Intelligent processing method and system for power consumption data in power system

Abstract

The invention provides an intelligent processing method and system for power consumption data in a power system, and relates to the technical field of power system data processing. Performing feature propagation calculation to obtain aggregation node features; abnormal nodes and normal nodes are identified, and an anomaly detection result with reliability measurement is generated; constructing a time sequence correlation constraint based on an abnormal detection result, and generating corrected power consumption data; and mapping the corrected power utilization data to a corresponding storage level according to the data quality grade and the access frequency characteristic, and configuring a differentiated redundancy protection mechanism. According to the invention, intelligent processing of the power consumption data is realized, and the data quality and the storage efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of power system data processing technology, and in particular to a method and system for intelligent processing of electricity consumption data in power systems. Background Technology

[0002] The operation of power systems generates a large amount of electricity consumption data, which is of great value for monitoring the power grid's operational status, diagnosing faults, and optimizing dispatch. With the advancement of smart grid construction, the number of power system data acquisition devices has increased dramatically, and the volume of data has exploded, posing significant challenges to data processing and analysis. Traditional electricity consumption data processing methods mainly rely on statistical models and expert experience rules, which are insufficient for effectively handling large-scale, high-dimensional, and multi-source heterogeneous power system data.

[0003] In recent years, with the application of artificial intelligence technology in the power sector, deep learning-based data processing methods have gradually emerged. In particular, graph neural networks, with their ability to process structured data, have provided a new technical path for power system data analysis. Power systems inherently possess distinct network topology characteristics, with complex interrelationships between devices, which aligns perfectly with the features of graph structures. Therefore, graph-based intelligent processing methods show broad application prospects in power system data analysis.

[0004] Existing power system electricity consumption data processing technologies suffer from the following defects and shortcomings: First, traditional data processing methods fail to fully utilize the topological information of the power system, often analyzing nodes in isolation and ignoring the physical connections and mutual influences between nodes, resulting in a lack of systematicity and accuracy in the analysis results. Second, existing anomaly detection algorithms generally lack reliability assessment mechanisms, failing to quantify the uncertainty of detection results and providing a lack of reliable reference for subsequent decision-making, thus increasing the risk of misjudgment and missed detection. Finally, existing technologies adopt a uniform strategy for data storage management, failing to implement layered storage based on data quality and access characteristics. This wastes storage resources and fails to provide sufficient protection mechanisms for critical data, affecting the stability and efficiency of the data processing system. Summary of the Invention

[0005] The present invention provides a method and system for intelligent processing of electricity consumption data in a power system, which can solve the problems in the prior art.

[0006] A first aspect of the present invention provides a method for intelligent processing of electricity consumption data in a power system, comprising: Based on the time-series measurement values ​​of electricity consumption data in the power system and the relationship between the power grid topology, a device association graph structure is constructed. Feature propagation calculation is performed on the device association graph structure, and the graph convolution propagation mechanism is used to aggregate multi-hop neighborhood information of graph nodes to obtain aggregated node features. Identify abnormal and normal nodes in the aggregated node features, decode the abnormal nodes into anomaly type identification results and anomaly confidence distribution, and perform uncertainty quantification evaluation on the anomaly type identification results based on the anomaly confidence distribution to generate anomaly detection results with reliability metrics; Based on the anomaly detection results, a time-series correlation constraint is constructed, and the abnormal nodes are correlated with the normal nodes. The time-series evolution features and physical constraint boundaries are extracted, and the corrected electricity consumption data is generated based on the time-series evolution features and the physical constraint boundaries. The corrected power consumption data is mapped to the corresponding storage level according to the data quality level and access frequency characteristics, and a differentiated redundancy protection mechanism is configured for each storage level. The corrected power consumption data is then written to the corresponding storage level according to the redundancy protection mechanism.

[0007] Based on the time-series measurements of electricity consumption data from the power system and the relationship with the power grid topology, a device association graph structure is constructed, including: The electricity consumption time-series measurements are divided into multi-scale time windows; Within each time window, the rate of change and fluctuation amplitude of the electricity consumption time-series measurement values ​​are calculated and combined into time-series statistical features. The electricity consumption time-series measurement values ​​are transformed from the time domain to the frequency domain through a frequency domain transformation algorithm to obtain a frequency domain coefficient sequence. The main frequency component and spectral energy distribution are extracted from the frequency domain coefficient sequence and combined into frequency domain transformation features. The time-series statistical features and the frequency domain transform features are fused to obtain time-series fused features; Based on the power grid topology, each power terminal device is mapped as a graph node, the time-series fusion feature is assigned as the node feature of the corresponding graph node, a weighted adjacency matrix is ​​constructed according to the electrical connection type and connection strength between power terminal devices, and the power terminal device pairs corresponding to the non-zero elements in the weighted adjacency matrix are mapped as graph edges, and the weight values ​​in the weighted adjacency matrix are assigned as the edge weights of the corresponding graph edges. An adaptive threshold is calculated based on the weight distribution of the weighted adjacency matrix and the graph connectivity constraints. The weighted adjacency matrix is ​​then sparsified, retaining graph edges with edge weights greater than the adaptive threshold. The sparsified weighted adjacency matrix, the graph nodes, and the node features are then structurally organized to construct the device association graph structure.

[0008] Feature propagation calculations are performed on the device association graph structure. A graph convolution propagation mechanism is used to aggregate multi-hop neighborhood information of graph nodes, resulting in aggregated node features including: Extract the node feature matrix and sparse adjacency matrix from the device association graph structure; The number of initial graph convolutional propagation layers controls the range of hops for neighborhood information aggregation, and the feature transformation parameter matrix controls the linear transformation of node features. The node feature matrix is ​​multiplied by the feature transformation parameter matrix to obtain the transformed node feature matrix. The sparse adjacency matrix is ​​multiplied by the transformed node feature matrix to achieve weighted aggregation of the features of each graph node to its direct neighbor nodes and perform nonlinear activation transformation to obtain single-hop activation features. Based on the number of graph convolution propagation layers, multi-hop neighborhood information aggregation is iteratively performed. In each iteration, the single-hop activation feature of the previous layer is used as the input node feature of the current layer. The linear transformation of the feature transformation parameter matrix, the neighborhood aggregation of the sparse adjacency matrix, and the nonlinear activation transformation are repeatedly performed to expand the neighborhood information aggregation range layer by layer and obtain the multi-hop aggregated feature sequence. The activation features of each layer in the multi-hop aggregated feature sequence are concatenated across layers, and the concatenated features are mapped to a unified feature space through feature dimensionality reduction transformation to obtain the aggregated node features that fuse multi-hop neighborhood information.

[0009] Identifying abnormal and normal nodes in the aggregated node features includes: By using feature space projection transformation, the aggregated node features are mapped to a low-dimensional feature space to obtain projected node features; the feature distance metric between corresponding graph node pairs in the projected node features is used as matrix elements to construct a feature similarity matrix; The local neighborhood density and neighborhood dispersion of each graph node are extracted from the feature similarity matrix and weighted together to obtain the anomaly score of each graph node. A temporal consistency test is performed on the features of the aggregated nodes. The feature change magnitude and direction of each graph node within adjacent time windows are calculated, and the feature change magnitude and direction are combined into a temporal fluctuation index. The abnormal score value and the time series fluctuation index are fused and judged. For graph nodes whose abnormal score value exceeds a preset spatial abnormality threshold and whose time series fluctuation index exceeds a preset time series abnormality threshold, they are marked as an abnormal candidate node set. The abnormal candidate node set is subjected to graph topology verification. It is detected whether the graph nodes in the abnormal candidate node set form a connected subgraph in the device association graph structure. If a connected subgraph is formed, all graph nodes in the connected subgraph are retained as abnormal nodes; otherwise, they are marked as normal nodes.

