Method and system for predicting short-term load probability of transformer area
By combining K-means clustering and LASSO feature selection with graph attention network and temporal convolutional network, the problem of multi-source data fusion and spatiotemporal correlation in transformer area load forecasting is solved, achieving high-precision load forecasting results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-09
- Publication Date
- 2026-03-31
AI Technical Summary
Existing technologies struggle to effectively integrate multi-source heterogeneous data such as population flow and meteorological conditions, neglecting the spatial similarity between transformer substations and the spatiotemporal dynamic correlation of load patterns. This results in limited spatiotemporal modeling capabilities of the GCN-RNN model, making it unable to accurately characterize the spatiotemporal dependence of transformer substation loads and lacking sufficient feature representation capabilities.
By extracting multidimensional features, the K-means clustering algorithm is used to identify transformer substations with similar load characteristics. LASSO is combined for feature selection, and a hybrid model of graph attention network (GAT) and temporal convolutional network (TCN) is constructed to combine spatial and temporal dependencies. Quantile regression is then used to quantify the uncertainty of prediction.
It significantly improves the accuracy and reliability of short-term load forecasting for transformer substations, and enhances the model's ability to express spatiotemporal characteristics and capture long-term trends.
Smart Images
Figure CN121769834A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of transformer area load forecasting technology, and in particular to a method and system for forecasting short-term load probability in transformer areas. Background Technology
[0002] In distribution network load forecasting, existing technologies struggle to effectively integrate multi-source heterogeneous data such as population flow and meteorological conditions, neglecting the spatial similarity between different distribution areas and the spatiotemporal dynamic correlation of load patterns. For example, distribution area load forecasting based on GCN-RNN primarily involves collecting historical load data and related external feature data from each distribution area, imputing missing values, and normalizing or standardizing features with different dimensions to ensure consistency in the input feature dimensions. Then, based on the historical load correlations between distribution areas, a graph structure is constructed. ,in V For a set of nodes, E Let be the set of edges. A The adjacency matrix is used to characterize the spatial dependence between distribution transformers. Spatial features extracted by the GCN are input into a recurrent neural network (RNN) in chronological order to capture the dynamic changes in transformer load over time. The RNN output is then passed through a fully connected layer to obtain future load predictions. This scheme directly uses individual transformers as graph nodes without clustering them, ignoring potential clustering relationships between transformers with similar electricity consumption characteristics. This results in a sparse graph structure and limited spatial feature representation, making it difficult to characterize the regional characteristics and spatial correlations of distribution network loads at a global level. Using raw multi-source features as input lacks feature selection and dimensionality reduction processes, easily introducing noisy features that affect the model's convergence speed and generalization performance. The GCN-RNN model has limited spatiotemporal modeling capabilities. GCN uses uniform weights when aggregating neighbor node features, only capturing averaged information from a fixed topology, failing to distinguish the importance of different neighbors. RNNs suffer from gradient vanishing and difficulties in modeling long sequences.
[0003] Therefore, it is necessary to study a prediction method and system for short-term load probability of transformer substations, which can accurately characterize the spatiotemporal dependence of load in transformer substations, extract key features, and achieve high-precision load prediction under the condition of multi-source heterogeneous data. Summary of the Invention
[0004] To address the problems in existing technologies, this invention provides a method and system for predicting short-term load probability in transformer substations. The specific technical solution is as follows: A method for predicting short-term load probability in a transformer substation includes the following steps: Step S1: Extract the load data of the transformer area including multi-dimensional features, including historical load data, historical population flow data, meteorological data, and date data, and calculate the peak-valley difference coefficient of the load. Step S2: Based on the extracted multidimensional features of the transformer area load data, obtain the correlation between the transformer area load and meteorological data and population flow data, and obtain the comprehensive spatial features based on meteorological factors and the comprehensive spatial features based on population flow. Step S3: Based on the comprehensive spatial characteristics based on meteorological factors, the comprehensive spatial characteristics based on population flow, and the peak-valley difference coefficient of the load, the K-means clustering algorithm is used to cluster the N transformer areas to obtain transformer area clusters with similar spatiotemporal characteristics. Step S4: Select multidimensional features based on LASSO to obtain the key features closely related to load changes after screening; Step S5: Graph attention network is used to model the spatial dependency relationship between transformer clusters, realize the adaptive aggregation of spatial features, and extract the updated features of transformer clusters. Step S6: Input the extracted and updated features of the transformer area into a temporal convolutional network. Through multi-layer dilated convolution, capture the multi-scale time dependence and long-term trend changes of the load sequence to obtain the load distribution of the transformer area at different confidence levels.
[0005] Preferably, in step S1, the peak-valley difference coefficient of the load is calculated based on the daily load curve. K pv The details are as follows: ; in, P max , P min These represent the peak and trough values of the daily load, respectively. This represents the daily average load.
