A commercial building energy consumption anomaly detection method based on a graph method

Abstract

The application provides a kind of commercial building energy consumption anomaly detection method based on graph method, it is related to intelligent building technical field, the method includes: based on the similarity graph structure is constructed, wherein node mapping is subentry partition subtable, and edge weight represents the time series similarity between corresponding nodes;By calculating the eigenvalue set of the Laplacian matrix of the graph structure, the spectral distribution representing the topological characteristics of the graph structure is generated, and the hierarchical clustering of multiple graph structures is performed based on the Battacharya distance between the spectral distributions, to obtain the energy consumption pattern clustering results of different time periods;The clustering results are analyzed by profile coefficient, and the number of class clusters corresponding to the maximum profile coefficient is selected as the corresponding division, and the class cluster set containing historical normal data is defined as the baseline class cluster.The application can improve the accuracy of building energy consumption anomaly detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent buildings, in particular to a commercial building energy consumption anomaly detection method based on a graph method. BACKGROUND

[0002] With the increasing demand for energy in the global building sector, building energy consumption management has become a key area for improving energy efficiency and reducing carbon emissions. The introduction of sub-metering and sub-zone metering technology in commercial buildings provides more detailed energy consumption data, but traditional methods have significant limitations in handling such complex data, especially in identifying abnormal usage patterns. Anomaly detection plays a crucial role in building energy consumption management. However, traditional statistical or supervised learning methods struggle to capture complex dependencies in sub-metering data.

[0003] Existing technologies have many limitations in building energy consumption anomaly detection, and their practicality is difficult to effectively adapt to complex energy consumption patterns in building environments.

[0004] Firstly, existing methods mainly rely on univariate analysis, usually focusing only on the energy consumption measured by a single meter, and cannot comprehensively analyze the complex relationships between multiple sub-meters. Secondly, existing technologies are inadequate in dealing with dynamic changes in building energy consumption. In practical applications, energy consumption patterns may fluctuate significantly due to changes in tenant behavior, environmental conditions, or equipment operating status, while traditional methods struggle to adapt flexibly to these dynamic patterns, resulting in high false positive rates or serious false negative rates in anomaly detection. In addition, most existing methods rely on large amounts of labeled data for model training, but in the field of building energy consumption management, it is often very difficult to obtain labeled data. Therefore, existing technologies have obvious defects in unsupervised detection capabilities, making it difficult to meet the needs of practical engineering applications. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a commercial building energy consumption anomaly detection method based on a graph method, which can improve the accuracy of building energy consumption anomaly detection.

[0006] To solve the above technical problems, the technical solutions of the present application are as follows:

[0007] In a first aspect, a commercial building energy consumption anomaly detection method based on a graph method, the method comprising:

[0008] Collecting real-time energy consumption data from sub-metering and sub-zone meters within the building, generating multiple groups of energy consumption time series with time series characteristics;

[0009] Performing dynamic time warping calculation on the energy consumption time series, quantifying the energy consumption time series similarity between each sub-metering and sub-zone meter, and constructing a graph structure based on the similarity, wherein the nodes are mapped to the sub-metering and sub-zone meters, and the edge weights represent the time series similarity between the corresponding nodes.

[0010] By calculating the eigenvalue set of the Laplacian matrix of the graph structure, a spectral distribution representing the topological characteristics of the graph structure is generated, and hierarchical clustering is performed on multiple graph structures based on the Battacharyya distance between the spectral distributions to obtain the clustering results of the energy consumption patterns in different time periods;

[0011] The clustering results are analyzed by the profile coefficient, and the number of clusters corresponding to the maximum profile coefficient is selected as the corresponding division. The cluster set containing historical normal data is defined as the baseline cluster.

[0012] A new energy consumption data is generated to generate a detected graph structure, and the minimum spectral distance between the new energy consumption data and all graph structures in the baseline cluster is calculated. If the minimum value exceeds a predetermined threshold, it is determined to be an abnormal graph structure.

[0013] The node connection strength analysis is performed on the abnormal graph structure, the edge weight drop amplitude of each node and its adjacent node is calculated, and the sub-item partition sub-table associated with the node with the maximum edge weight loss is marked as an abnormal source.

[0014] According to the spatial distribution characteristics and energy consumption mutation time sequence characteristics of the abnormal source, a detection report containing equipment maintenance prompt and energy efficiency optimization strategy is generated.

[0015] Further, the energy consumption time sequence is calculated by dynamic time warping, the energy consumption time sequence similarity between each sub-item partition sub-table is quantified, and a graph structure is constructed based on the similarity, wherein the node is mapped to the sub-item partition sub-table, and the edge weight represents the time sequence similarity between the corresponding nodes, including:

[0016] Based on the collected energy consumption data, all sub-item partition sub-tables in the building are identified, each sub-table corresponds to an independent energy consumption monitoring unit of a specific area or equipment, and the energy consumption time sequence of each sub-table is extracted;

[0017] For the energy consumption time sequence of any two sub-item partition sub-tables, dynamic time warping calculation is performed: multiple matching paths are generated by nonlinear time axis alignment, the Euclidean distance square of adjacent matching elements in each path is calculated, the total cumulative distance is obtained by accumulating all local distance squares on the path, the corresponding path with the minimum total cumulative distance is selected, and the square root of the total cumulative distance of the path is taken as the dynamic time warping distance of the two sequences, which represents the similarity of the energy consumption time sequences of the two sub-tables.

[0018] The dynamic time warping distance is mapped to the interval (0, 1) through a linear normalization function to generate an edge weight value, wherein the smaller the distance, the closer the similarity score after mapping to 1, and the larger the distance, the closer the score to 0.

[0019] A full connection graph structure is constructed by taking the sub-tables as nodes and the normalized similarity scores as edge weights, wherein the weight value of the edge between any two nodes reflects the similarity degree of the corresponding sub-tables' energy consumption time series.

