Multi-model fusion electricity charge abnormity identification method, device and equipment

Abstract

The invention provides a multi-model fusion electricity charge anomaly identification method, device and equipment, and relates to the technical field of machine learning. The method comprises the following steps: acquiring electric charge data of each user; an LSTM model, a graph neural network model and an isolated forest algorithm model are respectively used to carry out abnormity identification on the electricity charge data, and abnormity identification results corresponding to different models of each user are obtained; for each user, based on the user attribute of the user, determining the abnormal weight corresponding to the abnormal recognition result of the user under different models; and based on the abnormal weight, carrying out weighted fusion on the abnormal identification results of the user under different models to obtain an electricity charge abnormal identification result corresponding to the user. According to the invention, the electricity charge abnormity identification precision can be improved, and false alarm and missing detection are avoided.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method, apparatus and device for identifying anomalies in electricity charges using multi-model fusion. Background Technology

[0002] In the operation and management of power systems, identifying abnormal electricity bills is a crucial step in ensuring operational efficiency, detecting metering faults, and even preventing electricity theft. Currently, the industry mainly relies on rule-based audit engines to identify abnormal electricity bills. This technology typically screens abnormal electricity bills by predefining a series of static thresholds and logical rules (e.g., a sudden change in electricity consumption exceeding a certain fixed percentage, or an excessive deviation in electricity bill amount from the historical average).

[0003] However, with the continuous expansion of the user base and the increasing complexity of electricity consumption behavior, the aforementioned fixed-rule-based technical solutions have revealed significant limitations. On the one hand, fixed rules lack the ability to learn and generalize to complex, nonlinear anomaly patterns, leading to missed detections of complex anomalies. On the other hand, fixed rules cannot dynamically adapt to the inherent differences among different user types (such as residential, commercial, and industrial users), resulting in false alarms for normal users with unique behaviors. In summary, existing electricity bill anomaly identification methods suffer from problems such as false alarms and missed detections. Summary of the Invention

[0004] This invention provides a multi-model fusion method, apparatus, and device for identifying electricity bill anomalies, in order to solve the problems of false alarms and missed detections in existing electricity bill anomaly identification methods.

[0005] In a first aspect, embodiments of the present invention provide a multi-model fusion method for identifying electricity fee anomalies, including: Obtain electricity bill data for each user; The electricity bill data was anomaly identified using the Long Short-Term Memory (LSTM) network model, the graph neural network model, and the isolated forest algorithm model, respectively, to obtain the anomaly identification results for each user under different models. For each user, based on the user's attributes, determine the anomaly weight corresponding to the anomaly identification results of the user under different models; Based on the aforementioned anomaly weights, the anomaly identification results for the user under different models are weighted and fused to obtain the corresponding electricity bill anomaly identification result for the user.

[0006] Optionally, the user attributes include: user type, user stability, and user group size; For each user, based on the user's attributes, determine the anomaly weight corresponding to the anomaly identification results of that user under different models, including: For each user, based on the user type, determine the basic weights corresponding to the anomaly identification results of that user under different models; Based on the user stability and the user group size, the base weights are adjusted to obtain the adjusted weights, and the adjusted weights are normalized to obtain the anomaly weights corresponding to the anomaly identification results of the user under different models.

[0007] Optionally, the user types include: residential users, commercial users, and industrial users; For each user, based on user type, determine the basic weights corresponding to the anomaly detection results of that user under different models, including: For each user, if the user is a resident user, the base weight corresponding to the anomaly identification result of the user under the LSTM model is determined to be greater than the base weight corresponding to the anomaly identification result of the user under the graph neural network, and also greater than the base weight corresponding to the anomaly identification result of the user under the isolated forest algorithm network. If the user is a commercial user, then the base weight corresponding to the anomaly detection result of the user under the graph neural network is greater than the base weight corresponding to the anomaly detection result of the user under the LSTM model; and the base weight corresponding to the anomaly detection result of the user under the LSTM model is greater than the base weight corresponding to the anomaly detection result of the user under the isolated forest algorithm network. If the user is an industrial user, then the base weights corresponding to the anomaly identification results of the user under different models are determined to be the same.

[0008] Optionally, based on the user stability, the base weights are adjusted, including: When the user stability index is higher than the first stability threshold, the base weight of the user's anomaly identification result under the LSTM model is reduced, and the base weight of the user's anomaly identification result under the Isolation Forest algorithm model is increased; the lower the user stability index, the higher the user stability. When the user stability index is lower than the first stability threshold but higher than the second stability threshold, the basic weight is maintained. When the user's stability index is lower than the second stability threshold, increase the base weight of the user's anomaly identification result under the LSTM model and decrease the base weight of the user's anomaly identification result under the isolated forest algorithm model.

