Wind power plant group power space-time prediction method based on dynamic graph convolution and frequency domain learning
By combining dynamic graph convolution and frequency domain learning, the problems of dynamic spatial dependencies among wind farm clusters and insufficient utilization of frequency domain information are solved, achieving high-precision and robust wind power prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 烟台哈尔滨工程大学研究院
- Filing Date
- 2026-01-28
- Publication Date
- 2026-05-08
AI Technical Summary
Existing wind power forecasting methods are unable to effectively capture the dynamic spatial dependencies between wind farm clusters and neglect video domain information, resulting in insufficient forecast accuracy, especially poor robustness under complex weather conditions.
A spatiotemporal prediction method for wind farm cluster power based on dynamic graph convolution and frequency domain learning is constructed. The method captures the time-varying topological relationship of wind farm cluster through adaptive graph network and refines the prediction results by using frequency domain learning module, thus integrating spatiotemporal features and frequency domain information.
It achieves high-precision spatiotemporal collaborative prediction of wind farm power, improving prediction accuracy and robustness, especially demonstrating higher accuracy and stability over long periods and under complex weather conditions.
Smart Images
Figure CN122000884A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of power system operation and control and artificial intelligence application technology, and in particular to a method for predicting the power generation of new energy sources using deep learning technology. Background Technology
[0002] With the advancement of the global energy transition, wind power, as an important component of clean energy, has seen its installed capacity continuously grow. However, wind power is characterized by strong volatility, randomness, and intermittency, posing significant challenges to the safe and stable operation and dispatch of the power system when large-scale wind power grid integration is implemented. High-precision wind power forecasting is a key technology for alleviating peak-shaving pressure and improving the grid's absorption capacity.
[0003] Current wind power forecasting methods are mainly categorized into physical modeling, statistical methods, and artificial intelligence methods. Early studies often employed time series models such as Autoregressive Integral Moving Average (ARIMA), Support Vector Machine (SVM), or Long Short-Term Memory (LSTM) networks. While these methods can capture the historical time dependence of individual sites, they typically neglect the correlation between geographically adjacent wind farms. In reality, significant spatial correlations exist between wind farm clusters; wind speed changes in upstream wind farms usually affect downstream wind farms after a certain time delay.
[0004] To leverage spatial features, researchers have introduced convolutional neural networks (CNNs) to model wind farm data as grid images for processing. However, wind farms are typically geographically distributed irregularly, often non-Euclidean, making it difficult for CNNs to effectively handle this graph topology. In recent years, graph convolutional networks (GCNs) have been widely used due to their advantages in processing unstructured data. Existing GCN-based prediction methods typically construct static adjacency matrices based on the geographical distance or Pearson correlation coefficient between wind farms. However, in actual operation, the correlation between wind farms is not constant but exhibits significant dynamic time-varying characteristics with changes in meteorological conditions such as wind direction and speed (e.g., the direction of the wake effect changes with wind direction, causing a reversal in upstream and downstream dependencies). Fixed static graph structures struggle to capture this complex dynamic spatiotemporal coupling mechanism, limiting further improvements in prediction accuracy.
[0005] Furthermore, most existing deep learning prediction models only perform end-to-end modeling in the time domain, directly outputting power predictions, while often neglecting the valuable information contained in the model prediction residuals. Prediction errors typically contain high-frequency random components or systematic biases that are not captured by the main model. These features are difficult to identify in the time domain, but often exhibit learnable patterns in the frequency domain. Current solutions lack effective mechanisms for in-depth frequency domain mining and correction of prediction errors, resulting in insufficient robustness of the models when faced with drastic wind speed fluctuations.
[0006] In summary, how to construct a graph structure that can adaptively capture the dynamic spatial dependencies of wind farm clusters and fully utilize frequency domain information to refine the prediction results is a technical challenge that urgently needs to be solved in the field of wind power prediction.