[0006] Preferably, step S2, which involves obtaining the correlation between the substation load and meteorological and population flow data based on the extracted multidimensional features of the substation load data, specifically includes the following steps: (1) Let the time series of the load in the transformer area be as follows: P t The meteorological data series is W t， The meteorological data includes various meteorological factors, specifically temperature, humidity, wind speed, normal irradiance, and diffuse irradiance. The correlation coefficient between the substation load and the meteorological data is: ; ; Among them, cov( Pt , W t，r ) is a load time series P t With the rth meteorological factor W t，r The covariance of , where r=1,2,3,4,5, corresponding to temperature, humidity, wind speed, normal irradiance, and diffuse irradiance, respectively; , , respectively, represent the standard deviations of the load and the r-th meteorological factor sequence; Based on the temporal characteristics of the r-th meteorological factor, this method is used to quantify the response intensity of different substation loads to climate change. (2) Let the time series of population movement be: M t The correlation coefficient between transformer load and population flow data is calculated as follows: ; These are the standard deviations of the population mobility time series, and cov( P t , M t ) is a load time series P t Covariance with the time series of population movement; This is based on the temporal characteristics of population movement.
[0007] Preferably, obtaining the comprehensive spatial characteristics based on meteorological factors and the comprehensive spatial characteristics based on population flow in step S2 specifically includes the following steps: (1) Using a sliding window [ t -L, t The user load curve and meteorological factor sequence curve are segmented, and the Pearson correlation coefficient between the load curve and different meteorological factors is calculated within each sliding window to obtain the time... t Based on meteorological factors r Time correlation The time correlation vector of meteorological factors r is constructed by concatenating the data in chronological order. As shown in the following formula: ; The Pearson correlation coefficient between the time correlation vectors of different transformer areas is calculated as follows: ; Where corr() represents the Pearson correlation coefficient calculation function, Indicates the area i Taiwan j Based on meteorological factorsr Spatial correlation; and They are respectively the Taiwan area i Taiwan j The time correlation vector of meteorological factors; The mean of the spatial correlation of each transformer area with all other transformer areas based on meteorological factors is used as a comprehensive spatial correlation feature of that transformer area based on meteorological factor r. The specific calculation method is as follows: ; Indicates the area i Based on the comprehensive spatial correlation characteristics of meteorological factor r; (2) Using a sliding window [ t -L, t The user load curve and population flow time series curve are segmented, and the Pearson correlation coefficient between the load curve and the population flow time series is calculated within each sliding window to obtain the time... t Temporal correlation of population flow time series Concatenate them in chronological order to form a population mobility time correlation vector As shown in the following formula: ; The Pearson correlation coefficient between the time correlation vectors of different transformer areas is calculated as follows: ; in, Indicates the area i Taiwan j Based on the spatial correlation of population mobility; and They are respectively the Taiwan area i Taiwan j The population mobility time correlation vector; The mean of the spatial correlation of each transformer area with all other transformer areas based on population flow is used as a comprehensive spatial correlation feature of that transformer area based on population flow. The specific calculation method is as follows: ; Indicates the area i Comprehensive spatial correlation characteristics based on population mobility.
[0008] Preferably, step S3, which uses the K-means clustering algorithm to cluster N transformer substations to obtain substation clusters with similar spatiotemporal characteristics, specifically includes the following steps: (1) In NK initial cluster centers were randomly selected from each of the transformer substations. Each cluster center includes the peak-to-valley difference coefficient of the load in the substation area and the comprehensive spatial correlation characteristics based on the meteorological factor r. Comprehensive spatial correlation characteristics based on population flow ; (2) Calculate the Euclidean distance from each substation to each cluster center, and assign each substation to the substation cluster containing the cluster center with the smallest Euclidean distance; (3) Recalculate the cluster centers of each station area cluster based on the current cluster allocation; (4) Repeat (2)-(3) until the maximum number of iterations is reached or the cluster center no longer changes, and finally obtain K clusters of stations with similar spatiotemporal characteristics.
[0009] Preferably, step S5 utilizes a graph attention network to model the spatial dependencies between transformer clusters, achieving adaptive aggregation of spatial features and extracting the spatial features of the transformer areas. Specifically, this includes the following steps: (1) Constructing a graph structure ,in V It is a graph structure G A set of nodes, where each node represents a cluster of stations. E It is a graph structure G The edge connection set, A Representative graph structure G An adjacency matrix describes the connection relationships between graph nodes and is constructed based on the correlation between graph nodes. (2) After obtaining the graph structure G, the spatial dependencies between the clusters are extracted using a graph attention network, and the feature matrix filtered based on LASSO is then used. Reconstructed as a node-time feature tensor H , ;in, K The number of clusters in the station area. T The number of time samples for each station cluster. The feature dimensions of the feature matrix after LASSO filtering; (3) Each station cluster i At time step t The initial features are ;in This indicates that the background region clusters have been selected through LASSO feature selection. i At time step t eigenvectors; Taiwan Cluster i With each of its adjacent station clusters j Attention coefficient needs to be calculated This coefficient is used to represent the cluster of adjacent station areas. j Taiwan Clusteri The importance of attention coefficients is calculated using the features of two clusters of stations, as shown in the following formula: ; in This represents the concatenation of node features, where W is the weight matrix. a T For the parameter vector, use the LeakyReLU activation function; (4) The calculated attention coefficient Normalize to weights a ij The calculation formula is as follows: ; in, N i For the cluster of the Taiwan region i The set of adjacent station clusters is used to achieve feature aggregation between station clusters by weighted aggregation of the features of each adjacent station cluster. The aggregation formula is as follows: ; in, For activation function, The features of the station cluster are updated using a graph attention network.