[0020] Further, the spectral distribution representing the topological characteristics of the graph structure is generated by calculating the eigenvalue set of the Laplacian matrix of the graph structure, and the hierarchical clustering of multiple graph structures is performed based on the Battacharyya distance between the spectral distributions, to obtain the energy consumption pattern clustering results of different time periods, including:

[0021] For each graph structure of each time period, a standardized Laplacian matrix is constructed based on the edge weights between nodes, and eigenvalue decomposition is performed thereon to extract the eigenvalue set arranged in ascending order, thereby generating the spectral distribution representing the topological characteristics of the graph structure;

[0022] For any two spectral distributions of graph structures, the Battacharyya distance is calculated: the difference between the eigenvalues at corresponding positions in the two eigenvalue sets is squared and accumulated, and the square root of the accumulated result is taken, and the obtained value is defined as the spectral distance between the two graph structures;

[0023] Based on the spectral distance matrix between all graph structures, a hierarchical clustering algorithm is used to gradually merge the graph structures or clusters with the closest distance in a single linkage strategy until a preset number of initial cluster sets is formed;

[0024] According to the tree diagram generated in the hierarchical clustering process, the hierarchical division result satisfying the preset merging threshold is intercepted, and the graph structures of different time periods are classified into corresponding energy consumption pattern clusters to form a set of clustering results with time continuity.

[0025] Further, the silhouette coefficient analysis is performed on the clustering results, and the cluster number corresponding to the maximum silhouette coefficient is selected as the corresponding division, and the cluster set containing historical normal data is defined as the baseline cluster, including:

[0026] Based on the clustering result set generated by hierarchical clustering, a division scheme is set for each candidate cluster number, and the silhouette coefficient of each data point in each division scheme is calculated: for a single data point, the average distance between the data point and the remaining data points in the cluster to which the data point belongs is calculated as the first distance, and the average distance between the data point and all data points in the nearest remaining cluster is calculated as the second distance, and the difference between the second distance and the first distance is divided by the maximum value of the two distances to obtain the silhouette coefficient of the data point;

[0027] The average value of the silhouette coefficients of all data points in the division scheme corresponding to each candidate cluster number is calculated, and the cluster number with the maximum average value is selected as the corresponding division scheme; in the cluster set corresponding to the corresponding division scheme, the cluster containing the graph structure during the historical normal operation period is selected and merged into the baseline cluster set representing the normal mode of building energy consumption.

[0028] Further, generate the to-be-detected graph structure from the newly collected energy consumption data, calculate the minimum spectral distance between the to-be-detected graph structure and all graph structures in the baseline cluster, and determine that the to-be-detected graph structure is an abnormal graph structure if the minimum distance exceeds a preset threshold, including:

[0029] Based on the newly collected energy consumption data, generate the to-be-detected graph structure, and extract the feature value set of the standardized Laplacian matrix thereof to generate a corresponding spectral distribution;

[0030] For each cluster in the baseline cluster set, calculate the Battacharya distance between the spectral distribution of the to-be-detected graph structure and the spectral distribution of all historical graph structures in the cluster, and record the minimum spectral distance in each cluster;

[0031] Traverse the minimum spectral distances of all baseline clusters, and extract the smallest value as the global minimum spectral distance between the to-be-detected graph structure and the baseline cluster set;

[0032] Compare the global minimum spectral distance with a preset abnormality determination threshold, and determine that the to-be-detected graph structure is an abnormal energy consumption pattern if the distance exceeds the threshold, output an abnormality determination result, and mark the abnormal graph structure and its corresponding time period as an energy consumption abnormal event.

[0033] Further, perform node connection strength analysis on the abnormal graph structure, calculate the edge weight drop amplitude of each node and its adjacent nodes, and mark the sub-item partition table associated with the node with the largest edge weight loss as an abnormal source, including:

[0034] For each node in the abnormal graph structure, obtain the edge weight data in the historical normal mode, and calculate the absolute value of the difference between the real-time edge weight and the historical edge weight of the node and all adjacent nodes, and accumulate the difference values of all adjacent edges to obtain the connection loss of the node;

[0035] Normalize the connection loss of all nodes to generate a loss strength index in the interval [0, 1], wherein the node corresponding to the maximum loss value is marked as a candidate abnormal source;

[0036] Based on the edge weight fluctuation range statistically calculated from the historical normal data, set a dynamic determination threshold for the connection loss, and if the loss strength index of the candidate abnormal source exceeds the threshold, the sub-item partition table associated with the candidate abnormal source is located as an abnormal source.

[0037] Further, generate a detection report containing device maintenance prompts and energy efficiency optimization strategies according to the spatial distribution characteristics and energy consumption mutation timing characteristics of the abnormal source, including:

[0038] The spectral features of the abnormal source positioning result and the abnormal graph structure are processed, the spatial distribution density of the sub-item partition sub-table corresponding to the abnormal source and the timing fluctuation mode of the energy consumption mutation time period are analyzed, and the spatial-time correlation abnormal features are extracted; according to the spatial-time correlation abnormal features, in combination with the fault mode case library in the historical energy consumption event database, the similarity between the abnormal features and the known fault modes is matched, the abnormal cause type is inferred, including device failure, human operation abnormality or environmental interference;

[0039] Based on the abnormal cause type, a targeted optimization suggestion is generated: for the device failure type abnormality, a device repair list and priority sorting are generated, for the energy consumption mode mutation type abnormality, an energy efficiency parameter adjustment strategy is generated, and the spatial position information of the area where the abnormal source is located is associated; the optimization suggestion, the abnormal cause analysis and the associated data features are processed into a structured detection report.

[0040] In a second aspect, a commercial building energy consumption anomaly detection system based on a graph method comprises:

[0041] An acquisition module is configured to collect real-time energy consumption data from sub-item partition sub-tables in a building, generate multiple groups of energy consumption time series with timing characteristics, perform dynamic time warping calculation on the energy consumption time series, quantify the energy consumption time series similarity between each sub-item partition sub-table, and construct a graph structure based on the similarity, wherein the nodes are mapped to the sub-item partition sub-tables, and the edge weights represent the timing similarity between the corresponding nodes.

[0042] A comparison module is configured to generate a spectral distribution representing the topological features of the graph structure by calculating the eigenvalue set of the Laplacian matrix of the graph structure, and perform hierarchical clustering on multiple graph structures based on the Battacharyya distance between the spectral distributions to obtain energy consumption mode clustering results in different time periods; the clustering results are analyzed by the profile coefficient, the number of clusters corresponding to the maximum profile coefficient is selected as the corresponding division, and the cluster set containing historical normal data is defined as the baseline cluster; a to-be-detected graph structure is generated for newly collected energy consumption data, the minimum spectral distance between the to-be-detected graph structure and all graph structures in the baseline cluster is calculated, and if the minimum value exceeds a preset threshold, the to-be-detected graph structure is determined as an abnormal graph structure.

