, and the global field-level feature weights are larger; when At lower (stable weather) levels, When the value approaches 1, the weight of local wind turbine characteristics is greater;

[0206] S413. Dual-scale feature weighted fusion, which integrates local wind turbine scale features With global field-level scale features Weighted summation is used to output multi-scale fused features, as shown in the formula:

[0207] ;

[0208] in, To preserve both the overall meteorological trends of the wind farm and the fine-grained characteristics of individual wind turbines after integration.

[0209] like Figure 6 As shown, S5 includes:

[0210] S511. Predictive power calculation, fusing features across multiple scales. The input is a fully connected network, and the output is a predicted function. The fully connected network consists of an input layer, a hidden layer, and an output layer.

[0211] Input layer to hidden layer:

[0212] ;

[0213] This is the weight matrix from the input layer to the hidden layer; The bias vector of the input layer; It is the sigmoid function;

[0214] Hidden layer to output layer:

[0215] ;

[0216] For predicting the power matrix; This is the weight matrix from the hidden layer to the output layer. The bias vector of the hidden layer;

[0217] S512. Power Curve Segment Deviation Penalty: The wind turbine power curve exhibits different characteristics in different wind speed ranges, which are divided into three segments: below the cut-in wind speed, the rated wind speed range, and above the cut-out wind speed. A weighted penalty is applied to the deviation between the predicted power and the actual power in each segment, using the following formula:

[0218] ;

[0219] To predict the time step; Indicates the segmentation of the power curve; Assigning weights to each segment, the rated wind speed range has a higher weight due to its stable power output and significant contribution to the power grid. The highest, such as Cut-in wind speed below and above the cut-out wind speed Relatively low, such as ;

[0220] S513. Perturbation Sparsity Regularization: For the perturbation feature matrix labeled S2, an L1 regularization term is introduced to constrain the sparsity of the perturbation features learned by the model, thus constraining the sparsity of the perturbation region.

[0221] ;

[0222] The perturbation characteristic matrix, The regularization coefficients are minimized. The model retains only the perturbation features that have a significant impact on power prediction, thus reducing noise interference;

[0223] By combining the piecewise deviation penalty of the power curve with the perturbation sparsity regularization, the total loss function is obtained:

[0224] ;

[0225] By minimizing It balances prediction accuracy with model generalization ability, and outputs power prediction results that conform to physical consistency.

[0226] Backpropagation and model optimization employ gradient descent based on the total loss function. The model parameters are updated via backpropagation; for all weight matrices involved in S2-S4 (such as those in spatial adjacency matrix learning, power curve models, fully connected networks, etc.), the update formula is:

[0227] ;

[0228] For learning rate, loss function For the updated weight matrix The gradient.

[0229] like Figure 7 As shown, S6 includes:

[0230] S611. Model Deployment and Real-time Update Trigger: The trained wind power prediction model is deployed on the wind turbine edge computing unit, and real-time prediction is achieved by utilizing the local computing capabilities of the edge device; the update cycle is set to trigger the model update process; each time an update is performed, the time-varying adjacency matrix in S1 is recalculated based on the latest collected meteorological data (wind direction, wind speed, air pressure, etc.).

[0231] Due to the dynamic changes in the meteorological environment of wind farms, the time-varying adjacency matrix can reflect the real-time changes in the spatial relationship between wind turbines (such as changes in the airflow influence range caused by wind direction changes, which in turn change the power transmission and influence relationship between wind turbines), enabling the model to adapt to the new environment.

[0232] S612. Incremental training and dynamic quantization: The model weights of S2-S4 are fine-tuned using data collected over a period of time. To reduce computation, only the parameters of the nearest neighbor nodes closely related to the new data are updated. For example, when the meteorological conditions in the vicinity of a wind turbine change abruptly, only the relevant parameters of that wind turbine and its neighboring wind turbines in the model are updated. Through incremental training, the model can quickly incorporate new information, adapt to changes in the operating status of the wind farm, and maintain overall stability and generalization ability.

[0233] To reduce storage and transmission overhead for edge devices, the updated weight matrix is ​​quantized from 32-bit floating-point to 8-bit integer (INT8). The quantization formula is as follows:

[0234] ;

[0235] in, This is the updated weight matrix; This is the quantized weight matrix; The minimum value of the updated weight matrix; The maximum value of the updated weight matrix; Rounding function;

[0236] The quantized weight matrix retains the relative size of the weights during the quantization process, which significantly reduces data storage and transmission bandwidth without significantly affecting model performance, thereby improving the efficiency of model operation at the edge.

