Wind power cluster power prediction method based on global information graph fusion continuous learning

Abstract

The application discloses a wind power cluster power prediction method based on global information graph fusion continuous learning, and the method comprises the following steps: step one, calculating the space-time correlation coefficient; step two, generating a dynamic adjacency matrix; step three, graph embedding based on Laplace mapping; step four, extracting the features of the embedded vector through a clustering algorithm; step five, constructing a global information graph; step six, denoising the global information graph by using an adaptive time sequence label smoothing method; step seven, establishing a graph attention network model of the fusion continuous learning strategy; and step eight, model training and testing. The application can accurately predict the power generation of large wind power clusters, and helps to improve the economy and safety of the operation of the wind power field. Meanwhile, the method provides reliable data support for power grid dispatching and wind power consumption, guarantees the safe and stable operation of the power system, and promotes the high proportion of renewable energy access and the construction of a green and low-carbon energy system.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and in particular to a wind power cluster power prediction method based on global information graph fusion and continuous learning. Background Technology

[0002] The uncertainty of wind energy poses a serious challenge to attempts to integrate large-scale wind power generation into modern power systems. Therefore, accurate wind power forecasting is one of the foundations for ensuring the stable operation of the power system and promoting the consumption of wind power.

[0003] Compared to power prediction for a single wind farm, power prediction for a wind farm cluster faces greater modeling challenges. On one hand, there is significant spatial correlation among multiple wind farms within a cluster, with meteorological conditions such as wind speed and direction often exhibiting strong synchronicity and propagation. On the other hand, wind power itself is highly nonlinear and time-varying; therefore, the model must not only possess strong time-series modeling capabilities but also effectively capture the spatial dependencies between wind farms. Furthermore, the differences in terrain, equipment parameters, and meteorological characteristics among different wind farms necessitate that the model possess a certain degree of generalization ability and adaptability when handling cluster predictions.

[0004] Currently, methods for predicting wind farm cluster power mainly fall into two categories. The first category is independent prediction based on individual wind farms. Each wind farm is treated as an independent entity, and its power time-series model is constructed separately. The prediction results are then aggregated to obtain the cluster prediction value. This type of method primarily relies on time-series models in deep learning, such as Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), and Transformers. The second category is joint prediction of multiple wind farms. This type explicitly considers the spatial correlation and synergistic effects between wind farms during modeling. Through joint modeling, geographical, meteorological, and historical power relationships between wind farms can be utilized to improve overall prediction accuracy. Representative methods include Graph Neural Networks (GNN) and Spatiotemporal Graph Neural Networks (ST-GNN), which can effectively capture the spatiotemporal dependencies between wind farms.

[0005] Currently, most wind farm cluster power prediction methods based on traditional spatiotemporal graph neural networks employ static graph structures, meaning they use fixed input wind turbine nodes and edges, neglecting the dynamic changes in the relationships between wind turbines within a wind farm due to factors such as wind direction and weather. In reality, the connections between wind turbines exhibit significant spatiotemporal nonlinearity and complexity, potentially displaying drastically different structural characteristics at different time scales. Static graphs struggle to effectively capture these dynamic relationships, making it difficult for models to reflect the graph structure evolution caused by meteorological changes, thus impacting prediction accuracy. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a wind power cluster power prediction method based on global information graph fusion and continuous learning, which addresses the shortcomings of the existing technology.

[0007] The technical solution adopted by this invention to solve its technical problem is:

[0008] This invention provides a wind power cluster power prediction method based on global information graph fusion and continuous learning, the method comprising the following steps:

[0009] Step 1: Collect wind speeds at each wind turbine in the wind farm cluster, and propose three levels: absolute level, growth level, and fluctuation level, to describe the wind speed relationship between wind turbines. Then, calculate the spatiotemporal correlation coefficient by weighting the three levels.

[0010] Step 2: Based on the spatiotemporal correlation coefficient, calculate the spatiotemporal correlation coefficient between all wind turbines at each time point and construct the adjacency matrix;

[0011] Step 3: Calculate the degree matrix of the nodes based on the adjacency matrix, and construct the Laplacian matrix accordingly; map the high-dimensional graph structure of the Laplacian matrix to the low-dimensional space by minimizing the optimization objective function to obtain the embedding vector;

[0012] Step 4: Cluster the embedding vectors and extract their features, including cluster labels and cluster centers;

[0013] Step 5: Based on the obtained cluster labels and cluster centers, replace the graph of each category with its corresponding cluster center graph to obtain the global information graph;

[0014] Step 6: Adopt the adaptive time series label smoothing method to optimize the clustering labels of the global information map. By introducing a sliding window strategy, the context information of the adjacent time in the time series is used to dynamically correct the initial clustering results, reduce the label discreteness and jump phenomenon caused by noise, and obtain the denoised global information map.

[0015] Step 7: Use a graph attention network to fit the graph information input of the denoised global information graph, fuse a long short-term memory network and introduce a continuous learning strategy to establish a graph attention network model that incorporates the continuous learning strategy.

[0016] Step 8: Input the test set data into the graph attention network model that incorporates the continuous learning strategy to train the model. After training, input the historical wind power cluster power data into the model to obtain the predicted power data for future times.