[0010] Preferably, step S6 specifically includes the following steps: The extracted and updated features of the transformer area are used as the initial input to the temporal convolutional network. z i,t Introducing dilated convolution, by adding a dilation factor to the convolution operation, expands the receptive field for clustered platforms. i The expression for calculating the dilated convolution at time t is: ; in, Where d is the kernel size and d is the dilation factor. W k For convolution kernel weights, For dilated convolution in t Output at any moment Indicates the corresponding cluster of the upper layer. i At time step td*k Output; expansion coefficient d It increases exponentially by 2 with the number of network layers; In a multi-layer TCN, each layer contains residual connections to ensure training stability, as detailed below: ; F() represents the TCN layer operations, including dilated convolution and activation functions. Indicates the upper-level cluster of the platform area.i At time step t The output, Indicates the current layer cluster i At time step t The output; Finally, the output of the last layer of the TCN is mapped to the predicted load values at different quantiles through a fully connected layer: ; in, and These correspond to quantiles. The output layer weights and biases; This represents the corresponding quantile. The predicted load value; Next, quantile regression is used as the loss function to characterize the load distribution at different confidence levels, thereby obtaining interval prediction results; its loss function is the total loss of each quantile loss, defined as: ; in, y i , t The actual load value of cluster i in the transformer area at time t. Let i be the predicted load value for cluster i in transformer area at time t. The selected set of quantiles; The quantile loss function, The definition is as follows: .
[0011] A prediction system for short-term load probability in a transformer substation, employing the method described herein, includes: The data extraction module is used to extract load data of the transformer area including multi-dimensional features, such as historical load data, historical population flow data, meteorological data, and date data, and to calculate the peak-valley difference coefficient of the load. The spatial feature calculation module is used to obtain the correlation between the load of the substation area and meteorological data and population flow data based on the extracted multi-dimensional features of the substation area load data, and to obtain the comprehensive spatial features based on meteorological factors and the comprehensive spatial features based on population flow. The transformer area clustering module is used to cluster N transformer areas based on comprehensive spatial characteristics based on meteorological factors, comprehensive spatial characteristics based on population flow, and peak-valley difference coefficient of load, using the K-means clustering algorithm to obtain transformer area clusters with similar spatiotemporal characteristics. The feature selection module is used to select multidimensional features based on LASSO to obtain key features closely related to load changes after screening. The feature update module is used to model the spatial dependencies between transformer clusters using a graph attention network, achieve adaptive aggregation of spatial features, and extract the updated features of the transformer clusters. The load prediction module is used to input the extracted and updated features of the transformer area into a temporal convolutional network. Through multi-layer dilated convolution, it captures the multi-scale time dependence and long-term trend changes of the load sequence, and obtains the load distribution of the transformer area at different confidence levels.
[0012] A computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to perform the aforementioned method for predicting short-term load probability for a distribution area.
[0013] A processor for running a program, wherein the program executes the aforementioned method for predicting short-term load probability for a distribution area.
[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention introduces spatiotemporal feature extraction and uses K-means clustering to identify transformer substations with similar load characteristics. It utilizes LASSO for feature selection, constructs a hybrid model of graph attention network (GAT) and temporal convolutional network (TCN) to combine spatial and temporal dependencies, and employs quantile regression to quantify prediction uncertainty, significantly improving the accuracy and reliability of load forecasting. Attached Figure Description
[0015] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.
[0016] Figure 1 This is a flowchart of the method of the present invention.
[0017] Figure 2 This is a diagram of the architecture of the present invention.
[0018] Figure 3 This is a system schematic diagram of the present invention. Detailed Implementation
[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0020] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.
[0021] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.
[0022] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.
[0023] Example 1: like Figure 1 As shown, this embodiment provides a method for predicting short-term load probability in a transformer area, including the following steps: Step S1: Extract the load data of the transformer area including multi-dimensional features, including historical load data, historical population flow data, meteorological data, and date data, and calculate the peak-valley difference coefficient of the load.
[0024] Among them, the peak-valley difference coefficient of load is calculated based on the daily load curve. K pv The details are as follows: ; in, P max , P min These represent the peak and trough values of the daily load, respectively. This represents the daily average load.
[0025] Step S2: Based on the extracted multidimensional features of the transformer area load data, obtain the correlation between the transformer area load and meteorological data and population flow data, and obtain the comprehensive spatial features based on meteorological factors and the comprehensive spatial features based on population flow.