[0043] A processing module is configured to analyze the node connection strength of the abnormal graph structure, calculate the edge weight drop range of each node and its adjacent nodes, and mark the sub-item partition sub-table associated with the node with the largest edge weight loss as an abnormal source; according to the spatial distribution features and energy consumption mutation timing features of the abnormal source, a detection report containing device repair prompts and energy efficiency optimization strategies is generated.

[0044] In a third aspect, a computing device comprises:

[0045] One or more processors;

[0046] a memory device for storing one or more programs, when executed by the one or more processors, cause the one or more processors to implement the method.

[0047] In a fourth aspect, a computer-readable storage medium stores a program, which, when executed by a processor, implements the method.

[0048] The above scheme of the present application at least has the following beneficial effects:

[0049] The present application calculates the time series similarity of building sub-table energy consumption data by dynamic time warping algorithm, and converts these data into a graph structure model composed of nodes and edges. This graph structure can fully capture the complex dependence between different sub-tables, including time dynamics and regional correlations.

[0050] The present application proposes an unsupervised anomaly detection method, which identifies overall anomalies and point anomalies by analyzing the overall structure and local features of the graph. This method does not rely on labeled data, greatly reduces the dependence on human intervention, and improves the feasibility of practical application.

[0051] The hierarchical clustering method is used to generate baseline clusters of building energy consumption, which can represent the typical energy consumption patterns of the building at different time periods. Combined with the real-time generated graph structure, the present application supports multi-level anomaly analysis from a single sub-table to the overall building. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 is a flowchart of a commercial building energy consumption anomaly detection method based on a graph method provided by an embodiment of the present application.

[0053] Figure 2 is a schematic diagram of a commercial building energy consumption anomaly detection system based on a graph method provided by an embodiment of the present application.

[0054] Figure 3 is a schematic diagram of the working process of the building sub-table energy consumption anomaly detection method of the present application. DETAILED DESCRIPTION

[0055] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be accurately conveyed to those skilled in the art.

[0056] As Figure 1As shown, the embodiment of the present application proposes a commercial building energy consumption anomaly detection method based on a graph method, which comprises the following steps:

[0057] Step 11, collecting real-time energy consumption data from the sub-item partition sub-tables in the building to generate multiple groups of energy consumption time series with time sequence characteristics;

[0058] Step 12, performing dynamic time warping calculation on the energy consumption time series to quantify the energy consumption time sequence similarity between each sub-item partition sub-table, and constructing a graph structure based on the similarity, wherein the nodes are mapped to the sub-item partition sub-tables and the edge weights represent the time sequence similarity between the corresponding nodes;

[0059] Step 13, generating a spectral distribution representing the topological characteristics of the graph structure by calculating the eigenvalue set of the Laplacian matrix of the graph structure, and performing hierarchical clustering on multiple graph structures based on the Battacharyya distance between the spectral distributions to obtain energy consumption mode clustering results for different time periods;

[0060] Step 14, performing silhouette coefficient analysis on the clustering results to select the number of clusters corresponding to the maximum silhouette coefficient as the corresponding division, and defining the cluster set containing historical normal data as the baseline cluster;

[0061] Step 15, generating a to-be-detected graph structure for newly collected energy consumption data, calculating the minimum spectral distance between the to-be-detected graph structure and all graph structures in the baseline cluster, and determining it as an abnormal graph structure if the minimum value exceeds a predetermined threshold;

[0062] Step 16, performing node connection strength analysis on the abnormal graph structure, calculating the edge weight drop amplitude of each node and its adjacent nodes, and marking the sub-item partition sub-table associated with the node with the largest edge weight loss as the abnormal source;

[0063] Step 17, generating a detection report containing device maintenance prompts and energy efficiency optimization strategies based on the spatial distribution characteristics and energy consumption mutation time sequence characteristics of the abnormal source.

[0064] The multi-dimensional and high-resolution energy consumption data is obtained through the building internal sub-partition sub-table, covering the real-time power consumption of different areas / equipment, the DTW algorithm can align the non-synchronous sub-table data with time offset, accurately capture the similarity of the timing mode, convert the abstract energy consumption data into a node-edge graph structure, intuitively reflect the dependency between sub-tables, and the spectral distance quantifies the difference in graph structure through the eigenvalue of the Laplacian matrix of the graph, which can capture the overall topological features and avoid the one-sidedness of relying only on local edge weights. The silhouette coefficient automatically selects the optimal cluster number by evaluating the intra-class compactness and inter-class separation, avoiding the subjectivity of manual setting, and quickly identifies the overall energy consumption mode deviation from the normal range by comparing the spectral distance of the real-time graph and the baseline cluster, with low detection delay. By calculating the edge weight change of a single node, the specific abnormal equipment or area can be located, the abnormal detection result directly drives the optimization measures, and is connected with the building management system in real time, shortening the "detection-response" cycle, reducing energy waste and equipment failure loss.

[0065] In a preferred example of the present application, step 11 above comprises:

[0066] Step 111, according to the building area collecting energy consumption data, record the energy consumption data of each area through the building management system;

[0067] Step 112, the energy consumption data of each sub-partition sub-table collected is summarized to form a preliminary data set;

[0068] Step 113, by preprocessing the preliminary data set, the preprocessed energy consumption data is obtained.

[0069] In the embodiment of the present application, by locating the energy consumption source and automatically recording the energy consumption data of each area by the building management system, the tediousness and errors of data collection are reduced, the efficiency of data collection is improved, the data of each sub-partition sub-table is processed to form a comprehensive data set and facilitate data management and analysis, and the preliminary data set is preprocessed to improve data quality and enhance data availability, thereby improving the efficiency of energy consumption management.

[0070] In a specific embodiment of the present application, the specific implementation steps include:

[0071] Step 111, the building area collecting energy consumption data is determined, including each floor, room, equipment area, etc. in the building, and the corresponding data collection point is set in the building management system, and the system can record the energy consumption data of each area.

[0072] Step 112, the energy consumption data of each sub-partition sub-table is extracted from the building management system, the extracted data is preliminarily cleaned to remove repeated, erroneous or invalid data records, and the cleaned data is summarized according to the area, time and other dimensions to form a preliminary data set.

[0073] Step 113, standardize the preliminary data set, fill or delete the missing energy consumption data, detect and process the outliers in the data.