[0017] Furthermore, in step one of this invention, the absolute level, growth level, and fluctuation level are specifically as follows:

[0018] Absolute grade: Used to describe the direct relationship between wind turbines at a given time point, reflecting the static relationship between the turbines at the current moment. Its formula is:

[0019]

[0020] In the formula, v it and v jt Let be the wind speeds of the i-th and j-th wind turbines at time t, respectively. This represents the absolute ranking between the i-th and j-th wind turbines at time t;

[0021] Growth Level: Used to describe the developmental relationship between wind turbines at a given time point, reflecting the dynamic relationship between wind turbines at the current moment. Its formula is:

[0022]

[0023] In the formula, Δv i,t and Δv j,t Let v represent the wind speed changes of the i-th and j-th wind turbines from time t-1 to time t, respectively. it and v jt Let be the wind speeds of the i-th and j-th wind turbines at time t-1, respectively. This represents the growth level between the i-th and j-th wind turbines at time t;

[0024] Fluctuation level: Used to describe the fluctuation level of wind speed change between wind turbines at a given time point, and its formula is:

[0025]

[0026] In the formula, and S i S j These are the average and standard deviation of the wind speed of the i-th and j-th wind turbines within a certain period interval, respectively. This represents the fluctuation level between the i-th and j-th wind turbines at time t.

[0027] Furthermore, the method for calculating the spatiotemporal correlation coefficient in step one of the present invention is specifically as follows:

[0028] The spatiotemporal correlation coefficient is composed of a weighted combination of absolute level, growth level, and fluctuation level, and its expression is as follows:

[0029]

[0030] In the formula, α, β, and δ are the weighting coefficients for the absolute level, growth level, and fluctuation level, respectively. Let be the spatiotemporal correlation coefficient between the i-th and j-th wind turbines at time t.

[0031] Furthermore, the method for constructing the adjacency matrix in step two of this invention is specifically as follows:

[0032] Using the spatiotemporal correlation coefficient established in step one, a dynamic adjacency matrix is ​​generated. The spatiotemporal correlation coefficient between all wind turbines at each time node is calculated, and an adjacency matrix is ​​constructed. A threshold method is introduced: when the spatiotemporal correlation coefficient is greater than or equal to the threshold, the value is set to 1; otherwise, it is 0, which indicates whether there is an edge connection between wind turbines at that time node. The continuous spatiotemporal correlation coefficient matrix is ​​converted into a binary adjacency matrix, thereby constructing a dynamic adjacency matrix sequence that describes the changes in network topology.

[0033] Furthermore, the method for calculating the embedding vector in step three of this invention is specifically as follows:

[0034] The degree matrix of the nodes is calculated using the adjacency matrix of the graph obtained in step two, and the Laplacian matrix is ​​constructed from this matrix. By minimizing the objective function, the high-dimensional graph structure is mapped to a low-dimensional space, resulting in the embedding vector.

[0035]

[0036] In the formula, W ij The elements of the weight matrix of the graph, y i and y j It is the representation of node i and node j in low-dimensional space.

[0037] Furthermore, the clustering method in step four of this invention is specifically as follows:

[0038] The K-Means algorithm is used to cluster graph embedding vectors. First, cluster centers are randomly initialized. Then, the positions of the cluster centers are updated iteratively. In each iteration, the algorithm assigns each data point to its nearest cluster center and recalculates the center position of each cluster. The formula for updating the cluster centers is:

[0039]

[0040] In the formula, |C i | represents the number of sample points in the i-th cluster. This process is repeated until the update magnitude of the cluster centers is less than a preset threshold or the maximum number of iterations is reached. Finally, the cluster label of the graph corresponding to each embedding vector and the positions of all cluster centers are output.

[0041] Furthermore, the adaptive time-series label smoothing method used in step six of this invention is specifically as follows:

[0042] For each time step t, first check the current label. tThe label is determined by whether it is inconsistent with the labels of its immediate and neighboring neighborhoods. If the current label is different from the labels of its immediate and neighboring neighborhoods, the label smoothing process is triggered. The label smoothing determination function Smooth(t) is defined as follows:

[0043]

[0044] In the formula, label t Label representing time step t t-1 and label t+1 These represent the labels for the previous and next time steps, respectively. When the smoothing condition is met, it means that the current label may contain noise or inconsistencies, which will be corrected through a sliding window.

[0045] The smoothing process relies on a sliding window of size ω, which extends to the labels of ω time steps before and after the current time step t. The specific range of the window is as follows:

[0046]

[0047] In the formula, Represents a unique set of labels within the window. It is a tag Number of times it appears in the window;

[0048] If the smoothing condition Smooth(t) is satisfied, meaning the current label is inconsistent with its neighboring labels, then the current label is updated to the most common label within the window. The label update rule is expressed as follows:

[0049]

[0050] In the formula, After the smoothing process, the labels at all time steps are updated to the smoothed label sequence `smoothed_labels`, which is represented as follows:

[0051]

[0052] Based on the smoothed sequence label information of all time steps, the global information graph is reconstructed by replacing the adjacency matrix of each time step with the adjacency matrix of the cluster center of the corresponding cluster label, thus obtaining the denoised global information graph.