[0010] The abnormal nodes are decoded into anomaly type identification results and anomaly confidence distributions. The anomaly type identification results are then subjected to uncertainty quantification assessment based on the anomaly confidence distributions to generate anomaly detection results with reliability metrics, including: The multi-class decoding network performs nonlinear transformation and feature mapping on the aggregate node features corresponding to the abnormal node, calculates the category probability value of the abnormal node belonging to the corresponding abnormal type on each classification output branch, and combines the category probability values ​​of all classification output branches into an abnormal type probability vector. The anomaly type with the highest probability value is selected from the anomaly type probability vector as the anomaly type identification result, and the anomaly type probability vector is used as the anomaly confidence distribution; the information entropy is obtained by performing a weighted logarithmic operation on each anomaly confidence value in the anomaly confidence distribution; The multi-class decoding network is subjected to multiple random perturbation inferences. In each inference, random perturbation is applied to the internal parameters of the multi-class decoding network and the anomaly type probability vector is recalculated to obtain multiple sets of anomaly type probability vector samples. The prediction variance index is calculated based on the multiple sets of anomaly type probability vector samples and fused with the information entropy to obtain a comprehensive uncertainty measure. A reliability score is calculated based on the comprehensive uncertainty metric, and the reliability score is inversely mapped to the comprehensive uncertainty metric. The anomaly type identification result, the anomaly confidence distribution, and the reliability score are combined to form the anomaly detection result with a reliability metric.

[0011] The abnormal nodes are correlated with the normal nodes to extract temporal evolution features and physical constraint boundaries. Based on the temporal evolution features and physical constraint boundaries, corrected electricity consumption data is generated, including: Based on the time-series correlation constraints, the electricity consumption time-series measurement value sequence of the abnormal node and the normal node within the historical time window is extracted and the time-series correlation coefficient is calculated; normal nodes whose time-series correlation coefficient exceeds the preset correlation threshold are selected to construct a strongly correlated normal node subset, and the time-series correlation coefficient is used as the correlation weight; Based on the electricity consumption time-series measurement sequence of the strongly correlated normal node subset and the correlation weight, the time-series predicted value sequence of the abnormal node is obtained by weighted time-series interpolation; the time-series deviation sequence between the electricity consumption time-series measurement sequence of the abnormal node and the time-series predicted value sequence is calculated, and the time-series evolution features and fluctuation components are extracted by time-series decomposition. Based on the node type and power grid topology location of the abnormal node in the device association graph structure, the corresponding upper power limit constraint and lower power limit constraint are determined as the physical constraint boundary. Based on the time-series evolution characteristics, the time-series predicted value sequence is trend-corrected to obtain a trend-corrected value. The trend-corrected value is then subjected to the physical constraint boundary for boundary clipping, and smoothing filtering is performed to suppress high-frequency noise in the fluctuation component, thereby obtaining the corrected electricity consumption data.

[0012] The corrected electricity consumption data is mapped to corresponding storage levels according to data quality level and access frequency characteristics, and a differentiated redundancy protection mechanism is configured for each storage level, including: Based on the reliability score in the anomaly detection results and the time-series deviation sequence of the corrected electricity consumption data, a data quality assessment value is calculated, and data quality levels are classified according to the data quality assessment value; the historical access count and access time distribution of the corrected electricity consumption data within a preset statistical period are statistically analyzed, access frequency feature values ​​are calculated, and access frequency levels are classified. Establish a multidimensional mapping table between the data quality level and the access frequency level and the storage level. Based on the data quality level and access frequency level of the corrected power consumption data, query the target storage level in the multidimensional mapping table and allocate the corrected power consumption data to the target storage level. For each storage tier, a data importance weight is determined based on the data quality level corresponding to the storage tier, and a data activity weight is determined based on the access frequency level corresponding to the storage tier. A redundancy requirement coefficient is calculated based on the data importance weight and the data activity weight. The number of redundant replicas, the redundancy distribution strategy, and the data verification strength are determined based on the redundancy requirement coefficient, and a redundancy protection mechanism corresponding to the storage tier is established.

[0013] A second aspect of the present invention provides an intelligent processing system for electricity consumption data in a power system, comprising: The first unit is used to construct a device association graph structure based on the time-series measurement values ​​of electricity consumption data of the power system and the topological relationship of the power grid; feature propagation calculation is performed on the device association graph structure, and the graph convolution propagation mechanism is used to aggregate the multi-hop neighborhood information of the graph nodes to obtain the aggregated node features. The second unit is used to identify abnormal nodes and normal nodes in the features of the aggregated nodes, decode the abnormal nodes into anomaly type identification results and anomaly confidence distribution, and perform uncertainty quantification evaluation on the anomaly type identification results based on the anomaly confidence distribution to generate anomaly detection results with reliability metrics. The third unit is used to construct time-series correlation constraints based on the anomaly detection results, perform correlation analysis between the abnormal nodes and the normal nodes, extract time-series evolution features and physical constraint boundaries, and generate corrected electricity consumption data based on the time-series evolution features and the physical constraint boundaries. The fourth unit is used to map the corrected power consumption data to the corresponding storage level according to the data quality level and access frequency characteristics, configure a differentiated redundancy protection mechanism for each storage level, and write the corrected power consumption data into the corresponding storage level according to the redundancy protection mechanism.

[0014] A third aspect of the present invention, An electronic device is provided, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0015] Fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0016] The beneficial effects of this application are as follows: By constructing a device association graph structure and performing feature propagation calculations, the system fully utilizes power grid topology relationships to achieve effective aggregation of multi-hop neighborhood information, improving the global perception and feature representation capabilities of data processing. Confidence distribution decoding and uncertainty quantification assessment are performed on abnormal nodes to generate anomaly detection results with reliability metrics, enhancing the accuracy and credibility of anomaly identification and reducing the risk of false alarms and missed alarms. Based on the anomaly detection results, temporal correlation constraints are constructed, and correlation analysis between abnormal and normal nodes is performed. Temporal evolution features and physical constraint boundaries are extracted, enabling precise correction of electricity consumption data and effectively improving data quality. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the intelligent processing method for electricity consumption data in a power system according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the abnormal node identification and classification process. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 embodiments of the present invention, and not all embodiments. 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.

[0019] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0020] Figure 1 This is a flowchart illustrating the intelligent processing method for electricity consumption data in a power system according to an embodiment of the present invention. Figure 1 As shown, the method includes: Based on the time-series measurement values ​​of electricity consumption data in the power system and the relationship between the power grid topology, a device association graph structure is constructed. Feature propagation calculation is performed on the device association graph structure, and the graph convolution propagation mechanism is used to aggregate multi-hop neighborhood information of graph nodes to obtain aggregated node features. Identify abnormal and normal nodes in the aggregated node features, decode the abnormal nodes into anomaly type identification results and anomaly confidence distribution, and perform uncertainty quantification evaluation on the anomaly type identification results based on the anomaly confidence distribution to generate anomaly detection results with reliability metrics; Based on the anomaly detection results, a time-series correlation constraint is constructed, and the abnormal nodes are correlated with the normal nodes. The time-series evolution features and physical constraint boundaries are extracted, and the corrected electricity consumption data is generated based on the time-series evolution features and the physical constraint boundaries. The corrected power consumption data is mapped to the corresponding storage level according to the data quality level and access frequency characteristics, and a differentiated redundancy protection mechanism is configured for each storage level. The corrected power consumption data is then written to the corresponding storage level according to the redundancy protection mechanism.