[0237] S613. Topology closed-loop correction: The quantized new weight matrix and the updated time-varying adjacency matrix are fed back to S1 to replace the original static topology information. This process forms a closed-loop feedback mechanism of "data acquisition-model update-prediction". When the wind turbine layout in the wind farm is adjusted, new wind turbines are added, or long-term changes in the meteorological environment cause changes in the spatial relationship between wind turbines, the model continuously updates its own structure and parameters according to the latest operating data and environmental changes through topology closed-loop correction. This optimizes the dynamic adaptability of the prediction chain, ensures the accuracy and reliability of wind power prediction, and meets the actual operation needs of the wind farm.

[0238] The above description is a preferred embodiment of the invention and is not intended to limit the scope of the invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A method for short-term wind power prediction in wind farms based on a spatiotemporal attention mechanism, characterized in that, Includes the following steps: S1. Multi-source heterogeneous data fusion and dynamic graph construction: Integrate the power generation of distributed photovoltaics, the power generation of wind power, the energy storage SOC and the electricity consumption data of flexible loads, construct a dynamic graph that evolves over time, and output a time-varying adjacency matrix. S2. Perturbation-sensitive spatial attention encoding uses a time-varying adjacency matrix to differentially label perturbation regions of node features, dynamically aggregates neighborhood perturbation information using edge weights, and outputs perturbation-enhanced spatial features. S3. Wind speed-power coupling sensitivity-driven time attention decoding sends the disturbance enhancement spatial features and the multi-step wind speed sequence within the same window to the inlet control loop unit. The sensitivity matrix is ​​generated by parameterized power curve and the time steps are dynamically weighted to output the spatiotemporal coupling features. S4. Dual-scale adaptive feature distillation and fusion: The spatiotemporal coupled features are split into two branches: local wind turbine scale and global field scale. The scale factor weighted fusion driven by meteorological change probability is used to output multi-scale fused features. S5. Backpropagation of loss for physical consistency constraint: Based on the predicted power obtained by multi-scale fusion features through a fully connected network, a loss function is constructed with the sum of power curve segment deviation penalty and perturbation sparsity regularization. The loss function is backpropagated to update all weight matrices of S2-S4. S6. Edge-end rolling learning and topology closed-loop correction: The model is deployed at the edge. At fixed intervals, the time-varying adjacency matrix is ​​recalculated using on-site weather data and S2-S4 are incrementally trained. After dynamic quantization, S1 is replaced by a closed loop, and the prediction chain continues to evolve.

2. The method for short-term wind power prediction of wind farms based on a spatiotemporal attention mechanism according to claim 1, characterized in that, Before merging, the power generation data of distributed photovoltaic, wind power, energy storage SOC, and flexible load need to be preprocessed. The preprocessing mainly includes data cleaning and standardization. During data cleaning, missing values ​​in the power generation data of distributed photovoltaic, wind power, energy storage SOC, and flexible load are filled using linear interpolation. Abnormal values ​​that are significantly outside the rated range of the equipment are directly removed and replaced with the average of adjacent time periods. During standardization, all data are mapped to the [0,1] interval.

3. The method for short-term wind power prediction of wind farms based on a spatiotemporal attention mechanism according to claim 1, characterized in that, The dynamic graph uses wind turbines as nodes, and the edge weights are calculated jointly by dynamic meteorological similarity and geographical adjacency.

4. A method for short-term wind power prediction of wind farms based on a spatiotemporal attention mechanism according to any one of claims 1 and 3, characterized in that, The steps for constructing the dynamic graph include: S111. Dynamic graph node definition: Each wind turbine in the wind farm is used as the basic node of the dynamic graph, and each node corresponds to a unique wind turbine number. Node feature vector construction: The features of each node consist of two parts. The first is the real-time operating status of the wind turbine, which specifically includes real-time active power, speed, pitch angle, generator temperature, and gearbox temperature. The second is the surrounding meteorological elements, including wind speed, wind direction, atmospheric pressure, relative humidity, and ambient temperature. By integrating these two parts of information, a node feature vector with unified dimensions is finally formed. S112. Edge weight calculation: The edge weight is calculated jointly by dynamic meteorological similarity and geographical adjacency. Dynamic meteorological similarity Based on the wind direction-power response function, the formula is: ； in, For wind turbine and The real-time wind direction angle; The power response coefficient; Geographic adjacency Based on the Euclidean distance calculation, the formula is as follows: ； ； Fan and The geographical straight-line distance; , This represents the difference between latitude and longitude. This represents the maximum distance between wind turbines within the wind farm. The formula for edge weight fusion is: ； For dynamic weights, the core is the "quantification value of the association strength"; The adjustment coefficient is dynamic; when the wind speed is ≥10m / s, When the wind speed is <10m / s, The edge weight ranges from [0,1], and the larger the value, the stronger the connection between the two nodes. S113. Dynamic diagram representation, dynamic diagram As it evolves over time, it can be represented as: ； Represents a set of nodes; Denotes the set of edges. express The relationships between all wind turbine pairs at any given time are represented by edges, indicating "which wind turbines are related" and "how strong the relationship is". The elements are , Time Node and The edge between; S114. Generation and output of time-varying adjacency matrix, time-varying adjacency matrix for A square matrix, the elements of which are defined as: ； Based on the dynamic changes in meteorological data, the latest meteorological data is used at fixed intervals. , combined with fixed Generate a new time-varying adjacency matrix and recalculate. Generate a new adjacency matrix; Sparse matrix storage is used, only preserving The elements are stored in the form of Triples; output the time-varying adjacency matrix at the current time step, which is quantized to represent... The dynamic coupling strength between all wind turbines in the wind farm at any given time.