[0026] The correlation between transformer load and meteorological data and population flow data, obtained from the extracted multidimensional features of transformer load data, specifically includes the following steps: (1) Let the time series of the load in the transformer area be as follows: P t The meteorological data series is W t， The meteorological data includes various meteorological factors, specifically temperature, humidity, wind speed, normal irradiance, and diffuse irradiance. The correlation coefficient between the substation load and the meteorological data is: ; ; Among them, cov( P t , W t，r ) is a load time series P t With the rth meteorological factor W t，r The covariance of , where r=1,2,3,4,5, corresponding to temperature, humidity, wind speed, normal irradiance, and diffuse irradiance, respectively; , , respectively, represent the standard deviations of the load and the r-th meteorological factor sequence; Based on the temporal characteristics of the r-th meteorological factor, this method is used to quantify the response intensity of different substation loads to climate change. (2) The impact of factors such as resident travel and holiday migration on the load characteristics of the transformer area is gradually becoming apparent. Using population flow data, the correlation between load and population flow is calculated to measure the degree of impact of population changes on load fluctuations. Let the population flow time series be... M t The correlation coefficient between transformer load and population flow data is calculated as follows: ; These are the standard deviations of the population mobility time series, and cov( P t , M t ) is a load time series P t Covariance with the time series of population movement; This is based on the temporal characteristics of population movement.
[0027] Changes in the correlation between the load of a distribution area and external meteorological and population flow factors can reflect changes in the load pattern of that area over time. Furthermore, changes in this correlation across different distribution areas can reflect the spatial correlation between the loads of different distribution areas. Obtaining the comprehensive spatial characteristics based on meteorological factors and the comprehensive spatial characteristics based on population flow specifically includes the following steps: (1) Using a sliding window [ t -L, t The user load curve and meteorological factor sequence curve are segmented, and the Pearson correlation coefficient between the load curve and different meteorological factors is calculated within each sliding window to obtain the time... t Based on meteorological factors r Time correlation The time correlation vector of meteorological factors r is constructed by concatenating the data in chronological order. As shown in the following formula: ; The Pearson correlation coefficient between the time correlation vectors of different transformer areas is calculated as follows: ; Where corr() represents the Pearson correlation coefficient calculation function, Indicates the area i Taiwan j Based on meteorological factors r Spatial correlation; and They are respectively the Taiwan area i Taiwan j The time correlation vector of meteorological factors; The mean of the spatial correlation of each transformer area with all other transformer areas based on meteorological factors is used as a comprehensive spatial correlation feature of that transformer area based on meteorological factor r. The specific calculation method is as follows: ; Indicates the area i Based on the comprehensive spatial correlation characteristics of meteorological factor r; (2) Using a sliding window [ t -L, t The user load curve and population flow time series curve are segmented, and the Pearson correlation coefficient between the load curve and the population flow time series is calculated within each sliding window to obtain the time... t Temporal correlation of population flow time series Concatenate them in chronological order to form a population mobility time correlation vector As shown in the following formula: ; The Pearson correlation coefficient between the time correlation vectors of different transformer areas is calculated as follows: ; in, Indicates the area i Taiwan j Based on the spatial correlation of population mobility; and They are respectively the Taiwan area i Taiwan j The population mobility time correlation vector; The mean of the spatial correlation of each transformer area with all other transformer areas based on population flow is used as a comprehensive spatial correlation feature of that transformer area based on population flow. The specific calculation method is as follows: ; Indicates the area i Comprehensive spatial correlation characteristics based on population mobility.
[0028] Step S3: Based on the comprehensive spatial characteristics based on meteorological factors, the comprehensive spatial characteristics based on population flow, and the peak-valley difference coefficient of the load, the K-means clustering algorithm is used to cluster the N transformer areas to obtain transformer area clusters with similar spatiotemporal characteristics.
[0029] To eliminate the influence of different feature dimensions on the clustering results, Z-Score normalization is first used to preprocess all features. Based on the normalized multidimensional features, K-means clustering algorithm is used to group the transformer areas. This algorithm iteratively optimizes... N Each district is divided into K In each cluster.
[0030] The K-means clustering algorithm is used to cluster N transformer substations to obtain substation clusters with similar spatiotemporal characteristics. The specific steps include: (1) In N K initial cluster centers were randomly selected from each of the transformer substations. Each cluster center includes the peak-to-valley difference coefficient of the load in the substation area and the comprehensive spatial correlation characteristics based on the meteorological factor r. Comprehensive spatial correlation characteristics based on population flow ; (2) Calculate the Euclidean distance from each substation to each cluster center, and assign each substation to the substation cluster containing the cluster center with the smallest Euclidean distance; (3) Recalculate the cluster centers of each station area cluster based on the current cluster allocation; (4) Repeat (2)-(3) until the maximum number of iterations is reached or the cluster center no longer changes, and finally obtain K clusters of stations with similar spatiotemporal characteristics.
[0031] By extracting the spatiotemporal features of the transformer substations, clustering was performed to obtain substation clusters with significant spatiotemporal correlation, thereby improving the spatiotemporal correlation of graph nodes.
[0032] Step S4: Select multidimensional features based on LASSO to obtain the key features closely related to load changes after screening.