[0053] where c t =smoothed_labels t ,

[0054] In the formula, Let represent the denoised adjacency matrix at time step t. c represents the cluster center t The corresponding standard adjacency matrix, c t =smoothed_labels t This represents the cluster label after smoothing at time step t.

[0055] Furthermore, the graph attention network model integrating continuous learning strategies established in step seven of this invention is specifically as follows:

[0056] A graph attention network is used to fit the unpredictable graph information input of the global information graph. Specifically, this involves obtaining the denoised adjacency matrix sequence. Then, it is compared with the corresponding node feature sequence X. t Combined, it is used as input to the Graph Attention Network (GAT).

[0057] At each time step t, the Graph Attention Network (GAT) first bases its input features X on... t ∈R N×F and the corresponding adjacency matrix Calculate the attention weight for each node and aggregate information from neighboring nodes. The core operations of the graph attention layer are as follows:

[0058]

[0059] In the formula, Let W represent the input features of node i at time step t, W be the learnable linear transformation matrix, α be the attention weight vector, and N be the input features of node i at time step t. i Let H be the set of neighbors of node i, and σ(·) be a non-linear activation function. After processing by the graph attention module, the resulting spatial representation is H. t ∈R N×F′ .

[0060] To further model the dynamic evolution of node representations in the sequence, a graph is embedded into the sequence H. t The input is fed into the LSTM network. The LSTM unit models the time dependency through a gating mechanism; the computation process is as follows:

[0061] i t =σ(W i H t +U i h t-1 +b i ),

[0062] f t =σ(W f H t +U f h t-1 +b f ),

[0063] o t =σ(W o H t +U o h t-1 +b o ),

[0064] c t =f t ⊙c t-1 +i t ⊙tanh(W c H t +U c h t-1 +b c ),

[0065] h t =o t ⊙tanh(c t )

[0066] Among them, i t f t o t These are the input gate, forget gate, and output gate, respectively. t For the unit state, h t This represents the current hidden state, where ⊙ denotes element-wise multiplication. All weight matrices W * U * With bias term b * These are the parameters that the model needs to learn.

[0067] The GAT-LSTM model ultimately yields a series of hidden states h with spatiotemporal awareness. t The predicted value is then output through a fully connected layer.

[0068] Furthermore, the method for model training in step eight of this invention is specifically as follows:

[0069] The test set data is input into the model to train it. At time step t, the model receives a new batch of training data D. t ={(x1,y1),(x2,y2),...,(x n ,y n )}, where x i The input data is y. i These are the corresponding labels. The training loss of the model is expressed as:

[0070]

[0071] In the formula, f(x) i ;θ) is the model's prediction, L(f(x) i ;θ),y i) is the loss function, and θ is the model parameter.

[0072] First, a sliding window is set up. Each time the model receives a new batch of data, the data window is slid to ensure that only the latest fixed-size data window is used for training.

[0073] Then, experience replay is performed. As the window slides and generates new data, the current window data is added to the experience replay buffer, which stores a limited number of past windows. During each training iteration, in addition to the current window data, the model also samples past data from the buffer for training. Combined with the experience replay mechanism, loss calculation depends not only on the current window data but also on data from the experience replay buffer. The experience replay buffer stores data from multiple past windows, and during replay, a portion of data is randomly sampled from these historical experiences for training. The specific training objective is:

[0074]

[0075] In the formula, k is the amount of data sampled from the experience playback buffer, (x i ,y i The samples are historical data, taken from previous training windows. In this way, the model optimizes the loss for both the current task and past tasks simultaneously.

[0076] After training, historical wind power cluster power data is input into the model to obtain predicted power data for future times.

[0077] This invention provides a wind power cluster power prediction method system based on global information graph fusion and continuous learning, comprising:

[0078] Memory, used to store executable computer programs;

[0079] When the processor executes the executable computer program stored in memory, it implements the above-mentioned wind power cluster power prediction method based on global information graph fusion and continuous learning.

[0080] The beneficial effects of this invention are:

[0081] 1. This invention presents a method for predicting the power output of large-scale wind farm clusters. This method can accurately predict the power generation of large-scale wind farm clusters, helping to improve the economic efficiency and safety of wind farm operation. Simultaneously, this method provides reliable data support for grid dispatching and wind power consumption, ensuring the safe and stable operation of the power system and promoting the high-proportion integration of renewable energy and the construction of a green and low-carbon energy system.

[0082] 2. This invention proposes a method for constructing an adjacency matrix (graph) to describe the relationships between wind power clusters. This method first calculates the spatiotemporal correlation coefficient between wind turbines using wind speed data. Then, based on the cluster labels and cluster centers obtained using graph embedding and clustering algorithms, the graph for each category is replaced with its corresponding cluster center graph, resulting in a global information graph.

[0083] 3. This invention proposes an adaptive temporal label smoothing method to optimize the clustering labels of the global information graph. By introducing a sliding window strategy, the initial clustering results are dynamically corrected by utilizing the contextual information of nearby time points in the time series, thereby reducing label discreteness and jump phenomena caused by noise and obtaining a denoised global information graph.