[0009] Optionally, the base weights can be adjusted based on the size of the user group, including: When the user group size index is higher than the first size threshold, the base weight corresponding to the anomaly identification result of the user under the graph neural network model is increased; the higher the user group size index, the larger the user group size. When the user group size indicator is lower than the first size threshold but higher than the second size threshold, the basic weight is maintained. When the user group size index is lower than the second size threshold, the base weight corresponding to the abnormal identification result of the user in the graph neural network model is reduced.

[0010] Optionally, the method for determining the user stability index includes: Obtain the user's electricity consumption data within the target time period, and calculate the average value and standard deviation of the electricity consumption data; The ratio of the standard deviation to the mean is determined as the user stability index.

[0011] Optionally, the method for determining the user group size indicator includes: For each user, calculate the similarity of electricity consumption behavior, electricity consumption attributes, and load characteristics between that user and each of the remaining users. The similarity of electricity consumption behavior, similarity of electricity consumption attributes, and similarity of load characteristics are weighted and summed to obtain the comprehensive similarity between the user and each of the remaining users; The number of remaining users whose overall similarity is greater than a set threshold is determined, and the number is normalized to obtain the user group size index.

[0012] Optionally, the method further includes: For each user, a comprehensive similarity sequence corresponding to that user is determined, and in the comprehensive similarity sequence, the comprehensive similarity corresponding to a set quantile is determined as the set threshold; The comprehensive similarity sequence contains all the comprehensive similarities corresponding to the user, and the comprehensive similarities are arranged in ascending order.

[0013] Secondly, embodiments of the present invention provide a multi-model fusion-based electricity fee anomaly identification device, comprising: The acquisition module is used to acquire electricity bill data for each user; The identification module is used to identify anomalies in the electricity bill data using LSTM model, graph neural network model and isolated forest algorithm model respectively, and to obtain the anomaly identification results for each user under different models. The weighted module is used for: For each user, based on the user's attributes, determine the anomaly weight corresponding to the anomaly identification results of the user under different models; Based on the aforementioned anomaly weights, the anomaly identification results for the user under different models are weighted and fused to obtain the corresponding electricity bill anomaly identification result for the user.

[0014] Thirdly, embodiments of the present invention provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect or any possible implementation thereof.

[0015] This invention achieves comprehensive coverage of electricity bill anomaly identification from three different but complementary dimensions: the LSTM model identifies anomalous users whose electricity bills significantly deviate from their historical electricity consumption patterns from a temporal continuity perspective; the graph neural network identifies anomalous users whose electricity bills significantly deviate from those of similar user groups from a spatial correlation perspective; and the isolated forest algorithm identifies users with unknown anomaly patterns in the global data feature space from a data feature distribution perspective. This multi-perspective collaborative detection mechanism ensures that temporal anomalies, group anomalies, and novel unknown anomalies can all be effectively identified, significantly improving the coverage and detection accuracy of anomaly identification and fundamentally avoiding the missed detection problem that may be caused by a single detection perspective.

[0016] Furthermore, this embodiment of the invention introduces a dynamic weight allocation mechanism that adapts to user attributes. Based on each user's specific attributes (such as user type, electricity usage stability, group size, etc.), the weight ratio of each model in the final decision is intelligently adjusted. For example, for residential users with strong electricity usage patterns, the system assigns a higher decision weight to the LSTM model; for commercial users with strong group characteristics, the recognition results of the graph neural network are prioritized. This personalized model fusion strategy enables the system to accurately adapt to the inherent characteristics of different user groups, improving the accuracy of anomaly identification while effectively reducing the risk of false alarms, achieving the technical effect of accurate identification. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the implementation of the multi-model fusion method for identifying abnormal electricity charges provided in this embodiment of the invention. Figure 2 This is a schematic diagram of the structure of the multi-model fusion electricity fee anomaly identification device provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0018] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0019] As the user base continues to expand and electricity consumption behavior becomes increasingly complex, abnormal electricity bill patterns are also gradually increasing. Existing rule-based electricity bill anomaly identification schemes are prone to missing complex anomalies. Furthermore, rule-based electricity bill anomaly identification schemes cannot dynamically adapt to the inherent differences of different user types (such as residential, commercial, and industrial users), which can lead to false alarms for normal users with unique behaviors.

[0020] To effectively avoid false alarms and missed detections in the electricity bill anomaly identification process, this invention utilizes LSTM, graph neural network, and isolated forest algorithms to identify electricity bill anomalies. These methods significantly improve the coverage and detection accuracy of anomaly identification from the perspectives of temporal continuity, spatial correlation, and data feature distribution. Furthermore, this invention dynamically assigns weights to the anomaly identification results of each model based on user attributes, accurately adapting to the inherent characteristics of different user groups and thus reducing the probability of false alarms.