[0021] In one optional implementation, constructing a device association graph structure based on the time-series measurements of power consumption data and the relationship between the power grid topology includes: The electricity consumption time-series measurements are divided into multi-scale time windows; Within each time window, the rate of change and fluctuation amplitude of the electricity consumption time-series measurement values ​​are calculated and combined into time-series statistical features. The electricity consumption time-series measurement values ​​are transformed from the time domain to the frequency domain through a frequency domain transformation algorithm to obtain a frequency domain coefficient sequence. The main frequency component and spectral energy distribution are extracted from the frequency domain coefficient sequence and combined into frequency domain transformation features. The time-series statistical features and the frequency domain transform features are fused to obtain time-series fused features; Based on the power grid topology, each power terminal device is mapped as a graph node, the time-series fusion feature is assigned as the node feature of the corresponding graph node, a weighted adjacency matrix is ​​constructed according to the electrical connection type and connection strength between power terminal devices, and the power terminal device pairs corresponding to the non-zero elements in the weighted adjacency matrix are mapped as graph edges, and the weight values ​​in the weighted adjacency matrix are assigned as the edge weights of the corresponding graph edges. An adaptive threshold is calculated based on the weight distribution of the weighted adjacency matrix and the graph connectivity constraints. The weighted adjacency matrix is ​​then sparsified, retaining graph edges with edge weights greater than the adaptive threshold. The sparsified weighted adjacency matrix, the graph nodes, and the node features are then structurally organized to construct the device association graph structure.

[0022] A sliding window technique is employed to segment electricity consumption time-series measurements into multi-scale time windows, setting different time windows of varying scales, such as short-term (5 minutes), medium-term (30 minutes), and long-term (2 hours). This multi-scale window segmentation method allows for the simultaneous capture of instantaneous changes, short-term fluctuations, and long-term trends in the power system. For each power terminal device, its electricity consumption time-series measurements are segmented according to a preset window size and sliding step, forming a series of time window sequences.

[0023] Within each time window, the rate of change and fluctuation amplitude of electricity consumption time-series measurements are calculated. The rate of change is obtained by dividing the difference in electricity consumption between adjacent time points by the time interval, reflecting the speed of change in electrical load. The fluctuation amplitude is obtained by calculating the difference or standard deviation between the maximum and minimum values ​​within the window, reflecting the severity of electricity consumption fluctuations. These two indicators are combined into time-series statistical features. Simultaneously, the Fast Fourier Transform (FFT) algorithm is used to transform the electricity consumption time-series measurements from the time domain to the frequency domain, obtaining a frequency domain coefficient sequence. From this sequence, the dominant frequency component and spectral energy distribution are extracted. The dominant frequency component reflects the periodicity of electricity consumption, while the spectral energy distribution characterizes the contribution of different frequency components. These frequency domain information are combined to form frequency domain transformation features.

[0024] A strategy combining feature concatenation and weight fusion is adopted to concatenate time-series statistical feature vectors and frequency-domain transform feature vectors to form a joint feature vector. Different feature components are weighted by set weight coefficients to highlight the influence of important features. For example, for power equipment with large fluctuations, the weight of fluctuation amplitude features can be appropriately increased; for power equipment with obvious periodicity, the weight of the dominant frequency component features can be increased.

[0025] The temporal fusion features are assigned as node features to the corresponding graph nodes, forming a node feature matrix. A weighted adjacency matrix is ​​constructed based on the electrical connection type and connection strength between power terminal equipment. The connection type can be a direct electrical connection or a connection via a transformer, while the connection strength can be calculated based on factors such as physical distance, electrical impedance, or power transmission capacity. For example, equipment within the same distribution circuit has a higher connection strength, while equipment connected via multiple transformers has a lower connection strength. The power terminal equipment pairs corresponding to the non-zero elements in the weighted adjacency matrix are mapped to graph edges, and the weight values ​​in the weighted adjacency matrix are assigned as the edge weights of the corresponding graph edges.

[0026] The distribution characteristics of the adjacency matrix weights are analyzed, and their mean, variance, and quantiles are calculated. Combined with graph connectivity constraints—ensuring the connectivity of the graph structure is not disrupted—an iterative optimization method is used to determine the optimal threshold. For example, a smaller threshold can be started, and the threshold value can be gradually increased until the preset graph connectivity requirement is met. The weighted adjacency matrix is ​​then sparsified, retaining edges with weights greater than the adaptive threshold and deleting connections with smaller weights, thereby reducing noise and highlighting key connections.

[0027] The sparsed weighted adjacency matrix, graph nodes, and node features are structurally organized to construct a device association graph structure. The topological information of the graph is stored using data structures such as adjacency lists or adjacency matrices, and node features are mapped one-to-one with nodes in vector form. The resulting device association graph structure includes the topological connections, connection strengths, and power consumption characteristics of power terminal equipment, providing a unified data representation for subsequent tasks such as power grid analysis, fault detection, and load forecasting.

[0028] In practical applications, a city's smart grid system comprises hundreds of substations and thousands of terminal electrical devices. By deploying the aforementioned method, a device association graph structure covering the entire network was successfully constructed. This graph structure clearly displays the connection relationships and power consumption characteristics between different areas and different types of devices, helping maintenance personnel quickly identify abnormal power consumption behavior and potential fault points, thereby improving the reliability and efficiency of grid operation.

[0029] In one optional implementation, feature propagation calculation is performed on the device association graph structure, and the graph convolution propagation mechanism is used to aggregate multi-hop neighborhood information of graph nodes to obtain aggregated node features, including: Extract the node feature matrix and sparse adjacency matrix from the device association graph structure; The number of initial graph convolutional propagation layers controls the range of hops for neighborhood information aggregation, and the feature transformation parameter matrix controls the linear transformation of node features. The node feature matrix is ​​multiplied by the feature transformation parameter matrix to obtain the transformed node feature matrix. The sparse adjacency matrix is ​​multiplied by the transformed node feature matrix to achieve weighted aggregation of the features of each graph node to its direct neighbor nodes and perform nonlinear activation transformation to obtain single-hop activation features. Based on the number of graph convolution propagation layers, multi-hop neighborhood information aggregation is iteratively performed. In each iteration, the single-hop activation feature of the previous layer is used as the input node feature of the current layer. The linear transformation of the feature transformation parameter matrix, the neighborhood aggregation of the sparse adjacency matrix, and the nonlinear activation transformation are repeatedly performed to expand the neighborhood information aggregation range layer by layer and obtain the multi-hop aggregated feature sequence. The activation features of each layer in the multi-hop aggregated feature sequence are concatenated across layers, and the concatenated features are mapped to a unified feature space through feature dimensionality reduction transformation to obtain the aggregated node features that fuse multi-hop neighborhood information.