[0007] To address the aforementioned problems, this invention provides a spatiotemporal prediction method for wind farm cluster power based on dynamic graph convolution and frequency domain learning. This invention achieves accurate characterization of the time-varying topological relationships of wind farm clusters by constructing a dynamic graph structure that includes node embedding and gating mechanisms. Simultaneously, it innovatively introduces a frequency domain learning module to map time-domain residuals to the frequency domain for feature decoupling and correction, thereby effectively overcoming the poor robustness of traditional models under complex meteorological conditions and achieving high-precision spatiotemporal collaborative prediction of wind farm cluster power. Summary of the Invention
[0008] Symbols and definitions:
[0009] A. Exponents, sets, and matrices:
[0010]
[0011] B. Variables and constants
[0012]
[0013] C. Abbreviation
[0014]
[0015] The purpose of this invention is to provide a spatiotemporal prediction method for wind farm cluster power based on dynamic graph convolution and frequency domain learning, which can accurately capture spatiotemporal dynamic coupling characteristics and effectively mine global patterns in the frequency domain. This method proposes a novel spatiotemporal prediction architecture for wind farm cluster power. First, addressing the multidimensional heterogeneity and periodicity of wind power data, a multidimensional spatiotemporal embedding model is established to achieve unified encoding of time, space, and meteorological elements. Then, considering the real-time changes in the topology between wind turbines with meteorological conditions, a dynamic graph network construction method based on an adaptive adjacency matrix is proposed to dynamically capture time-varying spatial dependencies. Second, addressing the shortcomings of traditional time-domain modeling in extracting long-range dependencies and global patterns, a frequency domain feature learning model based on discrete Fourier transform is established to achieve decoupling analysis of the periodic and trend components of the power sequence. This method has advantages in effectiveness, superiority, and robustness in solving problems such as the difficulty in fusing spatiotemporal features and characterizing dynamic patterns in wind power prediction, and meets the practical needs of safe and stable grid operation under large-scale wind power grid integration.
[0016] To achieve the above objectives, this invention provides a spatiotemporal prediction method for wind farm cluster power based on dynamic graph convolution and frequency domain learning. The method is characterized by constructing a deep learning model that integrates an adaptive graph network and a frequency domain multilayer perceptron for wind power prediction. The steps are as follows:
[0017] S1. Obtain historical power data and meteorological data of wind farm clusters, construct a multi-dimensional spatiotemporal embedding layer, jointly encode time period, node location and meteorological characteristics, and generate an initial spatiotemporal feature vector containing multi-dimensional information.
[0018] S2. Construct a Dynamic Adaptive Graph Network (DAP) to capture the inherent topological structure between wind turbines using a learnable global adjacency matrix and generate time-varying spatial dependencies by combining local temporal features, thereby extracting the dynamic spatial features of the wind farm cluster.
[0019] S3. Design a frequency-domain multilayer perceptron (MLP) to project node features into the frequency domain through discrete Fourier transform, decouple the periodic and trend components of the power sequence at the node dimension, and restore them through inverse transform after complex number processing to extract the frequency domain evolution features of the power sequence.
[0020] S4. The extracted dynamic spatial features are fused with the frequency domain evolution features, and the power prediction values of the wind farm cluster at future moments are generated through the prediction output layer.
[0021] Preferred spatiotemporal correlation analysis module:
[0022] (1) Time-related
[0023] Mutual information (MI) is used to characterize the dependence between two random variables; a larger MI value indicates a stronger correlation. When mutual information is close to 0, it indicates that the two variables are approximately independent. For time series, the time correlation can be quantified by calculating the mutual information between the current value and its lagged values, as shown in the following formula:
[0024]
[0025] In the formula: is the joint probability distribution, and are the marginal probability distributions. In time series analysis, the temporal correlation is assessed by analyzing the mutual information between the sequence and its lagged sequences.
[0026] (2) Lag time correlation
[0027]
[0028] In the formula: represents the value of the time series at the current time, represents the value at the lagged time, is the joint probability distribution of and , and and are the marginal probability distributions of and , respectively.
[0029] (3) Spatial correlation
[0030] To quantify spatial correlation, this paper uses the Pearson correlation coefficient to measure the linear correlation between power sequences of different wind farms, as shown in the following formula:
[0031]
[0032] In the formula: , These are the power values of different wind turbines or wind farms. , It is its mean. A correlation greater than 0.8 indicates a strong correlation, while a correlation less than 0.5 indicates a weak correlation. A correlation of ≤0.8 indicates a moderate correlation. A value ≤0.5 indicates a weak correlation.