[0021] See Figure 1 The flowchart illustrating the implementation of the multi-model fusion electricity bill anomaly identification method provided in this embodiment of the invention is described in detail below: Step 101: Obtain electricity bill data for each user.

[0022] Here, the electricity billing data can include: electricity consumption and electricity cost for each billing cycle within the current time period. A rolling window can be used to determine the current time period. The rolling window can be 12 months in length, defining the most recent 12 months as the current time period. The billing cycle can be one month.

[0023] Based on the obtained electricity consumption and electricity cost, this embodiment of the invention can further extract the monthly average electricity consumption, the monthly average electricity cost, and the electricity consumption variation coefficient for subsequent model prediction.

[0024] The coefficient of variation for electricity consumption is the ratio between the standard deviation of electricity consumption and the average value of electricity consumption during the current period.

[0025] In addition, embodiments of the present invention can also obtain load data of each user, peak-hour electricity consumption and valley-hour electricity consumption of each user, as well as user type of each user, for subsequent model prediction.

[0026] This invention allows for the determination of the average load factor for each user based on their load data. The average load factor is the ratio of the monthly average load to the monthly maximum load.

[0027] The embodiments of the present invention can determine the peak-valley electricity consumption ratio based on the ratio of peak-hour electricity consumption to valley-hour electricity consumption for each user.

[0028] In this embodiment of the invention, user type coding can also be used to represent different user types. For example, user types may include residential users, commercial users, and industrial users. Specifically, the user type coding for residential users can be [10 0]. The user type coding for commercial users can be [0 1 0]. The user type coding for residential users can be [0 0 1].

[0029] The aforementioned monthly average electricity consumption and monthly average electricity bill can be used as characteristics of electricity consumption scale. The coefficient of variation in electricity consumption can be used as a characteristic of electricity consumption stability. The average load factor can be used as a characteristic of load characteristics. The peak-valley electricity consumption ratio can be used as a characteristic of time distribution. The user type code can be used as a characteristic of user type. These characteristics—electricity consumption scale, electricity consumption stability, load characteristics, time distribution, and user type—can reflect a user's electricity consumption characteristics from different dimensions and are used for subsequent identification of electricity bill anomalies.

[0030] Step 102: Use the LSTM model, graph neural network model and isolated forest algorithm model respectively to identify anomalies in electricity bill data and obtain the anomaly identification results for each user under different models.

[0031] For the LSTM model, the user's electricity bill sequence in the previous historical period is input into the LSTM model, and the predicted electricity bill sequence for the user in the current period is output.

[0032] Obtain the user's actual electricity bill sequence for the current time period, calculate the relative deviation rate between the actual electricity bill sequence and the predicted electricity bill sequence, and use this relative deviation rate as the user's anomaly score, i.e., the anomaly identification result.

[0033] The relative deviation rate is calculated as follows: The absolute deviation between the actual electricity cost series and the predicted electricity cost series is calculated separately to obtain the absolute deviation series. Then, the average value of the absolute deviations in the absolute deviation series is calculated to obtain the average absolute deviation. Subsequently, the average value of the actual electricity cost in the actual electricity cost series is calculated. Finally, the average value of the average absolute deviation and the actual electricity cost is determined as the relative deviation rate.

[0034] Here, the electricity billing amount sequence for the previous historical period includes the electricity billing amount for each billing cycle within the previous historical period. The predicted electricity billing amount sequence for the current period includes the predicted electricity billing amount for each billing cycle within the current period. The actual electricity billing amount sequence for the current period includes the actual electricity billing amount for each billing cycle within the current period. For example, the current period can be the most recent 12 months. The previous historical period can be the 12 months preceding the current period.

[0035] For the graph neural network model, each user is treated as a node, and edge connections are established based on user similarity to obtain a user relationship graph. The electricity consumption characteristics of each user are used as the initial features of each node, and the initial features and user relationship graph are input into the graph neural network model for message passing and feature aggregation to obtain the embedded representation of each node.

[0036] For each user, the comprehensive similarity between that user and each of the remaining users can be calculated. From the remaining users, users with a comprehensive similarity greater than a set threshold are identified as the user's neighbors, thereby establishing edge connections between users and obtaining a user relationship graph.

[0037] Here, the electricity consumption characteristics of each user are the electricity consumption scale characteristics, electricity consumption stability characteristics, load characteristic characteristics, time distribution characteristics, and user type characteristics of each user.