[0030] The device association graph structure can be represented as G = (V, E), where V represents the set of nodes and E represents the set of edges. Each node represents a device, and the edges between nodes represent the association relationships between devices. A node feature matrix X is extracted from this graph structure, where each row represents the original feature vector of a device. These features include device attribute information, historical behavior data, etc. Simultaneously, a sparse adjacency matrix A is extracted. The elements A[i][j] of this matrix represent the connection weight between node i and node j; if the two nodes are not connected, the value is 0. In practical applications, since the associations between devices are usually sparse, using a sparse matrix for storage can significantly reduce storage space and computational costs.

[0031] Initialize the graph convolution propagation layer number and feature transformation parameter matrix. The graph convolution propagation layer number L determines the range of hops for neighborhood information aggregation; for example, L=2 indicates aggregation of neighborhood information with a maximum of 2 hops. For each layer l (1≤l≤L), initialize the corresponding feature transformation parameter matrix W[l]. This matrix is ​​used to control the linear transformation of node features. The dimension of matrix W[l] is d_in×d_out, where d_in is the input feature dimension and d_out is the output feature dimension. These parameter matrices can be obtained through random initialization and training.

[0032] In the first layer of graph convolution propagation, the node feature matrix X is multiplied by the first layer feature transformation parameter matrix W[1] to obtain the transformed node feature matrix H[1]_temp = X·W[1]. The sparse adjacency matrix A is multiplied by the transformed node feature matrix H[1]_temp, i.e., H[1]_agg = A·H[1]_temp. This step realizes the weighted aggregation of the features of each graph node to its direct neighbor nodes. In order to enhance the expressive power of the model, a nonlinear activation function, such as the ReLU (Modified Linear Unit) function, is applied to the aggregation result to obtain the single-hop activation feature H[1] = ReLU(H[1]_agg). The ReLU function can introduce nonlinearity and improve the model's ability to express complex features.

[0033] Based on the set number of graph convolution propagation layers L, multi-hop neighborhood information aggregation is iteratively performed. In each iteration, the single-hop activation feature H[l-1] of the previous layer is used as the input node feature of the current layer. For the l-th layer (2≤l≤L), a linear transformation of the feature transformation parameter matrix W[l] is performed to obtain H[l]_temp = H[l-1]·W[l]. Neighborhood aggregation of the sparse adjacency matrix A is performed, i.e., H[l]_agg = A·H[l]_temp. A nonlinear activation function is applied to obtain the activation feature H[l]=ReLU(H[l]_agg) of the current layer. In this way, with each additional graph convolution layer, the range of neighborhood information aggregation is expanded by one hop, and the range of neighborhood information aggregation is expanded layer by layer, finally obtaining the multi-hop aggregation feature sequence {H[1], H[2], ..., H[L]}.

[0034] In practical applications, neighborhoods with different hop counts contain different types of information. For example, in device security analysis, a one-hop neighborhood reveals information about directly interacting devices, while a two-hop or more distant neighborhood reveals broader network structure information. Therefore, it is crucial to perform cross-layer feature concatenation of activation features from different layers in a multi-hop aggregated feature sequence, which can fuse information from neighbors at different distances.

[0035] The feature sequence {H[1], H[2], ..., H[L]} is concatenated along the feature dimension to form H_concat = [H[1]|| H[2] || ... || H[L]], where "||" represents the concatenation operation of the feature dimension. The concatenated feature dimension is L×d_out, which is too high and will lead to increased computational complexity and overfitting risk. Therefore, the concatenated features are mapped to a unified feature space through feature dimensionality reduction transformation, i.e., H_final = H_concat·W_final, where W_final is the dimensionality reduction transformation matrix with a dimension of (L×d_out)×d_unified, and d_unified is the dimension of the final unified feature space. Finally, the aggregated node feature H_final, which integrates multi-hop neighborhood information, is obtained.

[0036] In practical applications, such as power equipment anomaly detection, this method can be used to capture complex interaction patterns between devices. By aggregating multi-hop neighborhood information, it can analyze the similarities and differences in behavior between the current device and its directly and indirectly connected devices, thereby more accurately identifying potential abnormal behavior patterns.

[0037] The advantage of this graph convolution-based feature propagation method is that it can effectively capture the relationships and topological information between devices and integrate them into the node feature representation, thereby improving the accuracy and effectiveness of subsequent tasks such as anomaly detection and device classification.

[0038] In one optional implementation, identifying abnormal and normal nodes in the aggregated node features includes: By using feature space projection transformation, the aggregated node features are mapped to a low-dimensional feature space to obtain projected node features; the feature distance metric between corresponding graph node pairs in the projected node features is used as matrix elements to construct a feature similarity matrix; The local neighborhood density and neighborhood dispersion of each graph node are extracted from the feature similarity matrix and weighted together to obtain the anomaly score of each graph node. A temporal consistency test is performed on the features of the aggregated nodes. The feature change magnitude and direction of each graph node within adjacent time windows are calculated, and the feature change magnitude and direction are combined into a temporal fluctuation index. The abnormal score value and the time series fluctuation index are fused and judged. For graph nodes whose abnormal score value exceeds a preset spatial abnormality threshold and whose time series fluctuation index exceeds a preset time series abnormality threshold, they are marked as an abnormal candidate node set. The abnormal candidate node set is subjected to graph topology verification. It is detected whether the graph nodes in the abnormal candidate node set form a connected subgraph in the device association graph structure. If a connected subgraph is formed, all graph nodes in the connected subgraph are retained as abnormal nodes; otherwise, they are marked as normal nodes.

[0039] like Figure 2 As shown, the method includes: The system identifies anomalous nodes in the aggregated node features of the device association graph structure. Through feature space projection transformation, these features are mapped to a low-dimensional feature space, thereby reducing feature dimensionality while preserving data structure information. Specifically, Principal Component Analysis (PCA) is used to reduce the dimensionality of the original high-dimensional features, selecting the top k principal components to form the low-dimensional feature space. In practical applications, the value of k can be determined based on the cumulative contribution rate, typically selecting the number of principal components with a cumulative contribution rate of 85% or higher. After dimensionality reduction, the projected node features corresponding to each graph node are obtained, with the dimensionality reduced from tens or hundreds of dimensions to typically no more than 20 dimensions.

[0040] Based on the obtained projected node features, a feature similarity matrix is ​​constructed. The feature distance between any two graph nodes is calculated using Euclidean distance. For the projected feature vectors Fi and Fj of node i and node j, their distance values ​​are calculated and converted into similarity values, which are then used as the values ​​of matrix elements (i, j). The similarity can be transformed using a Gaussian kernel function, resulting in higher similarity values ​​for nodes that are closer together.

[0041] Local neighborhood information for each graph node is extracted from the feature similarity matrix, including local neighborhood density and neighborhood dispersion. Local neighborhood density reflects the closeness between a node and its surrounding nodes; it is calculated by counting the number of neighboring nodes whose similarity to the current node is higher than a preset threshold. Neighborhood dispersion measures the distance between a node and its nearest high-density nodes, reflecting the degree of isolation between the node and dense regions. These two indicators are fused using a weighted combination to obtain the anomaly score for each graph node. The weight coefficients can be determined through validation set tuning; generally, a density weight to dispersion weight ratio of 4:6 is reasonable.