[0033] After spatiotemporal feature extraction and substation clustering, the input data includes historical load, meteorological elements (including temperature, humidity, wind speed, normal irradiance, diffuse irradiance), population flow, date type, and other multi-dimensional features. These features contain redundant information to some extent.
[0034] To further improve the training efficiency and generalization performance of subsequent models, the Least Absolute Shrinkage and Selection Operator (LASSO) is introduced for feature selection. LASSO is a feature selection and regularization method for regression analysis. It achieves sparsity of model parameters by adding L1 norm constraints to the loss function, thereby automatically selecting key features.
[0035] LASSO, through the introduction of L1 regularization, makes some regression coefficients exactly zero during the optimization process, thereby achieving automatic feature screening. It can be used to screen features that have a significant impact on transformer load forecasting.
[0036] For each station cluster, construct a feature matrix. ,in M=K*T , K The number of clusters in the station area. T The number of time samples for each station cluster. is the feature dimension. Its corresponding regression model can be expressed as: ; This is the load forecast value. For constant terms, The regression coefficients are the characteristics.
[0037] Introduce a regularization parameter By optimizing the objective function ; in Y For the load value, preprocessed data on different meteorological factors and external characteristics such as population flow are subjected to LASSO feature selection, and the regularization parameter is adjusted accordingly. LASSO can balance the model's fitting accuracy with feature sparsity, and select the key features most closely related to load changes. , These are the filtered feature dimensions.
[0038] The LASSO feature selection mechanism is used to optimize multi-source features, which significantly reduces the interference of redundant features on model training and improves prediction accuracy and generalization ability.
[0039] Step S5 involves using a graph attention network (GAT) to model the spatial dependencies between transformer clusters, achieving adaptive aggregation of spatial features and extracting updated features for each transformer cluster. GAT processes graph-structured data, introducing an attention mechanism to adaptively calculate the importance weights of each node's neighbors, thus achieving differentiated aggregation of neighbor information. Unlike traditional graph convolutional networks (GCNs) that assign fixed weights to all neighbors, GAT can dynamically learn complex dependencies between nodes, more accurately capturing local spatial features in the graph structure.
[0040] Graph attention networks such as Figure 2 As shown, the specific steps include: (1) Constructing a graph structure ,in V It is a graph structure G A set of nodes, where each node represents a cluster of stations. E It is a graph structure G The edge connection set, A Representative graph structure G An adjacency matrix describes the connection relationships between graph nodes and is constructed based on the correlation between graph nodes. (2) After obtaining the graph structure G, the spatial dependencies between the clusters are extracted using a graph attention network, and the feature matrix filtered based on LASSO is then used. Reconstructed as a node-time feature tensor H , ;in, K The number of clusters in the station area. T The number of time samples for each station cluster. The feature dimensions of the feature matrix after LASSO filtering; (3) Each station cluster i At time step t The initial features are ;in This indicates that the background region clusters have been selected through LASSO feature selection. i At time step t eigenvectors; Taiwan Cluster i With each of its adjacent station clustersj Attention coefficient needs to be calculated This coefficient is used to represent the cluster of adjacent station areas. j Taiwan Cluster i The importance of attention coefficients is calculated using the features of two clusters of stations, as shown in the following formula: ; in This represents the concatenation of node features, where W is the weight matrix. a T The parameter vector is LeakyReLU, which is the activation function, introducing non-linear characteristics to enhance the model's expressive power. (4) The calculated attention coefficient Normalize to weights a ij The calculation formula is as follows: ; in, N i For the cluster of the Taiwan region i The set of adjacent station clusters is used to achieve feature aggregation between station clusters by weighted aggregation of the features of each adjacent station cluster. The aggregation formula is as follows: ; in, For activation function, The features of the station cluster are updated using a graph attention network.
[0041] Step S6 involves inputting the extracted and updated features of the transformer area into a temporal convolutional network. Through multi-layer dilated convolution, the network captures the multi-scale time dependence and long-term trend changes of the load sequence, yielding the load distribution of the transformer area at different confidence levels. Recurrent Neural Networks (RNNs) are commonly used neural networks for processing time-series data. However, traditional RNNs suffer from problems such as vanishing and exploding gradients, limiting their effectiveness on long-series data. Temporal Convolutional Networks (TCNs), as an improved structure of RNNs, can expand the receptive field through dilated convolution and improve training stability through residual connections, making them more efficient at modeling long-term dependencies.
[0042] Specifically, the following steps are included: After extracting spatially dependent features using a graph attention network (GAT), this paper uses a temporal convolutional network (TCN) to model the temporal features of the transformer clusters in order to capture the load patterns of the transformers as they evolve over time.