[0038] Based on the determined embedding representation of each node, for each node, the K most similar neighboring nodes among its directly connected neighboring nodes are identified, and the average of the embedding representations of these K neighboring nodes is calculated. Finally, the Euclidean distance between the node's embedding representation and this average value is calculated, and after normalization, this Euclidean distance is used as the user's anomaly score, i.e., the anomaly identification result. Here, K can take any integer value between 5 and 15.

[0039] For the Isolation Forest algorithm model, the monthly average electricity consumption, monthly average electricity bill, electricity consumption variation coefficient, average load factor, peak-valley electricity consumption ratio, and user type code of each user can be used as user features input into the Isolation Forest algorithm model. The Isolation Forest algorithm outputs anomaly scores for each user. However, considering that the anomaly scores output by the Isolation Forest algorithm model are between [-0.5, 0.5], this embodiment of the invention can further convert the anomaly scores output by the Isolation Forest algorithm model to [0, 1] to finally obtain the final anomaly scores for each user.

[0040] Step 103: For each user, based on the user's attributes, determine the anomaly weight corresponding to the anomaly identification results of the user under different models.

[0041] Here, user attributes can include: user type, user stability, and user group size. User type can include residential users, commercial users, and industrial users. Considering that different user types have different electricity consumption behavior characteristics, their potential anomaly patterns also differ. Therefore, this embodiment of the invention can dynamically adjust the anomaly weights corresponding to the anomaly identification results of each model based on user type, thereby achieving targeted detection for different types of users.

[0042] User stability reflects the stability of a user's electricity consumption behavior. User group size reflects the size of a group of users with similar electricity consumption behavior.

[0043] Here, the LSTM model focuses on learning individual user time-series patterns and is more suitable for stable users. Graph neural network models excel at discovering relative anomalies within a group and are more suitable for larger user groups. The Isolation Forest algorithm focuses on identifying users with unknown anomaly patterns in the global data feature space from the data distribution dimension.

[0044] Step 104: Based on the anomaly weight, the anomaly identification results of the user under different models are weighted and fused to obtain the electricity bill anomaly identification result corresponding to the user.

[0045] For each user, the anomaly identification results (i.e. anomaly scores) of that user under different models can be obtained separately, and the anomaly scores can be weighted and summed to obtain the final anomaly identification result (i.e. anomaly score) of that user.

[0046] Compared to existing technologies, this invention achieves comprehensive coverage of electricity bill anomaly identification from three different but complementary dimensions: the LSTM model identifies anomalous users whose electricity bills significantly deviate from their historical electricity consumption patterns from a temporal continuity perspective; the graph neural network identifies anomalous users whose electricity bills significantly deviate from those of similar user groups from a spatial correlation perspective; and the isolated forest algorithm identifies users with unknown anomaly patterns in the global data feature space from a data feature distribution perspective. This multi-perspective collaborative detection mechanism ensures that temporal anomalies, group anomalies, and novel unknown anomalies can all be effectively identified, significantly improving the coverage and detection accuracy of anomaly identification and fundamentally avoiding the missed detection problem that may be caused by a single detection perspective.

[0047] Furthermore, this embodiment of the invention introduces a dynamic weight allocation mechanism that adapts to user attributes. Based on each user's specific attributes (such as user type, electricity usage stability, group size, etc.), the weight ratio of each model in the final decision is intelligently adjusted. For example, for residential users with strong electricity usage patterns, the system assigns a higher decision weight to the LSTM model; for commercial users with strong group characteristics, the recognition results of the graph neural network are prioritized. This personalized model fusion strategy enables the system to accurately adapt to the inherent characteristics of different user groups, improving the accuracy of anomaly identification while effectively reducing the risk of false alarms, achieving the technical effect of accurate identification.

[0048] The method for determining abnormal weights will be explained in detail below.

[0049] Specifically, for each user, based on the user type, the basic weights corresponding to the anomaly identification results of the user under different models are determined; then, based on the user stability and the size of the user group, the basic weights are adjusted to obtain the adjusted weights, and the adjusted weights are normalized to obtain the anomaly weights corresponding to the anomaly identification results of the user under different models.

[0050] Here, user types include: residential users, commercial users, and industrial users.

[0051] For each user, if the user is a resident user, the base weight corresponding to the anomaly identification result of the user under the LSTM model is determined to be greater than the base weight corresponding to the anomaly identification result of the user under the graph neural network, and also greater than the base weight corresponding to the anomaly identification result of the user under the isolated forest algorithm network.

[0052] In other words, if the user is a resident user, then the base weight corresponding to the anomaly identification result of the user under the LSTM model is the largest.