[0042] A temporal consistency test is performed on the features of aggregated nodes, analyzing the change patterns of nodes over time. The time series is divided into multiple consecutive time windows, typically ranging from 1 hour to 24 hours, adjusted according to the business scenario. The changes in the feature vectors of each graph node within adjacent time windows are calculated, including the magnitude and direction of change. The magnitude of change is calculated using the norm of the vector difference, while the direction of change is determined by the cosine of the angle between the vectors. The magnitude and direction information are combined to form a temporal fluctuation index. A larger index indicates that the node behavior has changed significantly within a short period of time.

[0043] Anomaly scores are fused with temporal fluctuation indicators to identify potential anomalous nodes. Spatial and temporal anomaly thresholds are set, determined through historical data analysis, typically selecting the 95th percentile of their respective distributions. Graph nodes whose anomaly scores exceed both the spatial and temporal fluctuation indicators are marked as a set of candidate anomalous nodes. This dual-threshold mechanism effectively reduces false alarms and improves detection accuracy.

[0044] A graph topology verification method is used to examine the connectivity of the set of candidate anomalous nodes within the device association graph. A node is randomly selected as a seed node, and a breadth-first search algorithm is used to explore other candidate anomalous nodes directly connected to it. If a connected subgraph containing multiple candidate anomalous nodes can be formed, all nodes in that subgraph are retained as the final confirmed anomalous nodes. If candidate anomalous nodes are isolated from each other in the graph structure, they are marked as normal nodes. This topology-based verification mechanism is based on the observation that anomalous events often exhibit propagation and correlation.

[0045] In one optional implementation, the abnormal node is decoded into an anomaly type identification result and an anomaly confidence distribution, and the anomaly type identification result is subjected to uncertainty quantification evaluation based on the anomaly confidence distribution to generate an anomaly detection result with reliability metrics, including: The multi-class decoding network performs nonlinear transformation and feature mapping on the aggregate node features corresponding to the abnormal node, calculates the category probability value of the abnormal node belonging to the corresponding abnormal type on each classification output branch, and combines the category probability values ​​of all classification output branches into an abnormal type probability vector. The anomaly type with the highest probability value is selected from the anomaly type probability vector as the anomaly type identification result, and the anomaly type probability vector is used as the anomaly confidence distribution; the information entropy is obtained by performing a weighted logarithmic operation on each anomaly confidence value in the anomaly confidence distribution; The multi-class decoding network is subjected to multiple random perturbation inferences. In each inference, random perturbation is applied to the internal parameters of the multi-class decoding network and the anomaly type probability vector is recalculated to obtain multiple sets of anomaly type probability vector samples. The prediction variance index is calculated based on the multiple sets of anomaly type probability vector samples and fused with the information entropy to obtain a comprehensive uncertainty measure. A reliability score is calculated based on the comprehensive uncertainty metric, and the reliability score is inversely mapped to the comprehensive uncertainty metric. The anomaly type identification result, the anomaly confidence distribution, and the reliability score are combined to form the anomaly detection result with a reliability metric.

[0046] A multi-class decoding network performs nonlinear transformations and feature mappings on the aggregate node features corresponding to abnormal nodes. In the actual implementation, the multi-class decoding network adopts a fully connected neural network structure, containing two hidden layers and one output layer. The first hidden layer contains 128 neurons, the second hidden layer contains 64 neurons, and the number of neurons in the output layer is equal to the number of predefined abnormal types. For the input aggregate node feature vector, it is transformed through the first hidden layer, and the ReLU activation function is used to enhance the nonlinear expressive power; the second hidden layer further extracts discriminative features; and the output layer maps it to the space of each abnormal type.

[0047] Each output branch of the multi-class decoding network corresponds to a predefined anomaly type, such as "hardware failure," "software anomaly," or "power connection interruption." At each branch, the probability value of an anomaly node belonging to the corresponding anomaly type is calculated using the Softmax function. Specifically, if there are K predefined anomaly types, the initial values ​​of the output layer are z1, z2, ..., z... k The probability values ​​for each type are obtained by transforming the data using the Softmax function: For type i, its probability value is calculated by dividing the current type's exponent value by the sum of all type exponent values. The combination of the category probability values ​​of all classification output branches forms the anomaly type probability vector P = [p1, p2, ..., p...]. k ].

[0048] The anomaly type with the highest probability value is selected from the anomaly type probability vector as the anomaly type identification result. If the anomaly type probability vector is P = [0.15, 0.75, 0.05, 0.05], then the identification result is the second anomaly type. Simultaneously, the anomaly type probability vector P is used as the anomaly confidence distribution for subsequent uncertainty quantification assessment.

[0049] To assess the uncertainty of the identification results, the information entropy of the anomaly confidence distribution is calculated. During the information entropy calculation, a weighted logarithmic operation is performed on each anomaly confidence value in the distribution. In practical implementation, for the anomaly type probability vector P = [p1, p2, ..., p...],... k The information entropy H is calculated based on the weighted logarithm of each probability value. A higher information entropy value indicates that the model has a lower certainty about the current recognition result.

[0050] To further quantify uncertainty, multiple random perturbation inferences were performed on the multi-class decoding network. In each inference process, Gaussian random noise was applied to the internal parameters of the multi-class decoding network, with the perturbation standard deviation set at 2% of the original parameter values. After 10 random perturbation inferences, the anomaly type probability vector was recalculated, resulting in 10 sets of anomaly type probability vector samples. This technique simulates the posterior distribution sampling of a Bayesian neural network.

[0051] Based on these 10 sets of anomaly type probability vector samples, a prediction variance index is calculated. Specifically, for each anomaly type location, the variance of the probability values ​​of the 10 sets of samples at that location is calculated, and the average of the variances across all locations is taken to obtain the average prediction variance. This prediction variance index reflects the impact of model parameter uncertainty on the prediction results.

[0052] Information entropy and prediction variance are fused using a weighted average to obtain a comprehensive uncertainty measure, U. The weights are adjusted according to the specific application scenario; by default, the information entropy weight is 0.6 and the prediction variance weight is 0.4. This fusion strategy combines two dimensions: data uncertainty and model uncertainty.

[0053] The reliability score R is calculated based on the comprehensive uncertainty metric. The reliability score and the comprehensive uncertainty metric have an inverse mapping relationship, which is transformed using an exponential decay function R = exp(-αU), where α is the decay coefficient, set to 5. This ensures that when uncertainty is high, the reliability score is low, and vice versa.

[0054] The anomaly type identification result "sensor drift," the anomaly confidence distribution [0.05, 0.08, 0.72, 0.15], and the reliability score of 0.83 are combined to form the final anomaly detection result, output as a triple <anomaly type, confidence distribution, reliability score>. This anomaly detection result with reliability measurement provides more comprehensive information support for subsequent decision-making, helping to improve the accuracy and efficiency of anomaly handling.

[0055] In one optional implementation, the abnormal nodes are correlated with the normal nodes to extract temporal evolution features and physical constraint boundaries. Based on these temporal evolution features and physical constraint boundaries, corrected electricity consumption data is generated, including: Based on the time-series correlation constraints, the electricity consumption time-series measurement value sequence of the abnormal node and the normal node within the historical time window is extracted and the time-series correlation coefficient is calculated; normal nodes whose time-series correlation coefficient exceeds the preset correlation threshold are selected to construct a strongly correlated normal node subset, and the time-series correlation coefficient is used as the correlation weight; Based on the electricity consumption time-series measurement sequence of the strongly correlated normal node subset and the correlation weight, the time-series predicted value sequence of the abnormal node is obtained by weighted time-series interpolation; the time-series deviation sequence between the electricity consumption time-series measurement sequence of the abnormal node and the time-series predicted value sequence is calculated, and the time-series evolution features and fluctuation components are extracted by time-series decomposition. Based on the node type and power grid topology location of the abnormal node in the device association graph structure, the corresponding upper power limit constraint and lower power limit constraint are determined as the physical constraint boundary. Based on the time-series evolution characteristics, the time-series predicted value sequence is trend-corrected to obtain a trend-corrected value. The trend-corrected value is then subjected to the physical constraint boundary for boundary clipping, and smoothing filtering is performed to suppress high-frequency noise in the fluctuation component, thereby obtaining the corrected electricity consumption data.