[0074] In a preferred example of the present application, the above step 12 comprises:

[0075] Step 121, based on the collected energy consumption data, identify all sub-tables of the building, wherein each sub-table corresponds to an independent energy consumption monitoring unit of a specific area or device, and extract the energy consumption time series of each sub-table;

[0076] Step 122, for any two sub-tables of the building, perform dynamic time warping calculation on the energy consumption time series of the two sub-tables: generate multiple matching paths through nonlinear time axis alignment, calculate the Euclidean distance square of adjacent matching elements in each path, accumulate the total cumulative distance of all local distance squares on the path, select the path corresponding to the minimum total cumulative distance, and take the square root of the total cumulative distance of the path as the dynamic time warping distance of the two sequences, which represents the similarity of the energy consumption time series of the two sub-tables;

[0077] Step 123, map the dynamic time warping distance to the interval (0, 1) through a linear normalization function to generate an edge weight value, wherein the smaller the distance, the closer the similarity score after mapping to 1, and the larger the distance, the closer the score to 0;

[0078] Step 124, taking the sub-tables as nodes and the normalized similarity scores as edge weights, construct a fully connected graph structure, wherein the weight value of the edge between any two nodes reflects the similarity of the corresponding sub-tables.

[0079] In an embodiment of the present application, by identifying the sub-tables, the analysis object is clearly analyzed, the analysis is improved in pertinence, the dynamic time warping algorithm is introduced, the dynamic changes in the energy consumption data are cleverly captured, the similarity of the time series is evaluated through the nonlinear alignment path, the dissimilarity is converted into the edge weight, the similarity of the energy consumption time series between the sub-tables is quantified, the graph structure is constructed, and the key elements are provided, so that the graph structure can accurately reflect the energy consumption relationship between the sub-tables, and finally the graph structure is constructed.

[0080] In a specific embodiment of the present application, the specific implementation steps include:

[0081] Step 121, obtain the preprocessed energy consumption data, analyze the structure of the energy consumption data, identify different sub-tables, each sub-table usually corresponds to a specific area or device in the building, and record the identification (such as name, number) of each sub-table and the area or device information monitored by each sub-table.

[0082] Step 122, through , compute the similarity between two time series, where, is a time series X and Y is the dynamic time warping distance between X and Y denote two time series, which are vectors consisting of a series of data points arranged in chronological order, is a warping path between X and Y is the minimum path among all possible warping paths is the sum of all point pairs on the path (i,j) is the square of the Euclidean distance between the th element in X and the i th element in Y j is the square of the Euclidean distance between the (i,j) th element in is an element in the path i is the index in the time series X j is the index in the time series Y For any two energy consumption time series of sub-tables, the corresponding warping path is calculated using the DTW algorithm, which is a path of non-linear alignment through one time axis, so that the elements of the sequences are matched in order. For all matching elements on the path, the square of the Euclidean distance is calculated, and the total cumulative distance of the path is obtained by accumulating all local distance squares on the path. Among all legal paths, the path with the smallest total cumulative distance is selected as the corresponding warping path. The square root of the total cumulative distance of the corresponding path is taken to obtain the dynamic time warping distance between the two sequences.

[0083] Step 123, determine the dissimilarity between any two sub-tables according to the dynamic time warping distance calculated in step 122, by , transforming the edge weight, where, is the dynamic time warping similarity between and is the dynamic time warping distance between and The original dynamic time warping distance value of the two sub-table energy consumption time series is mapped to the interval (0, 1) by , where ，d is the original DTW distance value,​​​​​​​​​ is the minimum value among all the calculated DTW distance values, is the maximum value among all the calculated DTW distance values s is the mapped similarity score, with a value range of (0, 1), and the formed intuitive similarity score is the transformed edge weight, representing the energy consumption relationship closeness between two sub-tables.

[0084] Step 124, each sub-table is regarded as a node in the graph structure, and the connection edges between nodes are established according to the edge weight calculated in step 123, which represents the energy consumption relationship closeness between two node sub-tables. All nodes and edges are combined to form a complete graph structure.

[0085] In the preferred example of the present application, the above step 13 comprises:

[0086] Step 131, for each time period graph structure, a standardized Laplacian matrix is constructed based on the edge weight between nodes, and eigenvalue decomposition is performed on it to extract a set of eigenvalues in ascending order to generate a spectral distribution representing the topological characteristics of the graph structure;

[0087] Step 132, for any two graph structure spectral distributions, calculate their Battacharyya distance: square the difference between the eigenvalues at corresponding positions in the two eigenvalue sets, and take the square root of the accumulated result, the obtained value is defined as the spectral distance between the two graph structures;

[0088] Step 133, based on the spectral distance matrix between all graph structures, use the agglomerative hierarchical clustering algorithm to gradually merge the graph structures or class clusters with the closest distance in a single linkage strategy until a preset number of initial class cluster set is formed;

[0089] Step 134, according to the tree diagram generated in the hierarchical clustering process, intercept the hierarchical division result that meets the preset merging threshold, and classify the graph structures of different time periods into corresponding energy consumption mode class clusters to form a set of clustering results with time continuity.

[0090] In the embodiment of the present application, the Laplace matrix can convert the topological relationship of nodes and edges into linear algebra structure, facilitate extraction of global features of the graph through matrix operation, capture the spectral characteristics of the graph through the eigenvalue sequence of the Laplace matrix, and accurately measure the global topological difference of different graph structures through the spectral distance. The eigenvalue corresponds to different scale features of the graph from large to small, and the spectral distance avoids one-sidedness by relying on only a single feature by comprehensively considering the difference of all eigenvalues, thereby improving the accuracy of clustering. The typical graph structure is selected as the initial clustering center through the spectral distance, so as to ensure that the baseline class cluster can cover the main normal mode of building energy consumption and avoid clustering deviation caused by random initialization. Through the iterative process of "initial center to extended class cluster", similar graph structures are gradually classified into the same class cluster, thereby forming a hierarchical baseline mode system.

[0091] In a specific embodiment of the present application, the specific implementation steps include:

[0092] Step 131, a plurality of graph structures are generated, each of which represents the relationship between different sub-table energy consumption time series in the building, a Laplace matrix is constructed, for each graph structure, a vertex set V and an edge set E are defined, a degree matrix D and an adjacency matrix A are constructed, and the Laplace matrix L is defined as .

[0093] Step 132, eigenvalue decomposition is performed on each Laplace matrix L to obtain eigenvalues, and the spectral distance is obtained through , wherein is the spectral distance between graphs and , and and represent two graph structures, and represent the first eigenvalue of the Laplace matrix of graphs and , respectively. k is the number of eigenvalues, which is usually equal to the number of vertices of the graph, is the sum of all eigenvalues from 1 to k , i is an index variable, is the eigenvalue of the Laplace matrix.