[0043] The extracted and updated features of the transformer area are used as the initial input to the temporal convolutional network. z i,tTo characterize the long-term dependencies of the load sequence, dilated convolution is introduced. By adding a dilation factor to the convolution operation, the receptive field is expanded, which is beneficial for clustered substations. i The expression for calculating the dilated convolution at time t is: ; in, Where d is the kernel size and d is the dilation factor. W k For convolution kernel weights, For dilated convolution in t Output at any moment Indicates the corresponding cluster of the upper layer. i At time step td*k Output; expansion coefficient d As the number of network layers increases exponentially by 2, long-term dependencies can be captured without increasing the number of network layers; In a multi-layer TCN, each layer contains residual connections to ensure training stability, as detailed below: ; F() represents the TCN layer operations, including dilated convolution and activation functions. Indicates the upper-level cluster of the platform area. i At time step t The output, Indicates the current layer cluster i At time step t The output; Finally, the output of the last layer of the TCN is mapped to the predicted load values at different quantiles through a fully connected layer: ; in, and These correspond to quantiles. The output layer weights and biases; This represents the corresponding quantile. The predicted load value; Next, quantile regression is used as the loss function to characterize the load distribution at different confidence levels, thereby obtaining interval prediction results; its loss function is the total loss of each quantile loss, defined as: ; in, y i , t The actual load value of cluster i in the transformer area at time t. Let i be the predicted load value for cluster i in transformer area at time t. The selected set of quantiles; The quantile loss function, The definition is as follows: .
[0044] This invention employs the GAT-TCN model method to achieve deep joint learning of the spatial dependence and temporal evolution characteristics of clusters in the station area, thereby enhancing the model's ability to express spatiotemporal features and capture long-term dependencies.
[0045] In summary, this invention first extracts spatiotemporal feature clustering. By fusing historical load, meteorological factors, and population flow data, it extracts the temporal and spatial features of transformer substations. K-means clustering is used to divide substations with similar electricity consumption behaviors into substation clusters. A graph structure is constructed using substation clusters as nodes, improving the performance of the prediction model from the perspective of graph node generation. Secondly, key features are selected based on LASSO. The Least Absolute Shrinkage and Selection Operator (LASSO) is introduced in the feature input stage for feature selection. L1 regularization constraints are used to achieve feature sparsity and automatically remove redundant features that contribute little to load changes. Finally, spatiotemporal modeling is achieved through GAT-TCN. The Graph Attention Network (GAT) is used to model the spatial dependencies between substation clusters. The attention mechanism adaptively assigns importance weights to different neighboring nodes to achieve adaptive aggregation of spatial features. Subsequently, the spatially updated node features are input into the Temporal Convolutional Network (TCN), which captures the multi-scale temporal dependencies and long-term trend changes of the load sequence through multi-layer dilated convolution.
[0046] Example 2: Based on the same inventive concept as Embodiment 1, this embodiment provides a prediction system for short-term load probability in a transformer area, such as... Figure 3 As shown, the method described includes: The data extraction module is used to extract load data of the transformer area including multi-dimensional features, such as historical load data, historical population flow data, meteorological data, and date data, and to calculate the peak-valley difference coefficient of the load. The spatial feature calculation module is used to obtain the correlation between the load of the substation area and meteorological data and population flow data based on the extracted multi-dimensional features of the substation area load data, and to obtain the comprehensive spatial features based on meteorological factors and the comprehensive spatial features based on population flow. The transformer area clustering module is used to cluster N transformer areas based on comprehensive spatial characteristics based on meteorological factors, comprehensive spatial characteristics based on population flow, and peak-valley difference coefficient of load, using the K-means clustering algorithm to obtain transformer area clusters with similar spatiotemporal characteristics. The feature selection module is used to select multidimensional features based on LASSO to obtain key features closely related to load changes after screening. The feature update module is used to model the spatial dependencies between transformer clusters using a graph attention network, achieve adaptive aggregation of spatial features, and extract the updated features of the transformer clusters. The load prediction module is used to input the extracted and updated features of the transformer area into a temporal convolutional network. Through multi-layer dilated convolution, it captures the multi-scale time dependence and long-term trend changes of the load sequence, and obtains the load distribution of the transformer area at different confidence levels.
[0047] Example 3: Based on the same inventive concept as Embodiment 1, this embodiment provides a computer-readable storage medium, which includes a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to execute the aforementioned method for predicting short-term load probability of a distribution area.
[0048] Example 4: Based on the same inventive concept as Embodiment 1, this embodiment provides a processor for running a program, wherein the program executes the aforementioned method for predicting short-term load probability of a transformer area.
[0049] Those skilled in the art will recognize that the modules of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components of the examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the invention.
[0050] In the embodiments provided by this invention, it should be understood that the division of modules is only a logical functional division. In actual implementation, there may be other division methods, such as multiple modules can be combined into one module, one module can be split into multiple modules, or some features can be ignored.
[0051] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.
[0052] If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.
[0053] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.