[0053] If the user is a commercial user, then the base weight corresponding to the anomaly detection result of the user under the graph neural network is greater than the base weight corresponding to the anomaly detection result of the user under the LSTM model; and the base weight corresponding to the anomaly detection result of the user under the LSTM model is greater than the base weight corresponding to the anomaly detection result of the user under the isolated forest algorithm network.

[0054] In other words, if the user is a business user, the user's anomaly detection result under the graph neural network model has the highest base weight, followed by the anomaly detection result under the LSTM model, and the anomaly detection result under the isolated forest algorithm model has the lowest base weight.

[0055] If the user is an industrial user, then the base weights for the anomaly detection results under different models are the same. That is, the base weights for the anomaly detection results under each model are the same for industrial users.

[0056] Different user types exhibit different electricity consumption behavior characteristics. For residential users, their electricity consumption behavior is driven by personal lifestyle habits and exhibits strong temporal regularity. LSTM models excel at learning individual temporal patterns; therefore, for residential users, this embodiment of the invention sets the base weights of the LSTM model to be the largest.

[0057] For commercial users, their electricity consumption behavior is significantly influenced by group factors such as competition among peers and regional economic conditions. The electricity consumption behavior of commercial users within the same business district and industry shows a high degree of consistency. Graph neural networks excel at detecting relative anomalies within a group; therefore, for commercial users, this embodiment of the invention sets the base weights of the graph neural network model to be the largest. Simultaneously, considering that commercial users also exhibit certain temporal patterns, this embodiment of the invention sets the base weights of the LSTM model to be the second largest.

[0058] For industrial users, their electricity consumption behavior is constrained by both production plans (i.e., time sequence) and industry standards (i.e., group size), and their abnormal electricity bill patterns are quite complex. Therefore, for industrial users, the embodiments of this invention set the same basic weights for the three models.

[0059] For example, for residential users, the base weights for the LSTM model can be 0.6, the base weights for the graph neural network model can be 0.2, and the base weights for the isolated forest algorithm model can be 0.2.

[0060] For commercial users, the base weights for the LSTM model can be 0.3, the base weights for the graph neural network model can be 0.5, and the base weights for the isolated forest algorithm model can be 0.2.

[0061] For industrial users, the base weights for the LSTM model, graph neural network model, and isolated forest algorithm model can all be set to 0.33. Considering the strong temporal nature of industrial users' production plans, the base weights for the LSTM model can be appropriately increased. For example, the base weights for the LSTM model could be 0.4, the graph neural network model 0.3, and the isolated forest algorithm model 0.3.

[0062] In this embodiment of the invention, the basic weights are adjusted based on user type and user stability and user group size to obtain the final abnormal weights.

[0063] In some embodiments, the base weights can be adjusted based on user stability, specifically: When the user's stability index is higher than the first stability threshold, the base weight corresponding to the user's anomaly identification result under the LSTM model is reduced, and the base weight corresponding to the user's anomaly identification result under the Isolation Forest algorithm model is increased.

[0064] When the user stability metric is lower than the first stability threshold but higher than the second stability threshold, the base weight is maintained. When the user's stability index is lower than the second stability threshold, increase the base weight of the user's anomaly identification results under the LSTM model and decrease the base weight of the user's anomaly identification results under the Isolation Forest algorithm model.

[0065] Here, the stability index is the aforementioned coefficient of variation in electricity consumption. Its specific calculation method is as follows: Obtain user electricity consumption data within the target time period and calculate the corresponding average and standard deviation. The ratio of the standard deviation to the average is used as the user stability index. A rolling window can be used to define the target time period, ensuring that electricity consumption data from the most recent 12 months is consistently used to calculate the coefficient of variation. The rolling window can have a duration of 12 months.

[0066] Here, the lower the user stability index, the higher the user stability. When the user stability index is higher than the first stability threshold, it indicates that the user is a low-stability user, meaning that the user's electricity consumption pattern is irregular. In this case, the base weights of the LSTM model can be reduced, and the base weights of the Isolation Forest model can be increased to avoid false alarms. Here, the amount of reduction in the base weights of the LSTM model is equal to the amount of increase in the base weights of the Isolation Forest model. For example, the increase = decrease = 0.15.

[0067] When a user's stability index falls below the second stability threshold, it indicates that the user is a highly stable user, meaning their electricity consumption pattern is highly regular with minimal fluctuations. In this case, the base weights of the LSTM model can be increased, while the base weights of the Isolation Forest model can be decreased to avoid false alarms. Here, the increase in the base weights of the LSTM model is equal to the decrease in the base weights of the Isolation Forest model. For example, the increase = decrease = 0.15.

[0068] When a user's stability index is lower than the first stability threshold but higher than the second stability threshold, it indicates that the user is a medium-stability user, meaning that the power consumption pattern has a certain regularity and there are reasonable fluctuations. In this case, the basic weight can be maintained.

