Load data spatio-temporal correlation analysis method based on spatio-temporal graph neural network
By analyzing the spatiotemporal correlation of wind and solar load data using the FastDTW algorithm and spatiotemporal graph neural network, the problem of insufficient accuracy of traditional prediction methods is solved, and more efficient feature extraction and prediction of load data are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 山西省能源互联网研究院
- Filing Date
- 2022-12-28
- Publication Date
- 2026-05-15
AI Technical Summary
Traditional wind and solar power prediction methods have limited room for accuracy improvement due to the single factor considered, making it difficult to effectively improve the prediction effect of solar power.
The FastDTW algorithm is used to analyze the spatiotemporal correlation of wind and solar load data. A correlation matrix is constructed as the model input, and multiple independent and parallel spatiotemporal graph neural network modules are used to extract the spatiotemporal features of the load data. Feature transformation and extraction are performed through a spatiotemporal graph convolutional network.
It improves the prediction accuracy of wind and solar load data, provides high-quality basic data support, and provides an important data analysis tool for wind and solar power consumption and stable operation of the power system.
Smart Images
Figure CN116167276B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to dynamic time warping and graph neural network technology. Specifically, it uses the FastDTW algorithm to analyze the spatiotemporal correlation of wind and solar load data, constructs a correlation matrix as model input, and uses multiple independent parallel spatiotemporal graph neural network modules to save time, capture more complex correlations, and finally extract the spatiotemporal features of historical load sequences. Background Technology
[0002] Large-scale wind and solar power grid integration poses a severe challenge to the safe and stable operation of the power system. Wind and solar power forecasting is an important way to cope with the uncertainty of solar power and promote the consumption of wind and solar power. Traditional forecasting methods have limited room for improvement in accuracy due to the single factor considered. Therefore, it is urgent to study new theories and methods to effectively improve the solar power forecasting effect by fully exploring and utilizing the spatiotemporal correlation between wind and solar fields over a wide area.
[0003] Deep learning algorithms are characterized by their strong generalization ability, possessing the capacity to extract abstract features that characterize the system's operational state from a large feature set. In recent years, graph neural networks have been widely applied in natural language processing, social networks, computer vision, and life sciences due to their ability to construct deep learning models for graph network structures. The topology of a power system is itself a graph network, and most of the collected wind and solar power data are time-series, with inherent connections between data points at different times. Spatiotemporal graph neural networks consider not only the relationships between nodes but also the spontaneous changes of each node over time. In recent years, deep learning and graph neural networks have played an increasingly important role in energy research, and utilizing computer-aided analysis of the spatiotemporal correlations of load data is of significant importance and value to energy research. Summary of the Invention
[0004] To address the technical problem that traditional wind and solar power prediction methods have limited room for accuracy improvement due to considering only one factor, this invention provides a method for analyzing the spatiotemporal correlation of load data based on a spatiotemporal graph neural network.
[0005] This invention is achieved using the following technical solution: a method for spatiotemporal correlation analysis of load data based on spatiotemporal graph neural networks, comprising the following steps: (1) preprocessing
[0006] The raw wind and solar load data obtained by the acquisition terminal are processed to complete the data: the average value of the raw load data at the same time on two adjacent days with missing values is taken; then the completed load data is processed by normalization method, and the normalization formula is shown in (1).
[0007]
[0008] Where x is the original data after completion, x min It is the minimum value of the original data after completion, x max x represents the maximum value of the original data after completion. norm The normalized load time series is [0,1]; the original data uses the same period.
[0009] (2) FastDTW Algorithm Analysis
[0010] The multiple load time series obtained in step (1) after preprocessing need to be analyzed using the FastDTW algorithm to obtain the spatiotemporal correlation matrix. FastDTW calculates the similarity between any two load time series by extending and shortening the load time series.