[0033] The preferred dynamic adaptive graph network module of the Spatiotemporal Frequency Network (STF-Net):
[0034] (1) Definition of graph convolution
[0035] To extract spatial structure features from wind farms, this paper employs a graph convolutional approach based on a first-order Chebyshev polynomial approximation.
[0036] The announcement is as follows:
[0037]
[0038] In the formula: Represents the graph structure matrix; yes The degree matrix; It is the input feature matrix; It is the output feature matrix; Indicates weight; Indicates bias.
[0039] Spatial feature extraction is performed using a graph convolutional network (GCN). For the th... Layer input features Using the generated adaptive adjacency matrix To conduct map dissemination.
[0040] (2) Adaptive dynamic graph network
[0041] a) Global adaptive adjacency matrix
[0042] This paper employs an adaptive adjacency matrix to learn global spatial dependencies from data. The specific method is as follows: it is implemented by randomly initializing two adjacency matrices with learnable parameters. The adaptive adjacency matrix is:
[0043]
[0044] In the formula: Embedded in the source node; Embed for the target node.
[0045] b) Local dynamic graph filtering
[0046] This paper proposes a dynamic graph convolution method for learning local graph information. At each time step, a local adjacency matrix is generated using historical window data to reflect the time-varying local spatial dependencies between nodes. At each time step, the short-term spatial dependencies at the current moment are learned based on the historical window data, and the input signal is transformed using a dynamically filtered MLP to obtain the dynamically filtered signal.
[0047]
[0048] In the formula: This is the filtered dynamic signal.
[0049] c) Dynamic graph information fusion
[0050] Within the time series, the dynamic graph features of each time step are fused to generate a dynamic feature representation of the node across the entire time window. Element-wise multiplication is then performed on the sum to generate dynamic graph information.
[0051]
[0052] In the formula: express Dynamic graph information at the time step.
[0053] d) Dynamic graph convolution update
[0054]
[0055] In the formula: This is the adjacency matrix of the dynamic graph at the current time step; It is a degree matrix; For time Input time; The weight matrix is a learnable matrix; This is the activation function.
[0056] The preferred frequency domain feature extraction module of the Spatiotemporal Frequency Network (STF-Net):
[0057] (1) Sequence frequency domain transformation
[0058] To extract the periodicity and trend patterns of the time series, a Fourier transform is used to project the original input into the frequency domain. The formula is as follows:
[0059]
[0060] In the formula: For frequency variables; For integration variables; The imaginary unit is defined as the square root of -1. The first part is... The real part, abbreviated as The second half is The imaginary part, abbreviated as Therefore, in the equation It can be represented as .
[0061] (2) Complex feature decoupling learning
[0062] Frequency-domain MLPs have been redesigned for complex frequency components to effectively capture key time-series modes and compress energy through a global view. Formally, for complex inputs… Given a complex weight matrix and complex deviation Then, a frequency domain MLP can be customized as:
[0063]
[0064] In the formula: This is the final output; Indicates the first layer, It is an activation function. Because and Since they are all complex numbers, according to the rules of complex number multiplication, the equation can be further generalized as follows:
[0065]
[0066] In the formula: , According to the above formula, the MLP in the frequency domain can be realized by calculating the real and imaginary parts of the frequency components separately, and then superimposing them into a complex number to obtain the final result.
[0067] (3) Frequency domain reconstruction
[0068] After processing the frequency domain signal, the inverse Fourier transform is used to restore it to the time domain, so as to retain the processing results of periodic information and global trend.