[0084] 4. This invention proposes a wind farm cluster power prediction method based on a Globally AwareDynamic Spatiotemporal Graph Continual Learning framework (GADSG-CL). This method inputs the denoised global information graph into a spatiotemporal graph neural network (GAT-LSTM) and introduces a continuous learning strategy during model training. Through a sliding window and experience replay mechanism, it effectively fits the dynamic graph input, avoiding the model's forgetting of historical dynamic features. Comprehensive experiments on short-term and medium-term prediction tasks were conducted on two different public datasets. In experiments on two public short-to-medium-term prediction datasets, the GADSG-CL method demonstrates superior performance compared to existing methods. Attached Figure Description

[0085] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0086] Figure 1 This is a schematic diagram illustrating the generation of the dynamic adjacency matrix in the spatiotemporal feature map of this patent.

[0087] Figure 2 This is a schematic diagram of graph embedding based on Laplacian mapping in this patent.

[0088] Figure 3 This is a schematic diagram of the extraction of cluster centers in this patent.

[0089] Figure 4 This is a schematic diagram of replacing the corresponding cluster center map with the global information map in this patent.

[0090] Figure 5This is a schematic diagram of the continuous learning strategy based on sliding window and experience replay mechanism in this patent.

[0091] Figure 6 This is a schematic diagram of the overall framework of this patent.

[0092] Figure 7 The results of this patent's experimental evaluation metrics on the DSWE dataset at 12, 24, and 36 steps are as follows: a) Predicting the results of the next 12 steps in 12 steps, b) Predicting the results of the next 24 steps in 24 steps, and c) Predicting the results of the next 36 steps in 36 steps.

[0093] Figure 8 Visualize the experimental results of this patent on the DSWE dataset at 12, 24, and 36 steps. a) 12 steps predict the next 12 steps, b) 24 steps predict the next 24 steps, c) 36 steps predict the next 36 steps.

[0094] Figure 9 The results of this patent's experiments on the SDWPF dataset are as follows: a) Predicting the results of the next 12 steps in 12 steps; b) Predicting the results of the next 24 steps in 24 steps; c) Predicting the results of the next 36 steps in 36 steps.

[0095] Figure 10 Visualization of the experimental results of this patent on the SDWPF dataset at 12, 24, and 36 steps. a) 12 steps predict the next 12 steps, b) 24 steps predict the next 24 steps, c) 36 steps predict the next 36 steps. Detailed Implementation

[0096] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0097] Example 1

[0098] This invention provides a wind farm cluster power prediction method based on a global information graph fusion continuous learning strategy, which includes the following steps.

[0099] Step 1: Calculate the spatiotemporal correlation coefficient. Wind speed, as the most direct input affecting wind turbine power generation, directly reflects its magnitude and changes in power output. To accurately describe the wind speed relationship between wind turbines and the spatiotemporal variation of wind speed, this invention proposes three levels to describe the wind speed relationship between wind turbines:

[0100] (1) Absolute level: The absolute level describes the direct relationship between the wind speeds of the wind turbines at that point in time, and reflects the static relationship between the wind turbines at the current moment.

[0101]

[0102] In the formula, v it and v jt Let be the wind speeds of the i-th and j-th wind turbines at time t, respectively. This represents the absolute ranking between the i-th and j-th wind turbines at time t.

[0103] (2) Growth level: The growth level describes the development relationship between wind turbines at this time point, reflecting the dynamic relationship between wind turbines at the current moment.

[0104]

[0105] In the formula, Δv i,t and Δv j,t Let v represent the wind speed changes of the i-th and j-th wind turbines from time t-1 to time t, respectively. it and v jt Let be the wind speeds of the i-th and j-th wind turbines at time t-1, respectively. This represents the growth level between the i-th and j-th wind turbines at time t.

[0106] (3) Fluctuation level: The fluctuation level describes the level of fluctuation in wind speed between wind turbines at this point in time.

[0107]

[0108] In the formula, and S i S j are the average and standard deviation of the wind speed of the i-th and j-th wind turbines within a certain period interval, respectively. This represents the fluctuation level between the i-th and j-th wind turbines at time t. The smaller the stability difference between two sequences, the stronger their similarity.

[0109] Next, the spatiotemporal correlation coefficient is calculated. This coefficient is a weighted combination of absolute level, growth level, and fluctuation level, and its expression is as follows:

[0110]

[0111] In the formula, α, β, and δ are the weighting coefficients for the absolute level, growth level, and fluctuation level, respectively. Let be the spatiotemporal correlation coefficient between the i-th and j-th wind turbines at time t.

[0112] Step two, generate the dynamic adjacency matrix. Using the spatiotemporal correlation coefficients established in Step one, generate the dynamic adjacency matrix, calculate the spatiotemporal correlation coefficients between all wind turbines at each time point, and construct the adjacency matrix. For example... Figure 1As shown, this method extends the static graph into a spatiotemporal feature graph that can cover global information, thereby capturing the spatial correlation between wind turbines and reflecting the system's changing patterns in the time dimension.

[0113] A threshold method is introduced: when the spatiotemporal correlation coefficient is greater than or equal to a threshold, the value is set to 1; otherwise, it is set to 0, representing whether there is an edge connection between wind turbines at that time point. This transforms the continuous-valued spatiotemporal correlation coefficient matrix into a binary adjacency matrix, thereby constructing a dynamic adjacency matrix sequence describing network topology changes.