[0011] FastDTW uses the normalized path distance to measure the similarity between two load time series. Let the two load time series to be similarized be X1 and Y1, with lengths |X1| and |Y1| respectively. The normalized path is in the form W = w1, w2, ..., w K Where Max(|X1|,|Y1|)<=K<=|X1|+|Y1|;
[0012] w k The form is (i,j), where i represents the i-th coordinate in X1 and j represents the j-th coordinate in Y1. The rectified path W must start from w1 = (1,1) and end at w1. K = (|X1|,|Y1|) to ensure that every coordinate in X1 and Y1 appears in W; in addition, i and j in w(i,j) in W must be monotonically increasing. Monotonically increasing means:
[0013] w k =(i,j), k+1 =(i′,j′)(2)
[0014] Finally, the reorganized path obtained by equation (3) is the shortest reorganized path. In equation (3), Dist(w) ki ,w kj ) represents any classic distance calculation method, w ki It refers to the i-th data point of X, w kj It refers to the j-th data point of Y (Dist(w) ki ,w kj (This has the same meaning as Dist(i,j);)
[0015] D(i,j=Dist(i,j)+min[D(i-1,j,D(i,j-1),D(i-1,j-1)](3
[0016] The final calculated path distance is D(|X|,|Y|), where D(i,j) represents the path distance between two load time series of lengths i and j.
[0017] The correlation coefficient between the two load time series can be calculated using the normalized path distance obtained above.
[0018]
[0019] (3) Spatiotemporal graph neural network
[0020] ①. Calculate the correlation coefficients of any two load time series using the method in step (2), and construct the correlation matrix as the model input;
[0021] ②. The spatiotemporal graph neural network consists of two spatiotemporal convolutional blocks, each including two temporal convolutional blocks and one spatial convolutional block, finally connected to an output layer. The temporal convolutional blocks transform the feature dimensions, with 64 filters on both sides and 16 filters in the middle. Graph convolution is used to scale the channels down and up, achieving size compression and feature transformation. Each temporal convolutional block consists of a one-dimensional convolution and a gated linear unit, with a temporal convolutional kernel width of K. t Each time a temporal convolution block is passed, the length of the load time series decreases by K. t -1, by using two stacked spatiotemporal convolutional blocks, the load time series length is reduced by 4×(K). t -1);
[0022] The output layer consists of a temporal convolutional layer and a fully connected layer. The output layer first merges the temporal dimension of the output data of the previous spatiotemporal convolutional blocks using temporal convolution, and then outputs the final data after another convolution.
[0023] This invention discloses the use of spatiotemporal graph neural networks to analyze the spatiotemporal correlation of load data. First, the FastDTW algorithm is used to analyze the spatiotemporal correlation of the acquired load data to construct a correlation matrix. Finally, the acquired dataset is used to train multiple independent and parallel spatiotemporal graph neural networks to obtain the spatiotemporal characteristics of the load data and provide high-quality basic data for prediction. Attached Figure Description
[0024] Figure 1 Weekly load curve.
[0025] Figure 2 Time series curve graph.
[0026] Figure 3 Schematic diagram of the shortest resetting path.
[0027] Figure 4 Spatiotemporal graph neural network model architecture diagram.
[0028] Figure 5 Spatiotemporal convolution block architecture diagram.
[0029] Figure 6 Temporal convolution block architecture diagram.
[0030] Figure 7 Output layer architecture diagram.
[0031] Figure 8 A schematic diagram of the principle of this invention. Detailed Implementation
[0032] This invention preprocesses the obtained wind and solar load data, uses the FastDTW algorithm to achieve linear complexity in a time series composed of long load data, generates a correlation matrix, and uses the generated correlation matrix as input to multiple independent parallel spatiotemporal graph neural network models, thereby achieving the purpose of extracting the spatiotemporal features of the load time series.
[0033] 1. Pretreatment
[0034] Typically, the raw wind and solar load data obtained from the data acquisition terminal is incomplete and heavily polluted by noise, so data completion processing is necessary. This is typically achieved by averaging the load data from the same time point on two consecutive days adjacent to the missing value. Neural network models are sensitive to the scale of the input training data during training; excessively large values can negatively impact training performance. In such cases, normalization can be used to process the completed data. The normalization formula is shown below.
[0035]
[0036] Where x is the original data, x min It is the minimum value of the original data, x max For the maximum value, x norm The data is normalized to [0,1].
[0037] 2. Analysis of the FastDTW Algorithm
[0038] After obtaining the preprocessed load data time series in the previous step, a corresponding correlation matrix needs to be generated. This invention uses the FastDTW algorithm to analyze the spatiotemporal correlation of the load data time series to obtain the correlation matrix.
[0039] Conventional load data exhibits a certain periodicity on load curves. Based on the length of the actual load curve's periodicity, it can generally be divided into daily, weekly, and monthly periods. Currently, the weekly period is commonly used for correlation analysis of the load curve's periodicity. A weekly periodicity chart is shown below. Figure 1 As shown.