[0069] Attached Figure Description
[0070] Figure 1 This is an overall flowchart of a spatiotemporal prediction method for wind farm cluster power based on dynamic graph convolution and frequency domain learning provided by an embodiment of the present invention;
[0071] Figure 2 This is a schematic diagram of the spatiotemporal correlation analysis of wind farm groups in an embodiment of the present invention;
[0072] Figure 3 This is an overall model architecture diagram of the Spatiotemporal Frequency Network (STF-Net) provided in the embodiments of the present invention;
[0073] Figure 4 This is a schematic diagram of the structure of the multidimensional spatiotemporal embedding module in an embodiment of the present invention;
[0074] Figure 5 This is a schematic diagram of the structure of the dynamic adaptive graph network module in an embodiment of the present invention;
[0075] Figure 6 This is a visualization heatmap of the adaptive adjacency matrix in an embodiment of the present invention;
[0076] Figure 7 This is a schematic diagram of the structure of the frequency domain feature extraction module (frequency domain MLP) in an embodiment of the present invention;
[0077] Figure 8 This is a schematic diagram of the frequency domain feature decoupling mechanism in an embodiment of the present invention;
[0078] Figure 9 This is a schematic diagram of the geographical distribution of the wind farm clusters selected for the experiment in this embodiment of the invention;
[0079] Figure 10This is a bar chart or line graph comparing the errors of the method and the benchmark model of this invention under different prediction step sizes.
[0080] Figure 11 This is a curve comparing the predicted wind power output with the actual value during a typical time period using the method of this embodiment of the invention. Detailed Implementation
[0081] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0082] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.
[0083] This invention provides a spatiotemporal prediction method for wind farm power based on dynamic graph convolution and frequency domain learning. This method constructs a deep learning model called STF-Net (Spatio-Temporal Frequency Network), which integrates the ability of dynamic graph networks to capture time-varying spatial dependencies and the ability of frequency domain learning to extract global long-range time patterns.
[0084] A. Test System and Parameter Settings
[0085] This example uses Dataset 1 from the Penmanshiel Wind Farm in the UK, covering the period from January 1st to June 30th, 2021, with data intervals of 10 minutes. This dataset contains historical data and spatial location information for 14 wind turbines in a specific wind farm, including wind speed and direction information for each turbine. Dataset 2 is from a wind farm in Anhui Province, covering the period from January 1st to June 30th, 2021, with data intervals of 5 minutes. This dataset contains historical data and spatial location information for 12 wind farms within a specific region. Root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation metrics. Smaller RMSE and MAE values indicate that the predicted results are closer to the actual values, and the model's prediction accuracy is higher.
[0086] B. Results Analysis
[0087] To verify the effectiveness of the proposed wind farm cluster power spatiotemporal prediction method (STF-Net) based on dynamic graph convolution and frequency domain learning, simulation experiments were conducted using two real-world wind farm datasets, and the experimental results were compared and analyzed with various mainstream benchmark models.
[0088] 1. Dataset Description: The experimental data are from the Penmanshiel Wind Farm in the UK (dataset 1) and a group of wind farms in a certain area of Anhui, China (dataset 2).
[0089] Dataset 1: Contains historical data and spatial location information of 14 wind turbines, with a sampling interval of 10 minutes.
[0090] Dataset 2: Contains historical data and spatial location information for 12 wind farms, with a sampling interval of 5 minutes. Detailed statistical information for the dataset (including maximum, minimum, mean, and sample division) is shown in Table 1.
[0091] Table 1. Dataset Description
[0092] Table 1 Dataset description
[0093]
[0094] 2. Comparative Analysis of Experimental Results This invention selected eight benchmark methods for comparison, including temporal models (FeDformer, PatchTsT, FreTs) and spatiotemporal models (STGCN, GraphWaveNet, StemGNN, STID, D2STGNN). Mean absolute error (MAE) and root mean square error (RMSE) were used as evaluation metrics.
[0095] Improved prediction accuracy (corresponding to Tables 2 and 3) As shown in Tables 2 (comparison of experimental results for dataset 1) and 3 (comparison of experimental results for dataset 2), the model of this invention achieved the best performance in all experimental settings.
[0096] Table 2 Comparison Experiments with Dataset 1
[0097] Table 2 Dataset 1 Comparative experiment
[0098]
[0099] Table 3 Comparison Experiment of Dataset 2
[0100] Table 3 Dataset 2 Comparative experiments
[0101]
[0102] In both short-term (3-step) and medium- to long-term (6-step and 12-step) forecasts, the error metrics of this model are significantly lower than those of other benchmark models.
[0103] Specifically, in Dataset 1, compared with the suboptimal model, the RMSE of the model of this invention is improved by 4.55% in 3-step prediction, 9.8% in 6-step prediction, and 15.66% in 12-step prediction. This indicates that the present invention can more accurately capture the long-term time dependence of wind power.