[0094] Step 133, the spectral distance between all graph structures is calculated through step 132 to form a spectral distance matrix, and a plurality of initial clustering centers are selected according to the spectral distance matrix, wherein the clustering center represents an initial baseline class cluster.

[0095] Step 134, for each remaining graph structure, assign it to the cluster center with the minimum spectral distance to it by calculating the spectral distance to all initial cluster centers in step 132, forming larger class clusters.

[0096] In a preferred embodiment of the present application, step 14 above comprises:

[0097] Step 141, based on the set of clustering results generated by hierarchical clustering, set a division scheme for each candidate class cluster number, and calculate the silhouette coefficient of all data points in each division scheme: for a single data point, calculate the average distance between it and the remaining data points in the class cluster to which it belongs as the first distance, and the average distance between it and all data points in the nearest neighboring class cluster as the second distance, divide the difference between the second distance and the first distance by the maximum value of the two distances to obtain the silhouette coefficient of the data point.

[0098] Step 142, calculate the average value of the silhouette coefficients of all data points in the division scheme corresponding to each candidate class cluster number, and select the class cluster number with the maximum average value as the corresponding division scheme; in the class cluster set corresponding to the corresponding division scheme, filter the class cluster containing the graph structure during the historical normal operation period, and merge it into the baseline class cluster set representing the normal mode of building energy consumption.

[0099] In an embodiment of the present application, the similarity of the graph structure in the class cluster is measured to ensure that the energy consumption patterns in the same class cluster have high consistency, and the difference between the graph structure and other class clusters is measured to avoid different patterns being mistakenly classified into the same class cluster, thereby improving the discrimination of the baseline class cluster. By quantifying the clustering rationality through mathematical formulas, the subjectivity of artificial experience judgment is avoided, the silhouette coefficient is calculated independently for each graph structure, the "fuzzy samples" at the edge of the class cluster can be identified, and if the number of class clusters is too large, it will cause fragmentation of the baseline mode and misjudge normal fluctuations as abnormal; the silhouette coefficient can suppress such situations through overall scoring, and if the number of class clusters is too small, it will mask the real difference and cause the abnormal detection sensitivity to decrease; the silhouette coefficient can automatically identify the pattern differentiation point. The calculation complexity of the silhouette coefficient has a linear relationship with the sample size, and is suitable for clustering and optimization of large-scale historical energy consumption data in commercial buildings.

[0100] In a specific embodiment of the present application, the specific implementation steps include:

[0101] Step 141, for each class cluster, calculate the silhouette coefficient by , wherein, and represent the average distance of the graph to other graphs in the same class cluster (intra-class closeness) and the average distance to the nearest neighboring class cluster (inter-class separation), respectively, i is an index variable, and the average value thereof is the overall silhouette coefficient under the class cluster number.

[0102] Step 142, after calculating the silhouette coefficients for different cluster numbers, the silhouette coefficient is an indicator for evaluating the clustering effect, which measures the tightness of any graph within its own cluster and the separation from the nearest neighbor cluster. The value range of the silhouette coefficient is wherein:

[0103] Close to 1 indicates that the graph is well classified, the internal cluster is dense, and it is significantly distinguished from other clusters.

[0104] Close to 0 indicates that the graph is located at the boundary of the cluster, and may belong to multiple clusters, and the clustering effect is not obvious.

[0105] Less than 0 indicates that the graph may be misclassified and belongs to the wrong cluster.

[0106] By averaging the silhouette coefficients of all graphs, the overall clustering quality score under the cluster number can be obtained. Then, traverse multiple possible cluster numbers, calculate the average silhouette coefficient corresponding to each cluster number, and draw the change curve. Finally, select the cluster number with the highest average silhouette coefficient as the corresponding scheme. The higher the silhouette coefficient, the tighter the data points within the cluster, and the more separated the data points between the clusters, i.e. the better the clustering effect.

[0107] In the preferred embodiment of the present application, step 15 of generating a to-be-detected graph structure from the newly collected energy consumption data, calculating the minimum spectral distance between the to-be-detected graph structure and all graph structures in the baseline cluster, and determining the to-be-detected graph structure as an abnormal graph structure if the minimum distance exceeds a preset threshold, comprises:

[0108] Step 151, based on the newly collected energy consumption data, generate a to-be-detected graph structure, and extract the eigenvalue set of the normalized Laplacian matrix of the to-be-detected graph structure to generate a corresponding spectral distribution; for each cluster in the baseline cluster set, calculate the Battacharyya distance between the spectral distribution of the to-be-detected graph structure and the spectral distribution of all historical graph structures in the cluster, and record the minimum spectral distance in each cluster; traverse the minimum spectral distances of all baseline clusters, extract the smallest value as the global minimum spectral distance between the to-be-detected graph structure and the baseline cluster set; compare the global minimum spectral distance with a preset abnormality determination threshold, if the distance exceeds the threshold, determine the to-be-detected graph structure as an abnormal energy consumption pattern, output the abnormality determination result, and mark the abnormal graph structure and its corresponding time period as an energy consumption anomaly event.

[0109] In a specific example, the specific implementation process of step 151 is as follows:

[0110] First, the newly collected energy consumption data is converted into a specific graph structure, which can reflect the inherent correlation and characteristics of the energy consumption data. This newly generated graph is referred to as the to-be-detected graph structure. Next, key mathematical features are extracted from the to-be-detected graph structure, that is, all eigenvalues of the normalized Laplacian matrix thereof are calculated, and these eigenvalues are collected together to form a spectral distribution that can describe the characteristics of the graph structure. This spectral distribution is like a "fingerprint" of the to-be-detected graph structure, representing its unique structural characteristics.

[0111] Then, the "fingerprint" of the to-be-detected graph structure is compared with each cluster in the baseline cluster set. The baseline cluster set contains multiple pre-established clusters representing normal energy consumption patterns, and each cluster contains a number of historical graph structures from past normal energy consumption data, each of which also has a corresponding spectral distribution.

[0112] For each baseline cluster, the spectral distribution of the to-be-detected graph structure is compared with the spectral distribution of each historical graph structure in the cluster one by one. The similarity measure method used here is the Bhattacharyya distance, which can measure the degree of similarity between two distributions. The smaller the distance, the more similar the two distributions, meaning the corresponding graph structures are closer. After calculating the Bhattacharyya distances between the to-be-detected graph structure and all historical graph structures in the cluster, only the smallest distance value needs to be recorded, as this smallest distance represents the closeness of the to-be-detected graph structure to the most similar historical graph structure in the cluster.