[0069] The specific values ​​of the first stability threshold and the second stability threshold in this embodiment of the invention can be determined according to the actual situation. For example, the first stability threshold can be 0.6 and the second stability threshold can be 0.3.

[0070] In some embodiments, the base weight can be adjusted based on the size of the user group, specifically: When the user group size index exceeds the first size threshold, increase the base weight corresponding to the anomaly identification results of the user under the graph neural network model.

[0071] When the user group size metric is below the first size threshold but above the second size threshold, the basic weight is maintained. When the user group size index is lower than the second size threshold, the base weight corresponding to the abnormal identification result of the user in the graph neural network model is reduced.

[0072] Here, the higher the user group size index, the larger the user group size. When the user group size index is higher than the first size threshold, it indicates that there are many users similar to that user, meaning that the user's user group size is very large. Graph neural networks are good at discovering abnormal users within a user group. Therefore, this embodiment of the invention increases the base weights corresponding to the anomaly identification results of users under the graph neural network model. For example, the increase in the base weights corresponding to the graph neural network can be 0.2.

[0073] Correspondingly, when the user group size index is below the second size threshold, it indicates that there are very few users similar to that user, meaning the user's user group size is small. In this case, the base weight corresponding to the user's anomaly identification result in the graph neural network model can be reduced. For example, the reduction in the base weight corresponding to the graph neural network can be 0.2.

[0074] Here, the method for determining the user group size metric is as follows: For each user, the similarity of electricity consumption behavior, electricity consumption attributes, and load characteristics between the user and each of the remaining users is calculated. Then, the similarity of electricity consumption behavior, electricity consumption attributes, and load characteristics is weighted and summed to obtain the comprehensive similarity between the user and each of the remaining users. Finally, the number of remaining users whose comprehensive similarity is greater than a set threshold is determined, and the number is normalized to obtain the user group size index.

[0075] In calculating the similarity of electricity consumption behavior between any two users, the similarity between the average monthly electricity consumption of the two users can be calculated and used as the similarity of electricity consumption behavior.

[0076] When calculating the similarity of electricity usage attributes between any two users, it can be checked whether the user types of the two users are the same. If the user types are the same, the similarity of electricity usage attributes is determined to be 1. If the user types are different, the similarity of electricity usage attributes is determined to be 0.

[0077] When calculating the load characteristic similarity between any two users, the similarity between the average load rates of the two users can be calculated and used as the load characteristic similarity.

[0078] When weighting and summing the similarities of electricity consumption behavior, electricity consumption attributes, and load characteristics, the weights of these three similarities can be determined based on the actual situation. For example, the weight of electricity consumption behavior similarity can be 0.5, the weight of electricity consumption attributes similarity can be 0.3, and the weight of load characteristics similarity can be 0.2.

[0079] For each user, calculate the overall similarity between that user and the remaining users. From all the remaining users, identify the target users whose overall similarity is greater than a set threshold. These target users are the neighboring users connected to that user in the user relationship graph.

[0080] In this embodiment of the invention, the number of neighboring users connected to the user is normalized, and the normalized result is used as an indicator of the user's user group size.

[0081] The method for determining the threshold is as follows: For each user, a comprehensive similarity sequence is determined. Here, the comprehensive similarity sequence contains all the comprehensive similarities corresponding to the user, and the comprehensive similarities are arranged in ascending order.

[0082] In this embodiment of the invention, the comprehensive similarity corresponding to a set quantile in the comprehensive similarity sequence is determined as a set threshold. For example, the set quantile can be the 75th quantile.

[0083] In this embodiment of the invention, based on the basic weight determined by user type, the basic weight can first be adjusted based on user stability to obtain the adjusted basic weight. Then, the adjusted basic weight is further adjusted based on the user group size to obtain the final adjustment result. Finally, the adjustment result is normalized to obtain the final anomaly weight.

[0084] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0085] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0086] Figure 2 A schematic diagram of the multi-model fusion electricity fee anomaly identification device provided in an embodiment of the present invention is shown. For ease of explanation, only the parts related to the embodiment of the present invention are shown, and are described in detail below: like Figure 2As shown, the multi-model fusion electricity fee anomaly identification device 2 includes: an acquisition module 21, an identification module 22, and a weighting module 23.

[0087] Module 21 is used to acquire electricity bill data for each user; The identification module 22 is used to identify anomalies in electricity bill data using LSTM model, graph neural network model and isolated forest algorithm model respectively, and to obtain the anomaly identification results of each user under different models; Weighted module 23 is used for: For each user, based on the user's attributes, determine the anomaly weight corresponding to the anomaly identification results of the user under different models; Based on the anomaly weight, the anomaly identification results of the user under different models are weighted and fused to obtain the corresponding electricity bill anomaly identification result for the user.