[0104] Compared to models such as STGCN and GraphWaveNet that only focus on time-domain modeling, this invention introduces frequency-domain MLP to mine global periodic and trend features, which significantly improves prediction accuracy.
[0105] Multi-step prediction stability (corresponding) Figures 8 to 11 ) manual Figures 8 to 11 The MAE and RMSE curves of different models under different prediction steps are shown respectively.
[0106] As the number of prediction steps increases, the errors of all models show an upward trend, reflecting the cumulative effect of non-stationarity in wind power prediction.
[0107] Compared with other models, the error growth curve of the model in this invention is the flattest, with a much smaller increase than the comparative models. This indicates that by combining spatiotemporal coding, dynamic graph modeling, and frequency domain learning, this invention can effectively control error accumulation and exhibit higher robustness and accuracy in multi-step prediction tasks.
[0108] C. Analysis of ablation test results
[0109] To verify the contributions of each core module in this invention, ablation experiments were designed, removing the frequency domain MLP module (wo / fremlp), the adaptive graph convolution module (wo / acgcn), and the spatiotemporal embedding layer (wo / stemb), respectively. The experimental results are recorded in Table 4 (ablation experiments on dataset 1) and Table 5 (ablation experiments on dataset 2).
[0110] Table 4 Ablation Experiments in Dataset 1
[0111] Table 4 Dataset 1 Ablation Experiment
[0112]
[0113] Table 5 Ablation Experiments in Dataset 2
[0114] Table 5 Dataset 2 Ablation Experiment
[0115]
[0116] The role of frequency domain MLP: After removing frequency domain MLP, the model's performance degrades significantly at longer prediction steps (e.g., 1 hour, 2 hours), with RMSE increasing by approximately 7-9%. This demonstrates that frequency domain MLP, through Fourier transform, extracts the periodicity and trend patterns of time series data, which is crucial for predictions over long time spans.
[0117] The role of adaptive graph convolution: Removing the adaptive graph convolution module has the greatest negative impact on model performance, especially in long-term span predictions, where RMSE increases by nearly 5%. This indicates that the module effectively models the complex dynamic spatial relationships between wind turbines through a global adaptive adjacency matrix and dynamic graph convolution.
[0118] The role of the spatiotemporal embedding layer: Removing this layer weakens the model's ability to capture temporal, spatial, and feature interaction relationships, leading to an increase in both MAE and RMSE.
Claims
1. A spatiotemporal prediction method for wind farm cluster power based on dynamic graph convolution and frequency domain learning, characterized in that, Wind power prediction is performed by constructing a deep learning model that integrates adaptive graph networks and frequency domain multilayer perceptrons. The steps are as follows: S1. Obtain historical power data and meteorological data of wind farm clusters, construct a multi-dimensional spatiotemporal embedding layer, jointly encode time period, node location and meteorological characteristics, and generate an initial spatiotemporal feature vector containing multi-dimensional information. S2. Construct a Dynamic Adaptive Graph Network (DAP) to capture the inherent topological structure between wind turbines using a learnable global adjacency matrix and generate time-varying spatial dependencies by combining local temporal features, thereby extracting the dynamic spatial features of the wind farm cluster. S3. Design a frequency-domain multilayer perceptron (MLP) to project node features into the frequency domain through discrete Fourier transform, decouple the periodic and trend components of the power sequence at the node dimension, and restore them through inverse transform after complex number processing to extract the frequency domain evolution features of the power sequence. S4. The extracted dynamic spatial features are fused with the frequency domain evolution features, and the power prediction values of the wind farm cluster at future moments are generated through the prediction output layer.