Claims
1. A method for predicting short-term load probability of a transformer area, characterized in that, The method comprises the following steps: Step S1, extracting the substation load data comprising multi-dimensional features, wherein the multi-dimensional features comprise historical load data, historical population flow data, meteorological data, and date data, and calculating a peak-valley difference coefficient of the load; Step S2, obtaining the correlation between the substation load and the meteorological data and the population flow data according to the extracted multi-dimensional feature substation load data, and obtaining a comprehensive spatial feature based on meteorological factors and a comprehensive spatial feature based on population flow; Step S3, clustering the N substations according to the comprehensive spatial feature based on meteorological factors and the comprehensive spatial feature based on population flow and the peak-valley difference coefficient of the load by using a K-means clustering algorithm, to obtain substation clusters with similar spatio-temporal features; Step S4, selecting the multi-dimensional features based on LASSO, to obtain key features closely related to the load change after screening; Step S5, modeling the spatial dependency relationship between the substation clusters by using a graph attention network, to realize adaptive aggregation of the spatial features and extract updated features of the substations; Step S6, inputting the extracted updated features of the substations into a time convolution network, capturing multi-scale time dependency and long-term trend change of the load sequence through multiple layers of dilated convolution, and obtaining the load distribution of the substations at different confidence levels.
2. The method for short-term load probability prediction of a transformer area according to claim 1, characterized in that, The peak-valley difference coefficient of the load is calculated in step S1 based on the daily load curve K pv The specific process is as follows: ； wherein P max , P min are the peak and valley of the daily load, respectively, is the average daily load.
3. The method for short-term load probability prediction of a transformer area as claimed in claim 1 wherein, The correlation between the substation load and the meteorological data and the population flow data obtained according to the extracted multi-dimensional feature substation load data in Step S2 comprises the following steps: (1) Set the time series of the load of the transformer area as P t , and the time series of the meteorological data as W t， wherein the meteorological data includes different meteorological factors, specifically including temperature, humidity, wind speed, normal irradiance, and diffuse irradiance, and the correlation coefficient of the load of the transformer area and the meteorological data is: ； ； cov (T, H, V, S, D) = cov (T, H, V, S, D) - cov (T, H, V, S, D) (1) P t , W t，r cov (T, H, V, S, D) = cov (T, H, V, S, D) - cov (T, H, V, S, D) (1) P t cov (T, H, V, S, D) = cov (T, H, V, S, D) - cov (T, H, V, S, D) (1) W t，r cov (T, H, V, S, D) = cov (T, H, V, S, D) - cov (T, H, V, S, D) (1) 、 cov (T, H, V, S, D) = cov (T, H, V, S, D) - cov (T, H, V, S, D) (1) cov (T, H, V, S, D) = cov (T, H, V, S, D) - cov (T, H, V, S, D) (1) (2) Set the population flow time series as M t The correlation coefficient between the district load and the population flow data is calculated as follows: ； respectively the standard deviation of the population flow time series, cov( P t , M t ) the load time series P t the covariance with the population flow time series; is a time feature based on population flow.
4. The method for short-term load probability prediction of a transformer area as claimed in claim 3 wherein, The comprehensive spatial feature based on meteorological factors and the comprehensive spatial feature based on population flow obtained in Step S2 comprise the following steps: (1) The user load curve and the meteorological factor sequence curve are segmented using a sliding window t -L, t ] and the Pearson correlation coefficient between the load curve and different meteorological factors is calculated in each sliding window to obtain a time t based on the time correlation of meteorological factors r , which is spliced into a meteorological factor r time correlation vector in time sequence , as shown in the following formula: ； The Pearson correlation coefficient between the time correlation vectors of different substations is calculated, as shown in the following formula: ； where corr() represents the Pearson correlation coefficient calculation function, denotes the substation i and the substation j based on the spatial correlation of meteorological factors; r and are the meteorological factor time correlation vectors of the substation i and the substation j respectively. averaging the spatial correlation of each station with all other stations based on the meteorological factor as a comprehensive spatial correlation feature of the station based on the meteorological factor r The specific calculation is as follows: ； representative station area i comprehensive spatial correlation degree features based on meteorological factors r (2) Segmentation of the user load curve and the population flow time series curve using a sliding window t -L, t and calculation of the Pearson correlation coefficient between the load curve and the population flow time series within each sliding window, resulting in a time t time correlation based on the population flow time series vector of population flow time correlations in chronological order as shown in the following equation: ； The Pearson correlation coefficient between the time correlation vectors of different substations is calculated, as shown in the following formula: ； wherein, denotes a station area i and a station area j a spatial correlation based on population flows; and denotes a station area i and a station area j a temporal correlation vector of population flows; averaging the population flow-based spatial correlation of each station with all other stations as a population flow-based integrated spatial correlation feature of the station The specific calculation is as follows: ； representative station area i comprehensive spatial correlation degree features based on population flow 5. The method for short-term load probability prediction of a transformer area as claimed in claim 4 wherein, The K-means clustering algorithm is used to cluster the N substations in Step S3 to obtain substation clusters with similar spatio-temporal features, which comprises the following steps: (1) In N K initial clustering centers are randomly selected in the substation area Each clustering center includes the peak-valley difference coefficient of the load of the substation, the comprehensive spatial correlation degree feature based on meteorological factors r , and the comprehensive spatial correlation degree feature based on population flow ; (2) calculating the Euclidean distance of each substation to each clustering center, and assigning each substation to the substation cluster where the clustering center with the smallest Euclidean distance is located; (3) recalculating the clustering center of each substation cluster according to the current cluster assignment; (4) repeating (2)-(3) until the maximum number of iterations set is reached or the clustering center no longer changes, and finally obtaining K substation clusters with similar spatio-temporal features.