[0114] Step 3, Graph Embedding Based on Laplacian Mapping. The degree matrix of each node is calculated using the adjacency matrix of the graph obtained in Step 2, and the Laplacian matrix is ​​constructed accordingly. This is achieved by minimizing the optimization objective function, such as... Figure 2 As shown, mapping the high-dimensional graph structure to a low-dimensional space yields the embedding vector:

[0115]

[0116] In the formula, W ij The elements of the weight matrix of the graph, y i and y j It is the representation of node i and node j in low-dimensional space.

[0117] Step four involves extracting features from the embedding vectors using a clustering algorithm. Each embedding vector obtained in step three reflects the dynamic characteristics of changes in wind turbine adjacency relationships within the corresponding time period. This invention employs the K-Means algorithm to cluster graph embedding vectors, grouping similar embedding vectors into a single class to extract key structural features for further analysis. Specifically, the K-Means algorithm first randomly initializes cluster centers and then iteratively updates their positions. In each iteration, the algorithm assigns each data point to its nearest cluster center and recalculates the center position of each cluster. The formula for updating the cluster centers is:

[0118]

[0119] In the formula, |C i | represents the number of sample points in the i-th cluster. This process is repeated until the update magnitude of the cluster centers is less than a preset threshold or the maximum number of iterations is reached. Finally, the cluster label of the graph corresponding to each embedding vector and the positions of all cluster centers are output.

[0120] Step 5: Construct a global information graph. In Step 4, we obtained cluster labels and cluster centers, such as... Figure 4 As shown, the graph for each category is replaced with its corresponding cluster center graph, thereby achieving a unified representation and construction of global information.

[0121] Step six involves denoising the global information graph using an adaptive temporal label smoothing method. Due to the complex spatiotemporal correlation characteristics, the global information graph obtained in step five inevitably contains noise and random fluctuations, which affect the prediction performance of the prediction model. This invention proposes an Adaptive Temporal Label Smoothing (ATLS) method to optimize the labels of the global information graph. By introducing a sliding window strategy, it fully utilizes the contextual information of nearby time steps in the time series to dynamically correct the initial clustering results, thereby effectively reducing label dispersion and jump phenomena caused by noise.

[0122] Specifically, for each time step t, first check the current label. t The label smoothing process is triggered if there is inconsistency between the current label and its neighboring labels. Under this condition, the label smoothing decision function Smooth(t) is defined as follows:

[0123]

[0124] In the formula, label t Label representing time step t t-1 and label t+1 These represent the labels for the previous and next time steps, respectively. When the smoothing condition is met, it means that the current label may contain noise or inconsistencies, so it will be corrected using a sliding window.

[0125] The smoothing process relies on a sliding window of size ω. The window's range is defined by the labels of ω time steps before and after the current time step t. The specific range of the window is given by equation (7):

[0126]

[0127] In the formula, Represents a unique set of labels within the window. It is a tag Number of times it appears in the window.

[0128] If the smoothing condition Smooth(t) is satisfied, meaning the current label is inconsistent with its neighboring labels, then the current label is updated to the most common label within the window. The label update rule can be expressed as:

[0129]

[0130] In the formula, The updated labels. After the above smoothing process, the labels at all time steps are updated to the smoothed label sequence `smoothed_labels`, which can be represented as:

[0131]

[0132] Based on the smoothed sequence label information of all time steps, the global information graph is reconstructed by replacing the adjacency matrix of each time step with the adjacency matrix of the cluster center of the corresponding cluster label, thus obtaining the denoised global information graph.

[0133] where c t =smoothed_labels t ,

[0134] In the formula, Let represent the denoised adjacency matrix at time step t. c represents the cluster center t The corresponding standard adjacency matrix, c t =smoothed_labels t This represents the cluster label after smoothing at time step t.

[0135] Step 7: Establish a graph attention network model incorporating a continuous learning strategy. The state of graph information in the global information graph gradually changes over time, reflecting the environmental state at different moments and recording changes in wind turbine operating status. This invention uses a graph attention network (GAT) to fit the denoised global information graph to the ever-changing graph information input. Simultaneously, to effectively capture the temporal dependencies in wind power prediction tasks, this invention integrates a long short-term memory (LSTM) network on top of GAT to enhance the model's ability to model temporal information. Furthermore, a continuous learning strategy is introduced to record the changing information depicted by the global information graph, preventing the forgetting of past state information, especially environmental changes and wind turbine performance fluctuations that occur during long-term operation.

[0136] A graph attention network is used to fit the unpredictable graph information input of the global information graph. Specifically, this involves obtaining the denoised adjacency matrix sequence. Then, it is compared with the corresponding node feature sequence X. t Combined, it is used as input to the Graph Attention Network (GAT).

[0137] At each time step t, the Graph Attention Network (GAT) first bases its input features X on... t ∈R N×F and the corresponding adjacency matrix Calculate the attention weight for each node and aggregate information from neighboring nodes. The core operations of the graph attention layer are as follows:

[0138]

[0139] In the formula, Let W represent the input features of node i at time step t, W be the learnable linear transformation matrix, α be the attention weight vector, and N be the input features of node i at time step t. i Let H be the set of neighbors of node i, and σ(·) be a non-linear activation function. After processing by the graph attention module, the resulting spatial representation is H. t ∈R N×F′ .