[0088] In one possible implementation, user attributes include: user type, user stability, and user group size; Weighted module 23 is specifically used for: For each user, based on the user type, determine the basic weights corresponding to the anomaly identification results of that user under different models; Based on user stability and user group size, the base weights are adjusted to obtain the adjusted weights. The adjusted weights are then normalized to obtain the anomaly weights corresponding to the anomaly identification results of the user under different models.

[0089] In one possible implementation, user types include: residential users, commercial users, and industrial users; Weighted module 23 is specifically used for: For each user, if the user is a resident user, the base weight corresponding to the anomaly identification result of the user under the LSTM model is determined to be greater than the base weight corresponding to the anomaly identification result of the user under the graph neural network, and also greater than the base weight corresponding to the anomaly identification result of the user under the isolated forest algorithm network. If the user is a commercial user, then the base weight corresponding to the anomaly detection result of the user under the graph neural network is greater than the base weight corresponding to the anomaly detection result of the user under the LSTM model; and the base weight corresponding to the anomaly detection result of the user under the LSTM model is greater than the base weight corresponding to the anomaly detection result of the user under the isolated forest algorithm network. If the user is an industrial user, then the base weights corresponding to the anomaly identification results of the user under different models are determined to be the same.

[0090] In one possible implementation, the weighting module 23 is specifically used for: When the user stability index is higher than the first stability threshold, the base weight of the user's anomaly identification result under the LSTM model is reduced, and the base weight of the user's anomaly identification result under the Isolation Forest algorithm model is increased; the lower the user stability index, the higher the user stability. When the user stability metric is lower than the first stability threshold but higher than the second stability threshold, the base weight is maintained. When the user's stability index is lower than the second stability threshold, increase the base weight of the user's anomaly identification results under the LSTM model and decrease the base weight of the user's anomaly identification results under the Isolation Forest algorithm model.

[0091] In one possible implementation, the weighting module 23 is specifically used for: When the user group size index is higher than the first size threshold, the base weight corresponding to the anomaly identification result of the user under the graph neural network model is increased; the higher the user group size index, the larger the user group size. When the user group size metric is below the first size threshold but above the second size threshold, the basic weight is maintained. When the user group size index is lower than the second size threshold, the base weight corresponding to the abnormal identification result of the user in the graph neural network model is reduced.

[0092] In one possible implementation, the method for determining user stability metrics includes: Obtain the user's electricity consumption data within the target time period, and calculate the average and standard deviation of the electricity consumption data; The ratio of the standard deviation to the mean is used as the user stability index.

[0093] In one possible implementation, the method for determining the user group size metric includes: For each user, calculate the similarity of electricity consumption behavior, electricity consumption attributes, and load characteristics between that user and each of the remaining users. The similarity between the user and all remaining users is obtained by weighted summation of the similarity of electricity consumption behavior, similarity of electricity consumption attributes, and similarity of load characteristics. The number of remaining users whose overall similarity is greater than a set threshold is determined, and the number is normalized to obtain a user group size index.

[0094] In one possible implementation, the weighting module 23 is also used for: For each user, determine the comprehensive similarity sequence corresponding to that user, and in the comprehensive similarity sequence, determine the comprehensive similarity corresponding to the set quantile as the set threshold; The overall similarity sequence contains all the overall similarities corresponding to the user, and the overall similarities are arranged in ascending order.

[0095] This device embodiment can be used to implement the above method embodiment, and its technical principle and implementation effect are the same as those of the above method embodiment, so they will not be repeated here.

[0096] Figure 3 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. For example... Figure 3 As shown, the electronic device 3 of this embodiment includes a processor 30 and a memory 31. The memory 31 stores a computer program 32. When the processor 30 executes the computer program 32, it implements the steps in the various method embodiments described above. Alternatively, when the processor 30 executes the computer program 32, it implements the functions of each module / unit in the various device embodiments described above.

[0097] For example, computer program 32 may be divided into one or more modules / units, which are stored in memory 31 and executed by processor 30 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 32 in electronic device 3.

[0098] Electronic device 3 may include, but is not limited to, processor 30 and memory 31. Those skilled in the art will understand that... Figure 3 This is merely an example of electronic device 3 and does not constitute a limitation on electronic device 3. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device 3 may also include input / output devices, network access devices, buses, etc.

[0099] For the sake of simplicity and clarity, only the above-described functional modules / units are used as examples. In practical applications, the functions described above can be assigned to different functional modules / units as needed. These modules / units can be implemented in hardware, software, or a combination of both.