6. The method for short-term load probability prediction of a transformer area as claimed in claim 1 wherein, The graph attention network is used to model the spatial dependency relationship between the substation clusters in Step S5 to realize adaptive aggregation of the spatial features and extract the spatial features of the substations, which comprises the following steps: (1) Constructing a graph structure wherein V is a node set of the graph structure G , each node representing a cluster of a power district, E is a connection set of edges of the graph structure G , and A represents an adjacency matrix of the graph structure G , describing the connection relationship between the nodes of the graph, and being constructed according to the correlation between the nodes of the graph; (2) After obtaining the graph structure G, the spatial dependence between the transformer cluster is extracted by using the graph attention network, and the feature matrix screened based on LASSO is reconstructed into a node-time feature tensor , H , ; wherein, K is the number of transformer clusters, T is the number of time samples of each transformer cluster, is the feature dimension of the feature matrix screened based on LASSO. (3) each cluster of zones i At time step t The initial feature is ; where is the feature vector at time step i At time step t is the feature vector at time step Cluster of exchanges i Each contiguous cluster of exchanges j The attention coefficient needs to be calculated which represents the importance of the contiguous cluster of exchanges j to the cluster of exchanges i The attention coefficient is calculated from the characteristics of the two clusters of exchanges as follows: ; wherein represents concatenation of node features, W is a weight matrix, a T is a parameter vector, LeakyReLU activation function; (4) The calculated attention coefficient is normalized as a weight a ij The calculation formula is as follows: ； wherein, N i For the cluster of transformer area i The set of adjacent transformer area clusters, the feature aggregation between the transformer area clusters is realized by weighted aggregation of the features of each adjacent transformer area cluster, and the aggregation formula is as follows: ； wherein, is an activation function, is the feature of the transformer cluster updated by the graph attention network.
7. The method for short-term load probability prediction of a transformer area as claimed in claim 6 wherein, Step S6 comprises the following steps: The extracted features of the updated station area are taken as initial inputs of the time convolution network z i,t ; introduce dilated convolution, expand the receptive field by adding a dilated factor in the convolution operation, for the station area cluster i The dilated convolution calculation expression at time t: ； in, Where d is the kernel size and d is the dilation factor. W k For convolution kernel weights, For dilated convolution in t Output at any moment Indicates the corresponding cluster of the upper layer. i At time step t-d*k Output; expansion coefficient d It increases exponentially by 2 with the number of network layers; In the multi-layer TCN, each layer of TCN contains a residual connection to ensure training stability, which is as follows: ； F() denotes TCN layer operations, including dilated convolutions and activation functions, denotes the output of the previous layer of clusters i at time step t , denotes the output of the current layer of clusters i at time step t ; Finally, the output of the last layer of TCN is mapped to the predicted load value at different quantiles through a fully connected layer: ； wherein, and are output layer weights and bias corresponding to quantile denotes a predicted load value corresponding to quantile ; Next, the loss function uses quantile regression to realize the description of the load distribution at different confidence levels, so as to obtain the interval prediction result; the loss function is the total loss of each quantile loss, which is defined as: ； wherein, y i , t is the actual load value of the cluster i at time t, is the predicted load value of the cluster i at time t, is the selected set of quantiles; For the quantile loss function, is defined as follows: 。 8. A system for predicting short-term load probability of a transformer area, characterized by, The method of any one of claims 1-7 is applied, comprising: The data extraction module is configured to extract feeder area load data including multi-dimensional features, wherein the multi-dimensional features include historical load data, historical population flow data, meteorological data, and date data, and a peak-valley difference coefficient of the load is calculated; The spatial feature calculation module is configured to obtain correlations between feeder area load and meteorological data and population flow data according to the extracted feeder area load data including multi-dimensional features, and obtain comprehensive spatial features based on meteorological factors and comprehensive spatial features based on population flow; The feeder area clustering module is configured to cluster N feeder areas according to the comprehensive spatial features based on meteorological factors and the comprehensive spatial features based on population flow and the peak-valley difference coefficient of the load, and obtain feeder area clusters with similar spatio-temporal features by using a K-means clustering algorithm; The feature selection module is configured to select multi-dimensional features based on LASSO, and obtain screened key features closely related to load changes; The feature updating module is configured to model spatial dependency relationships between feeder area clusters by using a graph attention network, realize adaptive aggregation of spatial features, and extract updated features of the feeder areas; The load prediction module is configured to input the extracted updated features of the feeder areas into a time convolution network, capture multi-scale time dependencies and long-term trend changes of load sequences by using multiple layers of dilated convolution, and obtain load distributions of the feeder areas at different confidence levels.
9. A computer-readable storage medium, characterized in that, The computer readable storage medium includes a stored program, wherein the program controls the device where the computer readable storage medium is located to execute the method for predicting feeder area short-term load probability according to any one of claims 1 to 7 when the program is running.
10. A processor, comprising: The processor is configured to run a program, wherein the program executes the method for predicting feeder area short-term load probability according to any one of claims 1 to 7 when the program is running.