[0140] To further model the dynamic evolution of node representations in the sequence, a graph is embedded into the sequence H. t The input is fed into the LSTM network. The LSTM unit models the time dependency through a gating mechanism; the computation process is as follows:

[0141] i t =σ(W i H t +U i h t-1 +b i ),

[0142] f t =σ(W f H t +U f h t-1 +b f ),

[0143] o t =σ(W o H t +U o h t-1 +b o ),

[0144] c t =f t ⊙c t-1 +i t ⊙tanh(W c H t +U c h t-1 +b c ),

[0145] h t =o t ⊙tanh(c t )

[0146] Among them, i tf t o t These are the input gate, forget gate, and output gate, respectively. t For the unit state, h t This represents the current hidden state, where ⊙ denotes element-wise multiplication. All weight matrices W * U * With bias term b * These are the parameters that the model needs to learn.

[0147] The GAT-LSTM model ultimately yields a series of hidden states h with spatiotemporal awareness. t The predicted value is then output through a fully connected layer.

[0148] like Figure 5 The steps of the continuous learning strategy based on sliding window and experience replay mechanism are as follows:

[0149] (1) Set a sliding window. Each time the model receives a new batch of data, the data window will slide to ensure that only the latest fixed-size data window is used for training.

[0150] (2) Perform experience replay: When the window slides and generates new data, the current window data is added to the experience replay buffer. The buffer stores a limited number of past windows. During each training session, in addition to the data of the current window, the model also samples past data from the buffer for training.

[0151] (3) Reset the loss calculation. The total loss of the model includes not only the loss of the current window, but also the loss of historical data, so as to ensure that the model can effectively learn the current task while maintaining the memory of past tasks.

[0152] Step 8, Model Training and Testing. Input the test set data into the model to train it. After training, input historical wind power cluster power data into the model to obtain predicted power data for future times. The specific method for model training is as follows:

[0153] The test set data is input into the model to train it. At time step t, the model receives a new batch of training data D. t ={(x1,y1),(x2,y2),...,(x n ,y n )}, where x i The input data is y. i These are the corresponding labels. The training loss of the model is expressed as:

[0154]

[0155] In the formula, f(x) i;θ) is the model's prediction, L(f(x) i ;θ),y i ) is the loss function, and θ is the model parameter.

[0156] First, a sliding window is set up. Each time the model receives a new batch of data, the data window is slid to ensure that only the latest fixed-size data window is used for training.

[0157] Then, experience replay is performed. As the window slides and generates new data, the current window data is added to the experience replay buffer, which stores a limited number of past windows. During each training iteration, in addition to the current window data, the model also samples past data from the buffer for training. Combined with the experience replay mechanism, loss calculation depends not only on the current window data but also on data from the experience replay buffer. The experience replay buffer stores data from multiple past windows, and during replay, a portion of data is randomly sampled from these historical experiences for training. The specific training objective is:

[0158]

[0159] In the formula, k is the amount of data sampled from the experience playback buffer, (x i ,y i The samples are historical data, taken from previous training windows. In this way, the model optimizes the loss for both the current task and past tasks simultaneously.

[0160] After training, historical wind power cluster power data is input into the model to obtain predicted power data for future times.

[0161] Example 2

[0162] The present invention compares the proposed method with widely used benchmark methods.

[0163] To verify the effectiveness of the proposed method in terms of generalization ability and professional adaptability, this invention was tested on two time scales: short-term and medium-term prediction. Medium-term prediction used the DSWE dataset, which is derived from the Wind Spatio-Temporal Dataset2 appended to the textbook *Data Science for Wind Energy*. This dataset was collected from a wind farm located in flat inland terrain, covering hourly spatiotemporal measurements of 200 randomly selected wind turbines from 2010 to 2011. The first 16 numbered wind turbines were selected as the research objects during the test. Short-term prediction used the SDWPF dataset, provided by China Longyuan Power Group Co., Ltd., with a recording frequency of once every 10 minutes. The first 16 numbered wind turbines were also selected for modeling and evaluation. Ten comparative algorithms were used: GCN, GAT, GCN-LSTM, GAT-LSTM, STGCN, ASTGCN, MSTGCN, GA-GCN, GA-GAT, GA-GCN-LSTM, and the proposed algorithm GADSG-CL.

[0164] Tables 1 and 2 compare the prediction performance of eleven algorithms at medium- and short-term time scales, corresponding to the DSWE and SDWPF datasets, respectively. The experimental setup includes three prediction tasks: predicting the power of a corresponding number of future time steps based on the past 12, 24, and 36 time steps. Evaluation metrics include mean absolute error (MAE), mean squared error (MSE), root mean square error (RMSE), and coefficient of determination (R²). 2 Among them, MAE measures the overall error magnitude, while MSE and RMSE are more sensitive to large errors, and R... 2 Indicates the goodness of fit of the model. The smaller the MAE, MSE, and RMSE, the better the R-value. 2 The closer the value is to 1, the better the model performance.