5. The method for short-term wind power prediction of wind farms based on a spatiotemporal attention mechanism according to claim 1, characterized in that, S2 includes the following steps: S211. Node Feature Difference and Perturbation Region Labeling: Based on the time-varying adjacency matrix output by S1, a difference operation is performed between the real-time feature vector and the historical steady-state feature vector of each wind turbine node, as shown in the formula: ； in, for Time Node eigenvectors; This represents the historical steady-state eigenvectors. S212. Mark the disturbance area and determine the set disturbance threshold based on the equipment rated parameters and historical disturbance data statistics. , At that time, mark the node If the disturbance source is identified, it is marked as 1; otherwise, if it is a non-disturbance region, it is marked as 0, thus forming a disturbance label matrix. , Number of extensions; S213. Dynamic edge weight aggregation of neighborhood perturbation information: Using the edge weights of the time-varying adjacency matrix, the neighborhood perturbation information of each node is weighted and aggregated. The formula is as follows: ； Here is the spatial attention weight matrix. For feature dimension, For activation functions; To show the first Typhoon generators are a source of disturbance; S214. Output perturbation-enhanced spatial features, aggregate the enhanced features of all nodes to form a perturbation-enhanced spatial feature matrix, which serves as the input for subsequent temporal attention decoding.

6. The method for short-term wind power prediction of wind farms based on spatiotemporal attention mechanism according to claim 1, characterized in that, The steps of time-attention decoding driven by wind speed-power coupling sensitivity in S3 include: S311. Construction of multi-source input sequence: The perturbation enhancement spatial features output by S2 are concatenated with the multi-step wind speed sequence within the same time window as the number of time steps to form the input sequence: S312. Gated Recurrent Unit (GRU) Feature Extraction: The input sequence is fed into the gated recurrent unit (GRU) to capture temporal dynamic features. S313. Generation of parameterized power curves and sensitivity matrices: Construct a parameterized power curve model and calculate the wind speed-power coupling sensitivity matrix; Parameterized power curve model Represented as: ； Multi-step wind speed sequence Elements in; This is the weight matrix for the power curve model; This is the bias vector for the power curve model; Wind speed-power coupling sensitivity matrix The element is defined as: ； Indicates the first Wind speed change on the first The degree of influence of step power; , Indicates the first Step and the first The wind speed value at each step; S314. Time-step dynamic weighting and spatiotemporal coupling feature output; Through the sensitivity matrix Temporal characteristics of GRU output Weighted, the formula is: ； To pass the sensitivity matrix Weighted time characteristics; Weighted time features splicing with spatial features This forms a spatiotemporal coupling feature matrix: ； This is the spatiotemporal coupling feature matrix after splicing.

7. The method for short-term wind power prediction of wind farms based on spatiotemporal attention mechanism according to claim 1, characterized in that, The local wind turbine scale extracts fine-grained features of a single wind turbine through 1D convolution, using the following formula: ； This refers to the dimensional characteristics of a local wind turbine. The weight matrix is ​​for 1D convolution; The convolution bias vector; Indicates the kernel size; This is a 1D convolution operation; The global field-level scale aggregates the full-field features through global mean pooling. The formula is: ； Number of wind turbines; In the spliced ​​spatiotemporal coupling feature matrix, the first... The eigenvectors of the row.