[0040] FastDTW calculates the similarity between any two weekly load time series by stretching and shortening the load time series: for example Figure 2 As shown, the two solid lines represent two weekly load time series, and the dashed lines between the time series represent similar points between the two time series. FastDTW uses the sum of the distances between all these similar points, called the Warp Path Distance, to measure the similarity between two time series.
[0041] Let the two time series whose similarity is to be calculated be X1 and Y1, with lengths |X1| and |Y1| respectively. The normalized path has the form W = w1, w2, ..., w K， Where Max(|X1|,|Y1|)<=K<=|X1|+|Y1|.
[0042] w k The form is (i,j), where i represents the coordinate of X1 and j represents the coordinate of Y1. The normalized path W must start from w1 = (1,1) and end at w1. K The expression ends with (|X1|,|Y1|) to ensure that every coordinate in X1 and Y1 appears in W. Furthermore, the i and j of w(i,j) in W must be monotonically increasing. Monotonically increasing means:
[0043] w k =(i,j), k+1 =(i′,j′)(2)
[0044] The final rectified path is the one with the shortest distance, where Dist(w ki ,w kj ) represents any classic distance calculation method, such as Euclidean distance. ki It refers to the i-th data point of X, w kj It refers to the j-th data point of Y.
[0045] D(i,j=Dist(i,j)+min[D(i-1,j,D(i,j-1),D(i-1,j-1)](3
[0046] The final calculated reduced path distance is D(|X|,|Y|), where D(i,j) represents the reduced path distance between two time series of lengths i and j, such as... Figure 3 As shown; Figure 3 The diagram shows the cost matrix D, which is also a schematic diagram of the minimum regularized path distance.
[0047] The correlation coefficient between the two time series can be calculated using equation (4) based on the normalized path distance obtained above.
[0048]
[0049] This invention uses the FastDTW algorithm because:
[0050] 1) Coarse-grained sizing. The FastDTW algorithm first abstracts the original time series data. The data abstraction is iterated multiple times from 1 / 1->1 / 2->1 / 4->1 / 16, which can obtain the average value of multiple corresponding fine-grained data points.
[0051] 2) Projection. The FastDTW algorithm can run the DTW algorithm on time series at a coarser granularity.
[0052] 3) Fine-grained scaling. The FastDTW algorithm can refine the squares traversed by the normalized path obtained at a coarse-grained level to a finer-grained time series. In addition to fine-grained scaling, it can also expand the granularity by K levels within the fine-grained space.
[0053] 3. Spatiotemporal Graph Neural Network
[0054] Graph Convolutional Networks (GCNs) update the feature state of the current node based on information from neighboring nodes and the topological graph structure. There are two main types of GCNs for processing graph-structured data: spatial domain-based graph convolution and spectral domain-based graph convolution. This invention employs a spatial domain-based Spatiotemporal Graph Convolutional Network (STGCN), which essentially defines graph convolution operators based on spatial relationships, continuously aggregating feature information from neighboring nodes to update the feature state of the target node, similar to traditional CNNs.
[0055] The network architecture used in this invention is as follows:
[0056] 1. Calculate the correlation coefficients of any two load data time series obtained in the previous step, and construct a correlation matrix as the model input.
[0057] 2. A spatiotemporal graphical neural network consists of two spatiotemporal convolutional blocks (ST-Conv Blocks), such as... Figure 4As shown, each ST-Conv Block consists of two temporal convolutional blocks and one spatial convolutional block, finally connected to an output layer. The temporal convolutional blocks transform the feature dimensions; there are 64 filters on either side and 16 filters in the middle (e.g., ...). Figure 5 As shown), graph convolution is used to scale channels down and up, achieving size compression and feature transformation. The temporal-gated-Conv block consists of a 1-D Conv and a GLU, where GLU stands for Gated Linear Unit (GLU). Figure 6 The temporal kernel width of each temporal convolutional block is K. t Each time a temporal convolution block is passed, the length of the time series decreases by K. t -1, by using two stacked spatiotemporal convolutional blocks, the time series length is reduced by 4×(K). t -1);
[0058] 3. The output layer consists of a temporal convolutional layer and a fully connected layer (FC layer) (e.g.) Figure 7 The output layer first uses a convolutional layer along the time dimension to merge the time dimensions of the previous output data. After merging, it undergoes another convolution to output the final data. A residual network is also used during feature extraction to improve model performance to some extent.