[0113] When all baseline clusters are processed, there will be a minimum spectral distance corresponding to each cluster. Next, the smallest value among all these minimum spectral distances needs to be found, which is the global minimum spectral distance of the to-be-detected graph structure to the entire baseline cluster set, reflecting the closeness of the to-be-detected graph structure to the most similar situation among all normal energy consumption patterns.

[0114] Finally, this global minimum spectral distance is compared with a pre-set abnormality judgment threshold. If the distance value exceeds the threshold, it means that the to-be-detected graph structure is not close enough to the most similar situation among all normal energy consumption patterns, and the difference is large. At this time, it can be determined that the to-be-detected graph structure represents an abnormal energy consumption pattern. Once an abnormality is determined, the abnormality judgment result is output, and the abnormal graph structure and the time period it corresponds to are marked as an energy consumption abnormality event for further analysis and processing.

[0115] In the preferred example of the present application, step 16 above includes:

[0116] Step 161: For each node in the abnormal graph structure, obtain its edge weight data in the historical normal mode, and calculate the absolute value of the difference between the real-time edge weight and the historical edge weight of the node and all its neighboring nodes. Accumulate the difference of all neighboring edges to obtain the connection loss of the node.

[0117] Step 162: Normalize the connection loss of all nodes to generate a loss intensity index in the interval [0, 1], where the node with the maximum loss value is marked as a candidate anomaly source.

[0118] Step 163: Based on the fluctuation range of edge weights in historical normal data statistics, set a dynamic threshold for determining the connection loss. If the loss intensity index of a candidate abnormal source exceeds the threshold, then the associated sub-table of the sub-item partition is identified as an abnormal source.

[0119] In this embodiment of the invention, by defining the node connection loss as the sum of the changes in edge weights with all neighboring nodes, the specific node with abnormal energy consumption can be accurately located by starting from the correlation between nodes. In commercial buildings, the energy consumption data of different areas or equipment are interconnected. The specific node with abnormality can be identified from the overall graph structure, achieving refined location of the abnormality. The connection loss is calculated by using the absolute value of the difference between historical edge weights and real-time edge weights. The data comparison method objectively reflects the changes in node connection relationships, avoiding subjective judgment. The analysis is based on the actual changes in data, accurately capturing subtle changes in node connection relationships, and effectively identifying energy consumption abnormalities caused by factors such as equipment failure and changes in operating modes. Thresholds are set based on historical data to flexibly adapt to different building characteristics and operating rules, enhance the versatility of the method, and mark abnormal nodes in a timely manner to achieve preventive maintenance and reduce the probability of equipment failure.

[0120] In a specific embodiment of the present invention, the specific implementation steps include:

[0121] Step 161: Based on the identified abnormal graph structure, for each node in the graph structure, through... Calculate the connection loss, where, f It is used to identify a specific node in a graph structure. index, r It is used for traversing nodes. The index of the neighboring nodes, yes t Time Node with neighboring nodes The weight of the edges between them. It is a historic moment node with neighboring nodes The weight of the edges between them. yes is a neighbor node of the node , is a set consisting of all neighbor nodes of the node , and the connection loss of each node is defined as the sum of the edge weight changes between the node and all neighbor nodes, wherein the neighbor node refers to a node connected to the node through a direct edge.

[0122] In step 162, according to each node in the abnormal graph structure, the historical edge weight and the real-time edge weight are obtained, the historical edge weight is the weight value of the edge between the node and the neighbor node in the normal energy consumption mode, and the real-time edge weight is the edge weight value corresponding to the current time, and the node connection loss is calculated through step.

[0123] In step 163, according to the historical collected energy consumption data and the corresponding graph structure information, the node connection loss is statistically analyzed, for example, the distribution of the node connection loss in different time periods and under different working conditions is analyzed, the fluctuation range of the connection loss under normal conditions is found out, a connection loss threshold is set, the connection loss calculated for each node is compared with the set threshold, and if the connection loss of a certain node exceeds the set threshold, it is determined that the connection relationship of the node has changed, and the node is marked as an abnormal node.

[0124] In the preferred example of the present application, the above step 17 comprises:

[0125] In step 171, the abnormal source positioning result and the spectral characteristics of the abnormal graph structure are processed, the spatial distribution density of the sub-item partition sub-table corresponding to the abnormal source and the timing fluctuation mode of the energy consumption mutation time period are analyzed, and the space-time correlation abnormal characteristics are extracted; according to the space-time correlation abnormal characteristics, the similarity between the abnormal characteristics and the known fault mode in the fault mode case library in the historical energy consumption event database is matched, the abnormal cause type is inferred, including device failure, human operation abnormality or environmental interference;

[0126] In step 172, based on the abnormal cause type, the targeted optimization suggestion is generated: for the device failure type abnormality, a device repair list and priority sorting are generated, for the energy consumption mode mutation type abnormality, an energy efficiency parameter adjustment strategy is generated, and the spatial position information of the area where the abnormal source is located is associated; the optimization suggestion, the abnormal cause analysis and the associated data characteristics are processed into a structured detection report.

[0127] In the embodiment of the present application, the overall anomaly is judged by combining the spectral distance with the confidence interval threshold, the point anomaly is located by using the node connection loss, the accurate anomaly diagnosis is realized, the anomaly is deeply analyzed, the accurate attribution is combined with the historical and real-time data, the problem is avoided to be solved on the surface, the optimization suggestion is generated based on the scientific analysis, the manager is assisted to make reasonable decision, the cost is reduced, the management intelligentization and automation level are enhanced through the real-time docking with the building management system, the anomaly processing is timely and efficient, the closed loop mode of detection, analysis, suggestion and execution is convenient for experience accumulation and strategy optimization, and the system is adapted to the building operation characteristics.

[0128] In a specific embodiment of the present application, the specific implementation steps include:

[0129] In step 171, the types of anomalies are identified by combining the historical energy consumption data and the corresponding graph structure information such as the energy consumption mode of different time periods in history, the node connection relationship and the like, and the current real-time collected energy consumption data and the graph structure, such as judging whether the energy consumption anomaly is caused by device failure, improper operation parameter setting, external environmental factor change or the like, and finding out the root cause of the anomaly.