[0165] Table 1. Comprehensive comparison of experimental results on the DSWE dataset (interim)

[0166]

[0167]

[0168] Table 2. Comprehensive comparison of experimental results on the SDWPF dataset (short-term).

[0169]

[0170] As shown in Tables 1 and 2, from the evaluation of evaluation indicators, GADSG-CL outperforms all comparative methods in terms of performance across different evaluation indicators at both the medium and short-term time scales. This invention provides a new paradigm for wind power cluster power prediction that balances dynamic correlation, temporal stability, and noise resistance, and has significant practical implications for improving the grid-connected dispatch efficiency of renewable energy and the stability of the power system.

[0171] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0172] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A wind power cluster power prediction method based on global information graph fusion and continuous learning, characterized in that, The method includes the following steps: Step 1: Collect wind speeds at each wind turbine in the wind farm cluster, and propose three levels: absolute level, growth level, and fluctuation level, to describe the wind speed relationship between wind turbines. Then, calculate the spatiotemporal correlation coefficient by weighting the three levels. Step 2: Based on the spatiotemporal correlation coefficient, calculate the spatiotemporal correlation coefficient between all wind turbines at each time point and construct the adjacency matrix; Step 3: Calculate the degree matrix of the nodes based on the adjacency matrix, and construct the Laplacian matrix accordingly; map the high-dimensional graph structure of the Laplacian matrix to the low-dimensional space by minimizing the optimization objective function to obtain the embedding vector; Step 4: Cluster the embedding vectors and extract their features, including cluster labels and cluster centers; Step 5: Based on the obtained cluster labels and cluster centers, replace the graph of each category with its corresponding cluster center graph to obtain the global information graph; Step 6: Adopt the adaptive time series label smoothing method to optimize the clustering labels of the global information map. By introducing a sliding window strategy, the context information of the adjacent time in the time series is used to dynamically correct the initial clustering results, reduce the label discreteness and jump phenomenon caused by noise, and obtain the denoised global information map. Step 7: Use a graph attention network to fit the graph information input of the denoised global information graph, fuse a long short-term memory network and introduce a continuous learning strategy to establish a graph attention network model that incorporates the continuous learning strategy. Step 8: Input the test set data into the graph attention network model that incorporates the continuous learning strategy to train the model. After training, input the historical wind power cluster power data into the model to obtain the predicted power data for future times. The specific method for model training in step eight is as follows: The test set data is input into the model to train it, at time step... t The model receives a new batch of training data. ,in It is the input data. These are the corresponding labels; the model's training loss is expressed as: In the formula, It is the model's prediction. It is a loss function. These are model parameters; First, a sliding window is set up. Each time the model receives a new batch of data, the data window is slid to ensure that only the latest fixed-size data window is used for training. Then, experience replay is performed. As the window slides and generates new data, the current window data is added to the experience replay buffer, which stores a limited number of past windows. During each training session, in addition to the current window data, the model also samples past data from the buffer for training. Combined with the experience replay mechanism, loss calculation depends not only on the current window data but also on the data from the experience replay buffer. The experience replay buffer stores data from multiple past windows, and during replay, a portion of the data is randomly sampled from these historical experiences for training. The specific training objective is: In the formula, It is the amount of data sampled from the experience playback buffer. These are samples of historical data, taken from previous training windows; the model simultaneously optimizes the loss for both the current task and past tasks. After training, historical wind power cluster power data is input into the model to obtain predicted power data for future times.

2. The wind power cluster power prediction method based on global information graph fusion and continuous learning according to claim 1, characterized in that, The absolute level, growth level, and fluctuation level in step one are specifically as follows: Absolute grade: Used to describe the direct relationship between wind turbines at a given time point, reflecting the static relationship between the turbines at the current moment. Its formula is: In the formula, and Let be the wind speeds of the i-th and j-th wind turbines at time t, respectively. This represents the absolute ranking between the i-th and j-th wind turbines at time t; Growth Level: Used to describe the developmental relationship between wind turbines at a given time point, reflecting the dynamic relationship between wind turbines at the current moment. Its formula is: In the formula, and Let $\frac{i}{j}$ represent the wind speed changes of the $i$-th and $j$-th wind turbines from time $t-1$ to time $t$. and Let be the wind speeds of the i-th and j-th wind turbines at time t-1, respectively. This represents the growth level between the i-th and j-th wind turbines at time t; Fluctuation level: Used to describe the fluctuation level of wind speed change between wind turbines at a given time point, and its formula is: In the formula, , and , These are the average and standard deviation of the wind speed of the i-th and j-th wind turbines within a certain period interval, respectively. This represents the fluctuation level between the i-th and j-th wind turbines at time t.

3. The wind power cluster power prediction method based on global information graph fusion and continuous learning according to claim 2, characterized in that, The specific method for calculating the spatiotemporal correlation coefficient in step one is as follows: The spatiotemporal correlation coefficient is composed of a weighted combination of absolute level, growth level, and fluctuation level, and its expression is as follows: In the formula, , , These are the weighted coefficients for absolute level, growth level, and fluctuation level, respectively. Let be the spatiotemporal correlation coefficient between the i-th and j-th wind turbines at time t.