[0059] This invention uses the Graph Convolutional Network STGCN because it has the following advantages:
[0060] 1) Spatiotemporal graph convolutional networks can simultaneously learn node feature information and structural information end-to-end, making them the best choice for graph data learning tasks.
[0061] 2) Spatiotemporal graph convolutional networks have wide applicability and are suitable for nodes and graphs with arbitrary topologies.
[0062] 3) Spatiotemporal graph convolutional networks can effectively extract spatial features from the topological graph for machine learning.
Claims
1. A method for spatiotemporal correlation analysis of load data based on spatiotemporal graph neural networks, characterized in that, The steps include: (1) Preprocessing The raw wind and solar load data obtained by the acquisition terminal are processed to complete the data: the average value of the raw load data at the same time on two adjacent days with missing values is taken; then the completed load data is processed by normalization method, and the normalization formula is shown in (1). Where x is the original data after completion, x min It is the minimum value of the original data after completion, x max x represents the maximum value of the original data after completion. norm The normalized load time series is [0,1]; the original data uses the same period. (2) FastDTW Algorithm Analysis The multiple load time series obtained in step (1) after preprocessing need to be analyzed using the FastDTW algorithm to obtain the spatiotemporal correlation matrix. FastDTW calculates the similarity between any two load time series by extending and shortening the load time series. FastDTW uses the normalized path distance to measure the similarity between two load time series. Let the two load time series to be similarized be X1 and Y1, with lengths |X1| and |Y1| respectively. The normalized path has the form W = w1, w2, ..., w K， Where Max(|X1|,|Y1|) <= K <= |X1|+|Y1|; w k The form is (i,j), where i represents the i-th coordinate in X1 and j represents the j-th coordinate in Y1. The rectified path W must start from w1 = (1,1) and end at w1. K = (|X1|,|Y1|) to ensure that every coordinate in X1 and Y1 appears in W; in addition, i and j in w(i,j) in W must be monotonically increasing. Monotonically increasing means: w k =(i,j),w k+1 =(i′,j′) (2) Finally, the reorganized path obtained by equation (3) is the shortest reorganized path. In equation (3), Dist(w) ki ,w kj ) represents any classic distance calculation method, w ki It refers to the i-th data point of X, w kj It refers to the j-th data point of Y; D(i,j)=Dist(i,j)+min[D(i-1,j,D(i,j-1),D(i-1,j-1)] (3 The final calculated path distance is D(|X|,|Y|), where D(i,j) represents the path distance between two load time series of lengths i and j. The correlation coefficient between the two load time series can be calculated using the normalized path distance obtained above. (3) Spatiotemporal graph neural network ①. Calculate the correlation coefficients of any two load time series using the method in step (2), and construct the correlation matrix as the model input; ②. The spatiotemporal graph neural network consists of two spatiotemporal convolutional blocks, each including two temporal convolutional blocks and one spatial convolutional block, finally connected to an output layer. The temporal convolutional blocks transform the feature dimensions, with 64 filters on both sides and 16 filters in the middle. Graph convolution is used to scale the channels down and up, achieving size compression and feature transformation. Each temporal convolutional block consists of a one-dimensional convolution and a gated linear unit, with a temporal convolutional kernel width of K. t Each time a temporal convolution block is passed, the length of the load time series decreases by K. t -1, by using two stacked spatiotemporal convolutional blocks, the load time series length is reduced by 4×(K). t -1); The output layer consists of a temporal convolutional layer and a fully connected layer. The output layer first merges the temporal dimension of the output data of the previous spatiotemporal convolutional blocks using temporal convolution, and then outputs the final data after another convolution.
2. The method for spatiotemporal correlation analysis of load data based on spatiotemporal graph neural networks as described in claim 1, characterized in that, The raw data of wind and solar loads obtained by the acquisition terminal can be divided into daily, weekly, and monthly cycles.
3. The method for spatiotemporal correlation analysis of load data based on spatiotemporal graph neural networks as described in claim 2, characterized in that, Correlation analysis was performed on the periodicity of wind and solar power load time series using the weekly cycle.
4. The method for spatiotemporal correlation analysis of load data based on spatiotemporal graph neural networks as described in any one of claims 1-3, characterized in that, The distance calculation method refers to the Euclidean distance calculation method.
5. The method for spatiotemporal correlation analysis of load data based on spatiotemporal graph neural networks as described in any one of claims 1-3, characterized in that, Step (3) ②. The output layer uses a residual network during the feature extraction process, which can improve the model performance.