[0130] In step 172, according to the analysis result of the anomaly cause, the corresponding optimization suggestion is made, for example, if the power consumption of the HVAC system of a certain area is abnormally high, the system will suggest adjusting the air conditioning operation parameter, and if some devices consume too much power, it can be suggested to arrange maintenance or replace the devices.

[0131] As shown in Figure 2 The embodiment of the present application also provides a commercial building energy consumption anomaly detection system based on a graph method, which includes:

[0132] The acquisition module 21 is used for collecting real-time energy consumption data from the sub-item and sub-zone tables in the building, generating a plurality of groups of energy consumption time sequences with time sequence characteristics; performing dynamic time warping calculation on the energy consumption time sequences, quantifying the energy consumption time sequence similarity between each sub-item and sub-zone table, and constructing a graph structure based on the similarity, wherein the nodes are mapped to the sub-item and sub-zone tables, and the edge weight represents the time sequence similarity between the corresponding nodes;

[0133] The comparison module 22 is used for generating a spectral distribution representing the topological characteristics of the graph structure by calculating the eigenvalue set of the Laplacian matrix of the graph structure, and performing hierarchical clustering on a plurality of graph structures based on the Batacharya distance between the spectral distributions to obtain the energy consumption mode clustering result of different time periods; performing profile coefficient analysis on the clustering result, screening the cluster number corresponding to the maximum profile coefficient as the corresponding division, and defining the cluster set containing the historical normal data as the baseline cluster; generating a to-be-detected graph structure for the newly collected energy consumption data, calculating the minimum spectral distance of the to-be-detected graph structure from all graph structures in the baseline cluster, and determining that the to-be-detected graph structure is an abnormal graph structure if the minimum spectral distance exceeds a preset threshold.

[0134] The processing module 23 is used for performing node connection strength analysis on the abnormal graph structure, calculating the edge weight drop range of each node and its adjacent node, marking the sub-item partition sub-table associated with the node with the largest edge weight loss as an abnormal source, and generating a detection report containing equipment maintenance prompts and energy efficiency optimization strategies according to the spatial distribution characteristics and energy consumption mutation timing characteristics of the abnormal source.

[0135] The above is the preferred embodiment of the present application, it should be noted that for those skilled in the art, without departing from the principles of the present application, can make a number of improvements and refinements, these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. A commercial building energy consumption anomaly detection method based on a graph method, characterized in that, The method comprises: Collecting real-time energy consumption data from sub-item sub-zone tables within the building to generate multiple groups of energy consumption time series with timing characteristics; Performing dynamic time warping calculation on the energy consumption time series to quantify the energy consumption time series similarity between each sub-item sub-zone table, and constructing a graph structure based on the similarity, wherein the nodes are mapped to the sub-item sub-zone tables, and the edge weight represents the timing similarity between the corresponding nodes; By calculating the eigenvalue set of the Laplacian matrix of the graph structure, a spectral distribution representing the topological features of the graph structure is generated, and hierarchical clustering is performed on multiple graph structures based on the Battacharyya distance between the spectral distributions to obtain energy consumption mode clustering results in different time periods; Performing silhouette coefficient analysis on the clustering results, selecting the cluster number corresponding to the maximum silhouette coefficient as the corresponding division, and defining the cluster set containing historical normal data as the baseline cluster; For newly collected energy consumption data, generate a to-be-detected graph structure, calculate the minimum spectral distance between it and all graph structures in the baseline cluster, and if the minimum value exceeds a preset threshold, determine it as an abnormal graph structure; Performing node connection strength analysis on the abnormal graph structure, calculating the edge weight drop amplitude of each node and its adjacent nodes, and marking the sub-item sub-zone table associated with the node with the maximum edge weight loss as the abnormal source; According to the spatial distribution characteristics and energy consumption mutation timing characteristics of the abnormal source, a detection report containing equipment maintenance prompts and energy efficiency optimization strategies is generated. 2.The graph-based method based commercial building energy consumption anomaly detection method of claim 1, wherein, Performing dynamic time warping calculation on the energy consumption time series to quantify the energy consumption time series similarity between each sub-item sub-zone table, and constructing a graph structure based on the similarity, wherein the nodes are mapped to the sub-item sub-zone tables, and the edge weight represents the timing similarity between the corresponding nodes, including: Based on the collected energy consumption data, all sub-item sub-zone tables within the building are identified, wherein each sub-table corresponds to an independent energy consumption monitoring unit of a specific area or device, and the energy consumption time series of each sub-table is extracted; Performing dynamic time warping calculation on the energy consumption time series of any two sub-item sub-zone tables: generating multiple matching paths through nonlinear time axis alignment, calculating the Euclidean distance square of adjacent matching elements in each path, accumulating the total cumulative distance of all local distance squares on the path to obtain the total cumulative distance, selecting the path corresponding to the minimum total cumulative distance, and taking the square root of the total cumulative distance of the path as the dynamic time warping distance of the two sequences, which represents the similarity of the energy consumption time series of the two sub-tables; Mapping the dynamic time warping distance to the (0, 1) interval through a linear normalization function to generate an edge weight value, wherein the smaller the distance, the closer the similarity score after mapping to 1, and the larger the distance, the closer the score to 0; Constructing a fully connected graph structure with sub-item sub-zone tables as nodes and normalized similarity scores as edge weights, wherein the weight value of the edge between any two nodes reflects the similarity of the corresponding sub-table energy consumption time series. 3.The graph-based method based commercial building energy consumption anomaly detection method of claim 2, wherein, By calculating the eigenvalue set of the Laplacian matrix of the graph structure, a spectral distribution representing the topological features of the graph structure is generated, and hierarchical clustering is performed on multiple graph structures based on the Battacharyya distance between the spectral distributions to obtain energy consumption mode clustering results in different time periods, including: For each time period graph structure, a standard Laplacian matrix is constructed based on the edge weights between nodes, and eigenvalue decomposition is performed on it to extract a set of eigenvalues in ascending order, generating a spectral distribution representing the topological characteristics of the graph structure; For any two spectral distributions of graph structures, calculate their Battacharyya distance: square the difference between the eigenvalues at corresponding positions in the two eigenvalue sets, and take the square root of the accumulated result, the resulting value is defined as the spectral distance between the two graph structures; Based on the spectral distance matrix between all graph structures, use the agglomerative hierarchical clustering algorithm to gradually merge the graph structures or clusters with the closest distance in a single link strategy until a preset number of initial cluster sets are formed; According to the tree diagram generated in the hierarchical clustering process, intercept the hierarchical division result that meets the preset merging threshold, and classify the graph structures of different time periods into corresponding energy consumption mode clusters to form a set of clustering results with time continuity. 4.The commercial building energy consumption anomaly detection method based on graph method according to claim 3, characterized in that, Perform silhouette coefficient analysis on the clustering results, select the cluster number corresponding to the maximum silhouette coefficient as the corresponding division, and define the cluster set containing historical normal data as the baseline cluster, including: Based on the clustering result set generated by hierarchical clustering, set a division scheme for each candidate cluster number, and calculate the silhouette coefficient of all data points in each division scheme: for a single data point, calculate its average distance with the remaining data points in the cluster as the first distance, and the average distance with all data points in the nearest remaining cluster as the second distance. The difference between the second distance and the first distance is divided by the maximum value of the two distances to obtain the silhouette coefficient of the data point; Calculate the average value of the silhouette coefficients of all data points in the division scheme corresponding to each candidate cluster number, and select the cluster number with the maximum average value as the corresponding division scheme. In the cluster set corresponding to the corresponding division scheme, filter the cluster containing the graph structure during the historical normal operation period, and merge it into the baseline cluster set representing the normal mode of building energy consumption. 5.The graph-based method based commercial building energy consumption anomaly detection method of claim 4, wherein, For the newly collected energy consumption data, generate the to-be-detected graph structure, calculate the minimum spectral distance between it and all graph structures in the baseline cluster, and if the minimum value exceeds the preset threshold, it is determined to be an abnormal graph structure, including: Based on the newly collected energy consumption data, generate the to-be-detected graph structure, and extract the eigenvalue set of its standard Laplacian matrix to generate the corresponding spectral distribution; For each cluster in the baseline cluster set, calculate the Battacharyya distance between the spectral distribution of the to-be-detected graph structure and the spectral distribution of all historical graph structures in the cluster, and record the minimum spectral distance in each cluster; Iterate through all the minimum spectral distances of the baseline clusters, extract the smallest value as the global minimum spectral distance between the to-be-detected graph structure and the baseline cluster set; Compare the global minimum spectral distance with the preset abnormal judgment threshold, if the distance exceeds the threshold, the to-be-detected graph structure is determined to be an abnormal energy consumption mode, output the abnormal judgment result, and mark the abnormal graph structure and its corresponding time period as an energy consumption anomaly event. 6.The commercial building energy consumption anomaly detection method based on graph method according to claim 5, characterized in that, The node connection strength analysis is performed on the abnormal graph structure, the edge weight drop range of each node and its adjacent nodes is calculated, the sub-item partition sub-table associated with the node with the maximum edge weight loss is marked as an abnormal source, and the node connection strength analysis includes: For each node in the abnormal graph structure, the edge weight data of the node in the historical normal mode is obtained, and the absolute value of the difference between the real-time edge weight of the node and all adjacent nodes and the historical edge weight is calculated, and the difference value of all adjacent edges is accumulated to obtain the connection loss of the node; The connection loss of all nodes is normalized to generate a loss strength index in the interval [0, 1], wherein the node corresponding to the maximum loss value is marked as a candidate abnormal source; Based on the edge weight fluctuation range of the historical normal data statistics, a dynamic determination threshold of the connection loss is set, and if the loss strength index of the candidate abnormal source exceeds the threshold, the sub-item partition sub-table associated with the candidate abnormal source is positioned as an abnormal source.