4. The wind power cluster power prediction method based on global information graph fusion and continuous learning according to claim 1, characterized in that, The method for constructing the adjacency matrix in step two is as follows: Using the spatiotemporal correlation coefficient established in step one, a dynamic adjacency matrix is ​​generated. The spatiotemporal correlation coefficient between all wind turbines at each time node is calculated, and an adjacency matrix is ​​constructed. A threshold method is introduced: when the spatiotemporal correlation coefficient is greater than or equal to the threshold, the value is set to 1; otherwise, it is 0, which indicates whether there is an edge connection between wind turbines at that time node. The continuous spatiotemporal correlation coefficient matrix is ​​converted into a binary adjacency matrix, thereby constructing a dynamic adjacency matrix sequence that describes the changes in network topology.

5. The wind power cluster power prediction method based on global information graph fusion and continuous learning according to claim 1, characterized in that, The specific method for calculating the embedding vector in step three is as follows: The degree matrix of the nodes is calculated using the adjacency matrix of the graph obtained in step two, and the Laplacian matrix is ​​constructed from this matrix. By minimizing the objective function, the high-dimensional graph structure is mapped to a low-dimensional space, resulting in the embedding vector. In the formula, The elements of the weight matrix of the graph are represented. and It is a node and nodes Representation in low-dimensional space.

6. The wind power cluster power prediction method based on global information graph fusion and continuous learning according to claim 1, characterized in that, The clustering method in step four is as follows: The K-Means algorithm is used to cluster graph embedding vectors. First, cluster centers are randomly initialized. Then, the positions of the cluster centers are updated iteratively. In each iteration, the algorithm assigns each data point to its nearest cluster center and recalculates the center position of each cluster. The formula for updating the cluster centers is: In the formula, For the first The number of sample points in each cluster sample is counted. This process is repeated until the update magnitude of the cluster center is less than a preset threshold or the maximum number of iterations is reached. Finally, the cluster label of the graph corresponding to each embedding vector and the positions of all cluster centers are output.

7. The wind power cluster power prediction method based on global information graph fusion and continuous learning according to claim 1, characterized in that, The adaptive time-series label smoothing method used in step six is ​​specifically as follows: For each time step First, check the current tag. Whether there is inconsistency between the current label and its neighboring labels; if the current label is different from the labels before and after it, the label smoothing process is triggered. The label smoothing determination function is as follows. The definition is as follows: In the formula, Indicates time step The tag, and These represent the labels for the previous and next time steps, respectively. When the smoothing condition is met, it means that the current label may contain noise or inconsistencies, which will be corrected through a sliding window. The smoothing process depends on a size of A sliding window whose range is the current time step. front and back The labels for each time step, and the specific scope of the window are as follows: In the formula, Represents a unique set of labels within the window. It is a tag Number of times it appears in the window; If smoothing condition If the current label is inconsistent with its neighboring labels, then the current label is updated to the most common label within the window. The label update rule is expressed as follows: In the formula, After a smoothing process, the labels at all time steps are updated to the smoothed label sequence to represent the new labels. The label sequence is represented as: Based on the smoothed sequence label information of all time steps, the global information graph is reconstructed by replacing the adjacency matrix of each time step with the adjacency matrix of the cluster centers corresponding to the cluster labels, resulting in the denoised global information graph: In the formula, Indicates time step The corresponding denoised adjacency matrix, Representing cluster centers The corresponding standard adjacency matrix, Indicates time step Smoothed cluster labels.

8. The wind power cluster power prediction method based on global information graph fusion and continuous learning according to claim 7, characterized in that, The graph attention network model that integrates the continuous learning strategy established in step seven is specifically as follows: A graph attention network is used to fit the unpredictable graph information input of the global information graph. Specifically, this involves obtaining the denoised adjacency matrix sequence. Then, it is compared with the corresponding node feature sequence. Combined, it is used as input to the Graph Attention Network (GAT); At each time step The Graph Attention Network (GAT) first bases its attention on input features. and the corresponding adjacency matrix Calculate the attention weight for each node and aggregate information from neighboring nodes; the operations of the graph attention layer are as follows: In the formula, Indicates time step Time node j Input features, Indicates time step Time node i Input features, It is a learnable linear transformation matrix. For attention weight vectors, For nodes The neighborhood group, The activation function is nonlinear; after processing by a graph attention network, the resulting spatial representation is: ; To further model the dynamic evolution of node representations in the sequence, a graph is embedded into the sequence. The input is fed into the Long Short-Term Memory (LSTM) network; the LSM network models the time dependency through a gating mechanism, and its calculation process is as follows: in, , , These are the input gate, the forget gate, and the output gate. For the unit state, Currently in a hidden state. Represents element-wise multiplication, all weight matrices , With bias term These are the parameters that the model needs to learn; The graph attention network model that integrates continuous learning strategies, namely the GAT-LSTM model, ultimately obtains a series of hidden states with spatiotemporal awareness. The predicted value is then output through a fully connected layer.

9. A wind power cluster power prediction method system based on global information graph fusion and continuous learning, characterized in that, include: Memory, used to store executable computer programs; A processor, when executing an executable computer program stored in memory, implements the wind power cluster power prediction method based on global information graph fusion continuous learning as described in any one of claims 1 to 8.

Images