[0056] The constraint strategy is determined based on the type of anomaly (data missing, mutation, or equipment failure) and the confidence distribution. For data missing anomalies, time continuity constraints are established; for mutation anomalies, statistical regularity constraints (such as the 3σ principle) are constructed; and for equipment failure anomalies, constraints are established based on equipment characteristics. Simultaneously, different weights are assigned according to the anomaly confidence level, and a priority mechanism (physical characteristics > statistical regularity > time continuity) is used when constraints conflict, resulting in time-series correlation constraints.

[0057] Based on time-series correlation constraints, the electricity consumption time-series measurement sequences of abnormal nodes and normal nodes within the historical time window are extracted. The past 30 days are selected as the historical time window. For each abnormal node and normal node, the daily electricity consumption data within the time window are extracted to form two time-series sequences of length 30. The time-series correlation coefficient between the abnormal node and each normal node is calculated by the Pearson correlation coefficient, which is the product of the covariance of the two sequences divided by their standard deviations.

[0058] A subset of strongly correlated normal nodes is constructed by selecting normal nodes whose time-series correlation coefficients exceed a preset correlation threshold. In practical applications, the preset correlation threshold can be set to 0.7, meaning that when the correlation between the electricity consumption patterns of an abnormal node and a normal node reaches 0.7 or higher, they are considered to have a strong correlation. Simultaneously, these time-series correlation coefficients are used as correlation weights for subsequent weighted calculations.

[0059] Based on the electricity consumption time-series measurement sequence and association weights of a strongly correlated normal node subset, a weighted time-series interpolation is used to calculate the time-series predicted value sequence of abnormal nodes. For time point t, the predicted value of the abnormal node can be obtained through the following steps: obtain the electricity consumption measurement value of each node in the strongly correlated normal node subset at time t; based on the association weights, perform a weighted average of these measurement values ​​to obtain the predicted value of the abnormal node at time t. Repeating this process yields the time-series predicted value sequence of the abnormal node throughout the entire time window.

[0060] The time-series deviation sequence between the measured and predicted electricity consumption sequences of anomalous nodes is calculated. The time-series deviation sequence equals the measured sequence minus the predicted sequence. A seasonal decomposition method is used to decompose this deviation sequence temporally, extracting time-series evolution characteristics and fluctuation components. The deviation sequence is decomposed into a trend term, a seasonal term, and a residual term. The trend term reflects the long-term trend and serves as the time-series evolution characteristic; the seasonal term and the residual term constitute the fluctuation components.

[0061] Based on the node type and grid topology location of abnormal nodes in the device association graph structure, corresponding upper and lower power constraints are determined as physical constraint boundaries. For example, for residential user nodes, the upper limit constraint can be determined based on the user's historical maximum electricity consumption and equipment capacity; for industrial user nodes, the constraint range can be determined based on their contracted capacity and operating characteristics. In practical applications, the upper limit constraint can be set to 1.2 times the historical peak value, and the lower limit constraint can be set to 0.8 times the historical valley value, to balance the accuracy and flexibility of data correction.

[0062] Trend correction values ​​are obtained by performing trend correction on the time-series predicted value sequence based on its temporal evolution characteristics. The extracted trend term is then added to the time-series predicted value sequence to obtain the preliminary trend correction value. Physical constraints are applied to the trend correction value for boundary pruning to ensure that the corrected data does not exceed a reasonable range. When the correction value exceeds the upper limit, it is set as the upper limit value; when the correction value is below the lower limit, it is set as the lower limit value.

[0063] The data after boundary clipping is smoothed and filtered to suppress high-frequency noise in the fluctuation component. A moving average filter is used to average the data at two time points before and after the data point to achieve a smooth transition of the data and obtain the corrected power consumption data.

[0064] In a real-world application scenario, an anomaly was observed in the electricity meter data of a household in a residential area. Using the method described above, it was identified that this household's electricity consumption patterns were highly correlated with those of five other households in the same area (correlation coefficients all exceeding 0.8), thus constructing a subset of strongly correlated normal nodes. Based on the data from these normal users, the normal electricity consumption curve for the abnormal user was predicted. Combined with the user's historical electricity consumption characteristics, physical constraints were determined, with an upper limit of 5 kW and a lower limit of 0.2 kW. The final corrected electricity consumption data not only preserved seasonal electricity consumption characteristics but also eliminated abnormal fluctuations in the original data, making the corrected data more consistent with the user's actual electricity consumption behavior.

[0065] The above methods effectively corrected abnormal electricity consumption data, ensuring the accuracy and continuity of power grid data and providing reliable data support for subsequent power system operation analysis and electricity consumption behavior research.

[0066] In one optional implementation, the corrected electricity consumption data is mapped to corresponding storage levels according to data quality level and access frequency characteristics, and a differentiated redundancy protection mechanism is configured for each storage level, including: Based on the reliability score in the anomaly detection results and the time-series deviation sequence of the corrected electricity consumption data, a data quality assessment value is calculated, and data quality levels are classified according to the data quality assessment value; the historical access count and access time distribution of the corrected electricity consumption data within a preset statistical period are statistically analyzed, access frequency feature values ​​are calculated, and access frequency levels are classified. Establish a multidimensional mapping table between the data quality level and the access frequency level and the storage level. Based on the data quality level and access frequency level of the corrected power consumption data, query the target storage level in the multidimensional mapping table and allocate the corrected power consumption data to the target storage level. For each storage tier, a data importance weight is determined based on the data quality level corresponding to the storage tier, and a data activity weight is determined based on the access frequency level corresponding to the storage tier. A redundancy requirement coefficient is calculated based on the data importance weight and the data activity weight. The number of redundant replicas, the redundancy distribution strategy, and the data verification strength are determined based on the redundancy requirement coefficient, and a redundancy protection mechanism corresponding to the storage tier is established.

[0067] The reliability score for each electricity consumption data record is obtained from the anomaly detection module. This score reflects the confidence level that the data is considered normal. Simultaneously, the time-series deviation sequence between the corrected electricity consumption data and the original data is calculated, including peak deviation, root mean square deviation, and continuity deviation. The reliability score is assigned a weight w1, and the comprehensive index of the time-series deviation sequence is assigned a weight w2. The weighted average is calculated as the data quality assessment value Q. Based on a preset quality threshold, Q is divided into three data quality levels: high, medium, and low.

[0068] The system statistically analyzes historical access data for electricity consumption within a preset statistical period after correction, recording the number of times each data point was queried or used, as well as the access time distribution characteristics, including the proportion of access during working hours, the proportion of access during non-working hours, and the concentration of access time. A weight v1 is assigned to the historical access frequency, and a weight v2 is assigned to the access time distribution characteristics to calculate the access frequency feature value F. Based on a preset frequency threshold, F is divided into three access frequency levels: high frequency, medium frequency, and low frequency.