2. The spatiotemporal prediction method for wind farm cluster power based on dynamic graph convolution and frequency domain learning according to claim 1, characterized in that, Spatiotemporal correlation analysis module: (1) Time-related Mutual information (MI) is used to characterize the dependence between two random variables; a larger MI value indicates a stronger correlation. When mutual information is close to 0, it indicates that the two variables are approximately independent. For time series, the time correlation can be quantified by calculating the mutual information between the current value and its lagged values, as shown in the following formula: In the formula: is the joint probability distribution, and are the marginal probability distributions. In time series analysis, temporal correlation is assessed by analyzing the mutual information between the sequence and its lagged sequences. (2) Lag time correlation In the formula: represents the value of the time series at the current time, represents the value at the lagged time, is the joint probability distribution of and , and and are the marginal probability distributions of and , respectively. (3) Spatial correlation To quantify spatial correlation, this paper uses the Pearson correlation coefficient to measure the linear correlation between power sequences of different wind farms, as shown in the following formula: In the formula: , These are the power values of different wind turbines or wind farms. , It is its mean. A correlation greater than 0.8 indicates a strong correlation, while a correlation less than 0.5 indicates a weak correlation. A correlation of ≤0.8 indicates a moderate correlation. A value ≤0.5 indicates a weak correlation.
3. The spatiotemporal prediction method for wind farm cluster power based on dynamic graph convolution and frequency domain learning according to claim 1, characterized in that, Dynamic adaptive graph network module of the Spatiotemporal Frequency Network (STF-Net): (1) Definition of graph convolution To extract spatial structure features from wind farms, this paper employs a graph convolution approach based on a first-order Chebyshev polynomial approximation. The announcement is as follows: In the formula: Represents the graph structure matrix; yes The degree matrix; It is the input feature matrix; It is the output feature matrix; Indicates weight; Indicates bias. Spatial feature extraction is performed using a graph convolutional network (GCN). For the th... Layer input features Using the generated adaptive adjacency matrix To conduct map dissemination. (2) Adaptive dynamic graph network a) Global adaptive adjacency matrix This paper employs an adaptive adjacency matrix to learn global spatial dependencies from data. The specific method is as follows: it is implemented by randomly initializing two adjacency matrices with learnable parameters. The adaptive adjacency matrix is: In the formula: Embedded in the source node; Embed for the target node. b) Local dynamic graph filtering This paper proposes a dynamic graph convolution method for learning local graph information. At each time step, a local adjacency matrix is generated using historical window data to reflect the time-varying local spatial dependencies between nodes. At each time step, the short-term spatial dependencies at the current moment are learned based on the historical window data, and the input signal is transformed using a dynamically filtered MLP to obtain the dynamically filtered signal. In the formula: This is the filtered dynamic signal. c) Dynamic graph information fusion Within the time series, the dynamic graph features of each time step are fused to generate a dynamic feature representation of the node across the entire time window. Element-wise multiplication is then performed on the sum to generate dynamic graph information. In the formula: express Dynamic graph information at the time step. d) Dynamic graph convolution update In the formula: This is the adjacency matrix of the dynamic graph at the current time step; It is a degree matrix; For time Input time; The weight matrix is a learnable matrix; This is the activation function.
4. The spatiotemporal prediction method for wind farm cluster power based on dynamic graph convolution and frequency domain learning according to claim 1, characterized in that, Frequency domain feature extraction module of Spatiotemporal Frequency Network (STF-Net): (1) Sequence frequency domain transformation To extract the periodicity and trend patterns of the time series, a Fourier transform is used to project the original input into the frequency domain. The formula is as follows: In the formula: For frequency variables; For integration variables; The imaginary unit is defined as the square root of -1. The first part is... The real part, abbreviated as The second half is The imaginary part, abbreviated as Therefore, in the equation It can be represented as . (2) Complex feature decoupling learning Frequency-domain MLPs have been redesigned for complex frequency components to effectively capture key time-series modes and compress energy through a global view. Formally, for complex inputs… Given a complex weight matrix and complex deviation Then, a frequency domain MLP can be customized as: In the formula: This is the final output; Indicates the first layer, It is an activation function. Because and Since they are all complex numbers, according to the rules of complex number multiplication, the equation can be further generalized as follows: In the formula: , According to the above formula, the MLP in the frequency domain can be realized by calculating the real and imaginary parts of the frequency components separately, and then superimposing them into a complex number to obtain the final result. (3) Frequency domain reconstruction After processing the frequency domain signal, the inverse Fourier transform is used to restore it to the time domain, so as to retain the processing results of periodic information and global trend.