8. The method for short-term wind power prediction of wind farms based on spatiotemporal attention mechanism according to claim 1, characterized in that, In S4, the step of outputting multi-scale fusion features through scale factor weighted fusion driven by meteorological abrupt change probability includes: S411. Calculation of the probability of meteorological sudden changes: Based on the rate of change of wind speed and wind direction over a recent period, a logistic regression model is used to calculate the probability of meteorological sudden changes. The formula is as follows: ； in, For wind speed change rate, For wind direction change rate, For model parameters, The sigmoid function is used when wind speed changes drastically and wind direction shifts significantly within a short period of time. The value will approach 1, indicating a high probability of sudden weather changes; S412. Scale factor weighted fusion, based on the probability of meteorological abrupt changes. As the global weight scaling factor, the corresponding local weight scaling factor is ,when At higher levels, the global weight scaling factor approaches 1, resulting in larger weights for global field-level features; when... At lower levels, When the value approaches 1, the weight of local wind turbine characteristics is greater; S413. Dual-scale feature weighted fusion, which integrates local wind turbine scale features With global field-level scale features Weighted summation is used to output multi-scale fused features, as shown in the formula: ； in, To preserve both the overall meteorological trends of the wind farm and the fine-grained characteristics of individual wind turbines after integration.

9. The method for short-term wind power prediction of wind farms based on spatiotemporal attention mechanism according to claim 1, characterized in that, S5 includes: S511. Predictive power calculation, fusing features across multiple scales. Input a fully connected network, output the predicted power. A fully connected network consists of an input layer, a hidden layer, and an output layer. S512. Power Curve Segment Deviation Penalty: The wind turbine power curve exhibits different characteristics in different wind speed ranges, which are divided into three segments: below the cut-in wind speed, the rated wind speed range, and above the cut-out wind speed. A weighted penalty is applied to the deviation between the predicted power and the actual power in each segment, using the following formula: ； To predict the time step; Indicates the segmentation of the power curve; Weights for each segment; For predicting the power matrix; S513. Perturbation Sparsity Regularization: For the perturbation feature matrix labeled S2, an L1 regularization term is introduced to constrain the sparsity of the perturbation features learned by the model, thus constraining the sparsity of the perturbation region. ； The perturbation characteristic matrix, The regularization coefficients are minimized. The model retains only the perturbation features that have a significant impact on power prediction, thus reducing noise interference; By combining the piecewise deviation penalty of the power curve with the perturbation sparsity regularization, the total loss function is obtained: ； By minimizing It balances prediction accuracy with model generalization ability, and outputs power prediction results that conform to physical consistency. Backpropagation and model optimization employ gradient descent based on the total loss function. The model parameters are updated via backpropagation; for all weight matrix updates involved in S2-S4, the update formula is: ； For learning rate, loss function For the updated weight matrix The gradient.

10. The method for short-term wind power prediction of wind farms based on spatiotemporal attention mechanism according to claim 1, characterized in that, The S6 mentioned above includes: S611. Model Deployment and Real-time Update Trigger: The trained wind power prediction model is deployed on the edge computing unit of the wind turbine, and real-time prediction is achieved by utilizing the local computing power of the edge device; an update cycle is set to trigger the model update process; each time an update is performed, the time-varying adjacency matrix in S1 is recalculated based on the latest collected meteorological data; due to the dynamic changes in the meteorological environment of the wind farm, the time-varying adjacency matrix can reflect the real-time changes in the spatial relationship between wind turbines in a timely manner, enabling the model to adapt to the new environment; S612. Incremental training and dynamic quantization: The model weights of S2-S4 are fine-tuned using data collected over a period of time, and the parameters of the nearest neighbor nodes closely related to the new data are updated. Through incremental training, the model can quickly incorporate new information, adapt to changes in the operating status of the wind farm, and maintain overall stability and generalization ability. The updated weight matrix is ​​quantized from 32-bit floating-point to 8-bit integer using the following formula: ； in, This is the updated weight matrix; This is the quantized weight matrix; The minimum value of the updated weight matrix; The maximum value of the updated weight matrix; Rounding function; S613. Topology closed-loop correction: The quantized new weight matrix and the updated time-varying adjacency matrix are fed back to S1 to replace the original static topology information. This process forms a closed-loop feedback mechanism of "data acquisition-model update-prediction". When the wind turbine layout in the wind farm is adjusted, new wind turbines are added, or long-term changes in the meteorological environment cause changes in the spatial relationship between wind turbines, the model continuously updates its own structure and parameters according to the latest operating data and environmental changes through topology closed-loop correction. This optimizes the dynamic adaptability of the prediction chain, ensures the accuracy and reliability of wind power prediction, and meets the actual operation needs of the wind farm.

Images