[0069] A multi-dimensional mapping table is established between data quality levels, access frequency levels, and storage tiers, employing a three-tier storage architecture: a high-performance solid-state storage tier, a medium-speed hybrid storage tier, and a low-cost archive storage tier. The mapping table is designed as follows: high-quality, frequently accessed data is mapped to the high-performance solid-state storage tier; high-quality data but with medium or low-frequency access, and medium-quality data but with high-frequency access, are mapped to the medium-speed hybrid storage tier; and the remaining combinations are mapped to the low-cost archive storage tier. The target storage tier for the corrected power consumption data is determined by looking up the table, and data allocation operations are performed.

[0070] For each storage tier, a differentiated redundancy protection mechanism is configured, and the data importance weight Wi is determined based on the data quality level corresponding to the storage tier. High-quality data is assigned a higher importance weight, medium-quality data a second higher, and low-quality data a lower weight. Simultaneously, the data activity weight Wa is determined based on the access frequency level. High-frequency access data is assigned a higher activity weight, medium-frequency data a second higher, and low-frequency data a lower weight.

[0071] The redundancy requirement coefficient R is calculated based on the data importance weight Wi and the data activity weight Wa: R = α×Wi + β×Wa, where α and β are balancing parameters. Specific redundancy protection strategies are determined based on the redundancy requirement coefficient R, including the number of redundant copies, redundancy distribution strategy, and data verification strength.

[0072] For the high-performance solid-state storage tier, the redundancy mechanisms include: maintaining three data copies distributed across different physical nodes; using strong checksums such as Reed-Solomon (10,4) encoding; and implementing a real-time synchronous replication strategy. For the medium-speed hybrid storage tier, the redundancy mechanisms include: maintaining two data copies; using medium-strength checksums such as Reed-Solomon (6,2) encoding; and implementing a near real-time synchronous replication strategy. For the low-cost archive storage tier, the redundancy mechanisms include: maintaining a single data copy and periodically backing it up; using basic checksums such as parity checking; and implementing a periodic batch replication strategy.

[0073] In actual operation, the data distribution and access pattern changes of each storage level are evaluated regularly, and the parameters and thresholds of the multidimensional mapping table are dynamically adjusted. When a significant change in the access frequency of a certain type of data or an adjustment in the quality assessment value is detected, the data migration mechanism is triggered to redistribute the data to a more suitable storage level and update the redundancy protection mechanism accordingly.

[0074] Through the aforementioned differentiated storage and protection mechanisms, the efficiency of storage resource utilization and the requirements of data security are effectively balanced, providing the most suitable storage strategy for power consumption data of different qualities and access characteristics, thereby improving the overall system performance and reliability.

[0075] This invention relates to an intelligent power consumption data processing system in a power system, the system comprising: The first unit is used to construct a device association graph structure based on the time-series measurement values ​​of electricity consumption data of the power system and the topological relationship of the power grid; feature propagation calculation is performed on the device association graph structure, and the graph convolution propagation mechanism is used to aggregate the multi-hop neighborhood information of the graph nodes to obtain the aggregated node features. The second unit is used to identify abnormal nodes and normal nodes in the features of the aggregated nodes, decode the abnormal nodes into anomaly type identification results and anomaly confidence distribution, and perform uncertainty quantification evaluation on the anomaly type identification results based on the anomaly confidence distribution to generate anomaly detection results with reliability metrics. The third unit is used to construct time-series correlation constraints based on the anomaly detection results, perform correlation analysis between the abnormal nodes and the normal nodes, extract time-series evolution features and physical constraint boundaries, and generate corrected electricity consumption data based on the time-series evolution features and the physical constraint boundaries. The fourth unit is used to map the corrected power consumption data to the corresponding storage level according to the data quality level and access frequency characteristics, configure a differentiated redundancy protection mechanism for each storage level, and write the corrected power consumption data into the corresponding storage level according to the redundancy protection mechanism.

[0076] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0077] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0078] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0079] 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; and these 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.

Claims

1. A method for intelligent processing of electricity consumption data in a power system, characterized in that, include: Based on the relationship between the time-series measurements of electricity consumption data from the power system and the power grid topology, a device association graph structure is constructed. Feature propagation calculation is performed on the device association graph structure, and the graph convolution propagation mechanism is used to aggregate multi-hop neighborhood information of graph nodes to obtain aggregated node features. Identify abnormal and normal nodes in the aggregated node features, decode the abnormal nodes into anomaly type identification results and anomaly confidence distribution, and perform uncertainty quantification evaluation on the anomaly type identification results based on the anomaly confidence distribution to generate anomaly detection results with reliability metrics; Based on the anomaly detection results, a time-series correlation constraint is constructed, and the abnormal nodes are correlated with the normal nodes. The time-series evolution features and physical constraint boundaries are extracted, and the corrected electricity consumption data is generated based on the time-series evolution features and the physical constraint boundaries. The corrected power consumption data is mapped to the corresponding storage level according to the data quality level and access frequency characteristics, and a differentiated redundancy protection mechanism is configured for each storage level. The corrected power consumption data is then written to the corresponding storage level according to the redundancy protection mechanism.

2. The method according to claim 1, characterized in that, Based on the time-series measurements of electricity consumption data from the power system and the relationship with the power grid topology, a device association graph structure is constructed, including: The electricity consumption time-series measurements are divided into multi-scale time windows; Within each time window, the rate of change and fluctuation amplitude of the electricity consumption time-series measurement values ​​are calculated and combined into time-series statistical features. The electricity consumption time-series measurement values ​​are transformed from the time domain to the frequency domain through a frequency domain transformation algorithm to obtain a frequency domain coefficient sequence. The main frequency component and spectral energy distribution are extracted from the frequency domain coefficient sequence and combined into frequency domain transformation features. The time-series statistical features and the frequency domain transform features are fused to obtain time-series fused features; Based on the power grid topology, each power terminal device is mapped as a graph node, the time-series fusion feature is assigned as the node feature of the corresponding graph node, a weighted adjacency matrix is ​​constructed according to the electrical connection type and connection strength between power terminal devices, and the power terminal device pairs corresponding to the non-zero elements in the weighted adjacency matrix are mapped as graph edges, and the weight values ​​in the weighted adjacency matrix are assigned as the edge weights of the corresponding graph edges. An adaptive threshold is calculated based on the weight distribution of the weighted adjacency matrix and the graph connectivity constraints. The weighted adjacency matrix is ​​then sparsified, retaining graph edges with edge weights greater than the adaptive threshold. The sparsified weighted adjacency matrix, the graph nodes, and the node features are then structurally organized to construct the device association graph structure.