7. The graph-based method for commercial building energy consumption anomaly detection method according to claim 6, wherein, According to the spatial distribution characteristics and energy consumption mutation timing characteristics of the abnormal source, a detection report containing device maintenance prompts and energy efficiency optimization strategies is generated, including: Processing the abnormal source positioning result and the spectral characteristics of the abnormal graph structure, analyzing the spatial distribution density of the sub-item partition sub-table corresponding to the abnormal source and the timing fluctuation mode of the energy consumption mutation time period, extracting the space-time correlation abnormal characteristics; according to the space-time correlation abnormal characteristics, combining the fault mode case library in the historical energy consumption event database, matching the similarity between the abnormal characteristics and the known fault mode, deducing the abnormal cause type, including device failure, human operation abnormality or environmental disturbance; Based on the abnormal cause type, a targeted optimization suggestion is generated: for the device failure type abnormality, a device maintenance list and priority sorting are generated, for the energy consumption mode mutation type abnormality, an energy efficiency parameter adjustment strategy is generated, and the spatial position information of the area where the abnormal source is located is associated; the optimization suggestion, abnormal cause analysis and associated data characteristics are processed into a structured detection report.

8. A graph-based method for commercial building energy consumption anomaly detection system, the system implements the method of any one of claims 1 to 7, characterized in that, Including: An acquisition module is configured to collect real-time energy consumption data from sub-item partition sub-tables in a building to generate multiple groups of energy consumption time series with timing characteristics; The energy consumption time series is subjected to dynamic time warping calculation to quantify the energy consumption time sequence similarity between each sub-item partition sub-table, and a graph structure is constructed based on the similarity, wherein the nodes are mapped to the sub-item partition sub-tables, and the edge weights represent the timing similarity between the corresponding nodes; A comparison module is configured to generate spectral distribution representing the topological characteristics of the graph structure by calculating the eigenvalue set of the Laplacian matrix of the graph structure, and perform hierarchical clustering on multiple graph structures based on the Batacharya distance between the spectral distributions to obtain energy consumption mode clustering results in different time periods; the clustering results are subjected to contour coefficient analysis, the number of clusters corresponding to the maximum contour coefficient is selected as the corresponding division, and the cluster set containing historical normal data is defined as a baseline cluster; a to-be-detected graph structure is generated for newly collected energy consumption data, the minimum spectral distance between the to-be-detected graph structure and all graph structures in the baseline cluster is calculated, and if the minimum value exceeds a preset threshold, the to-be-detected graph structure is determined as an abnormal graph structure. The processing module is configured to perform node connection strength analysis on the abnormal graph structure, calculate a descending amplitude of an edge weight of each node and its adjacent node, and mark a sub-item sub-table associated with a node with a largest edge weight loss as an abnormal source. According to the spatial distribution characteristics and the energy consumption mutation timing characteristics of the abnormal source, a detection report containing equipment maintenance prompts and energy efficiency optimization strategies is generated.

9. A computing device, comprising: The method comprises the steps of: one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, so that the one or more processors implement the method as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a program which, when executed by a processor, implements the method as claimed in any one of claims 1 to 7.

Images