3. The method according to claim 1, characterized in that, Feature propagation calculations are performed on the device association graph structure. A graph convolution propagation mechanism is used to aggregate multi-hop neighborhood information of graph nodes, resulting in aggregated node features including: Extract the node feature matrix and sparse adjacency matrix from the device association graph structure; The number of initial graph convolutional propagation layers controls the range of hops for neighborhood information aggregation, and the feature transformation parameter matrix controls the linear transformation of node features. The node feature matrix is ​​multiplied by the feature transformation parameter matrix to obtain the transformed node feature matrix. The sparse adjacency matrix is ​​multiplied by the transformed node feature matrix to achieve weighted aggregation of the features of each graph node to its direct neighbor nodes and perform nonlinear activation transformation to obtain single-hop activation features. Based on the number of graph convolution propagation layers, multi-hop neighborhood information aggregation is iteratively performed. In each iteration, the single-hop activation feature of the previous layer is used as the input node feature of the current layer. The linear transformation of the feature transformation parameter matrix, the neighborhood aggregation of the sparse adjacency matrix, and the nonlinear activation transformation are repeatedly performed to expand the neighborhood information aggregation range layer by layer and obtain the multi-hop aggregated feature sequence. The activation features of each layer in the multi-hop aggregated feature sequence are concatenated across layers, and the concatenated features are mapped to a unified feature space through feature dimensionality reduction transformation to obtain the aggregated node features that fuse multi-hop neighborhood information.

4. The method according to claim 1, characterized in that, Identifying abnormal and normal nodes in the aggregated node features includes: By using feature space projection transformation, the aggregated node features are mapped to a low-dimensional feature space to obtain projected node features; the feature distance metric between corresponding graph node pairs in the projected node features is used as matrix elements to construct a feature similarity matrix; The local neighborhood density and neighborhood dispersion of each graph node are extracted from the feature similarity matrix and weighted together to obtain the anomaly score of each graph node. A temporal consistency test is performed on the features of the aggregated nodes. The feature change magnitude and direction of each graph node within adjacent time windows are calculated, and the feature change magnitude and direction are combined into a temporal fluctuation index. The abnormal score value and the time series fluctuation index are fused and judged. For graph nodes whose abnormal score value exceeds a preset spatial abnormality threshold and whose time series fluctuation index exceeds a preset time series abnormality threshold, they are marked as an abnormal candidate node set. The abnormal candidate node set is subjected to graph topology verification. It is detected whether the graph nodes in the abnormal candidate node set form a connected subgraph in the device association graph structure. If a connected subgraph is formed, all graph nodes in the connected subgraph are retained as abnormal nodes; otherwise, they are marked as normal nodes.

5. The method according to claim 1, characterized in that, The abnormal nodes are decoded into anomaly type identification results and anomaly confidence distributions. The anomaly type identification results are then subjected to uncertainty quantification assessment based on the anomaly confidence distributions to generate anomaly detection results with reliability metrics, including: The multi-class decoding network performs nonlinear transformation and feature mapping on the aggregate node features corresponding to the abnormal node, calculates the category probability value of the abnormal node belonging to the corresponding abnormal type on each classification output branch, and combines the category probability values ​​of all classification output branches into an abnormal type probability vector. The anomaly type with the highest probability value is selected from the anomaly type probability vector as the anomaly type identification result, and the anomaly type probability vector is used as the anomaly confidence distribution; the information entropy is obtained by performing a weighted logarithmic operation on each anomaly confidence value in the anomaly confidence distribution; The multi-class decoding network is subjected to multiple random perturbation inferences. In each inference, random perturbation is applied to the internal parameters of the multi-class decoding network and the anomaly type probability vector is recalculated to obtain multiple sets of anomaly type probability vector samples. The prediction variance index is calculated based on the multiple sets of anomaly type probability vector samples and fused with the information entropy to obtain a comprehensive uncertainty measure. A reliability score is calculated based on the comprehensive uncertainty metric, and the reliability score is inversely mapped to the comprehensive uncertainty metric. The anomaly type identification result, the anomaly confidence distribution, and the reliability score are combined to form the anomaly detection result with a reliability metric.

6. The method according to claim 1, characterized in that, The abnormal nodes are correlated with the normal nodes to extract temporal evolution features and physical constraint boundaries. Based on the temporal evolution features and physical constraint boundaries, corrected electricity consumption data is generated, including: Based on the time-series correlation constraints, the electricity consumption time-series measurement value sequence of the abnormal node and the normal node within the historical time window is extracted and the time-series correlation coefficient is calculated; normal nodes whose time-series correlation coefficient exceeds the preset correlation threshold are selected to construct a strongly correlated normal node subset, and the time-series correlation coefficient is used as the correlation weight; Based on the electricity consumption time-series measurement sequence of the strongly correlated normal node subset and the correlation weight, the time-series predicted value sequence of the abnormal node is obtained by weighted time-series interpolation; the time-series deviation sequence between the electricity consumption time-series measurement sequence of the abnormal node and the time-series predicted value sequence is calculated, and the time-series evolution features and fluctuation components are extracted by time-series decomposition. Based on the node type and power grid topology location of the abnormal node in the device association graph structure, the corresponding upper power limit constraint and lower power limit constraint are determined as the physical constraint boundary. Based on the time-series evolution characteristics, the time-series predicted value sequence is trend-corrected to obtain a trend-corrected value. The trend-corrected value is then subjected to the physical constraint boundary for boundary clipping, and smoothing filtering is performed to suppress high-frequency noise in the fluctuation component, thereby obtaining the corrected electricity consumption data.

7. The method according to claim 1, characterized in that, The corrected electricity consumption data is mapped to corresponding storage levels according to data quality level and access frequency characteristics, and a differentiated redundancy protection mechanism is configured for each storage level, including: Based on the reliability score in the anomaly detection results and the time-series deviation sequence of the corrected electricity consumption data, a data quality assessment value is calculated, and data quality levels are classified according to the data quality assessment value; the historical access count and access time distribution of the corrected electricity consumption data within a preset statistical period are statistically analyzed, access frequency feature values ​​are calculated, and access frequency levels are classified. Establish a multidimensional mapping table between the data quality level and the access frequency level and the storage level. Based on the data quality level and access frequency level of the corrected power consumption data, query the target storage level in the multidimensional mapping table and allocate the corrected power consumption data to the target storage level. For each storage tier, a data importance weight is determined based on the data quality level corresponding to the storage tier, and a data activity weight is determined based on the access frequency level corresponding to the storage tier. A redundancy requirement coefficient is calculated based on the data importance weight and the data activity weight. The number of redundant replicas, the redundancy distribution strategy, and the data verification strength are determined based on the redundancy requirement coefficient, and a redundancy protection mechanism corresponding to the storage tier is established.

8. An intelligent power consumption data processing system in a power system, used to implement the method as described in any one of claims 1-7, characterized in that, include: The first unit is used to construct a device association diagram structure based on the time-series measurement values ​​of electricity consumption data of the power system and the relationship between the power grid topology; Feature propagation calculation is performed on the device association graph structure, and the graph convolution propagation mechanism is used to aggregate multi-hop neighborhood information of graph nodes to obtain aggregated node features. The second unit is used to identify abnormal nodes and normal nodes in the features of the aggregated nodes, decode the abnormal nodes into anomaly type identification results and anomaly confidence distribution, and perform uncertainty quantification evaluation on the anomaly type identification results based on the anomaly confidence distribution to generate anomaly detection results with reliability metrics. The third unit is used to construct time-series correlation constraints based on the anomaly detection results, perform correlation analysis between the abnormal nodes and the normal nodes, extract time-series evolution features and physical constraint boundaries, and generate corrected electricity consumption data based on the time-series evolution features and the physical constraint boundaries. The fourth unit is used to map the corrected power consumption data to the corresponding storage level according to the data quality level and access frequency characteristics, configure a differentiated redundancy protection mechanism for each storage level, and write the corrected power consumption data into the corresponding storage level according to the redundancy protection mechanism.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images