[0100] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0101] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for identifying electricity bill abnormality by multi-model fusion, characterized in that, The method comprises: obtaining electricity data of each user; respectively using an LSTM model, a graph neural network model and an isolation forest algorithm model to perform anomaly identification on the electricity data, and obtaining corresponding anomaly identification results of each user under different models; for each user, based on the user attributes of the user, determining an anomaly weight corresponding to the anomaly identification result of the user under different models; based on the anomaly weight, performing weighted fusion on the anomaly identification results of the user under different models to obtain an electricity anomaly identification result corresponding to the user. 2.The multi-model fused electricity charge anomaly identification method according to claim 1, characterized in that, The user attributes include: user type, user stability and user group size; for each user, based on the user attributes of the user, determining an anomaly weight corresponding to the anomaly identification result of the user under different models, comprising: for each user, based on the user type, determining a basic weight corresponding to the anomaly identification result of the user under different models; based on the user stability and the user group size, adjusting the basic weight to obtain an adjusted weight, and normalizing the adjusted weight to obtain an anomaly weight corresponding to the anomaly identification result of the user under different models. 3.The multi-model fused electricity bill anomaly identification method of claim 2, characterized in that, The user type includes: residential user, commercial user and industrial user; for each user, based on the user type, determining a basic weight corresponding to the anomaly identification result of the user under different models, comprising: for each user, if the user is a residential user, determining that the basic weight corresponding to the anomaly identification result of the user under the LSTM model is greater than the basic weight corresponding to the anomaly identification result of the user under the graph neural network, and greater than the basic weight corresponding to the anomaly identification result of the user under the isolation forest algorithm network; if the user is a commercial user, determining that the basic weight corresponding to the anomaly identification result of the user under the graph neural network is greater than the basic weight corresponding to the anomaly identification result of the user under the LSTM model; and the basic weight corresponding to the anomaly identification result of the user under the LSTM model is greater than the basic weight corresponding to the anomaly identification result of the user under the isolation forest algorithm network; if the user is an industrial user, determining that the basic weights corresponding to the anomaly identification results of the user under different models are the same.

4. The multi-model fused electricity charge anomaly identification method according to claim 2 or 3, characterized in that, Adjusting the basic weight based on the user stability, comprising: when the user stability index is higher than a first stability threshold, reducing the basic weight corresponding to the anomaly identification result of the user under the LSTM model, and increasing the basic weight corresponding to the anomaly identification result of the user under the isolation forest algorithm model; the lower the user stability index, the higher the user stability; when the user stability index is lower than the first stability threshold and higher than a second stability threshold, maintaining the basic weight; when the user stability index is lower than the second stability threshold, increasing the basic weight corresponding to the anomaly identification result of the user under the LSTM model, and reducing the basic weight corresponding to the anomaly identification result of the user under the isolation forest algorithm model.

5. The multi-model fused electricity charge anomaly identification method according to claim 2 or 3, characterized in that, Adjusting the basic weight based on the user group size, comprising: When the user group size indicator is higher than the first size threshold, increase the base weight corresponding to the abnormality identification result of the user under the graph neural network model; the higher the user group size indicator, the larger the user group size; When the user group size indicator is lower than the first size threshold and higher than the second size threshold, maintain the base weight; When the user group size indicator is lower than the second size threshold, decrease the base weight corresponding to the abnormality identification result of the user under the graph neural network model.

6. The multi-model fused electricity charge anomaly identification method according to claim 2 or 3, characterized in that, The method for determining the user stability indicator comprises: Obtain the electricity consumption data of the user in a target period, and calculate the average value and the standard deviation corresponding to the electricity consumption data; Determine the ratio of the standard deviation to the average value as the user stability indicator.

7. The multi-model fused electricity charge anomaly identification method according to claim 2 or 3, characterized in that, The method for determining the user group size indicator comprises: For each user, respectively calculate the electricity consumption behavior similarity, electricity consumption attribute similarity and load characteristic similarity between the user and each remaining user; Weighted sum the electricity consumption behavior similarity, electricity consumption attribute similarity and load characteristic similarity to obtain the comprehensive similarity between the user and each remaining user; Determine the number of remaining users whose comprehensive similarity is greater than a set threshold, and normalize the number to obtain the user group size indicator. 8.The multi-model fused electricity cost anomaly identification method of claim 7, wherein, The method further comprises: For each user, determine the comprehensive similarity sequence corresponding to the user, and determine the comprehensive similarity corresponding to a set quantile in the comprehensive similarity sequence as the set threshold; The comprehensive similarity sequence contains all comprehensive similarities corresponding to the user, and each comprehensive similarity is arranged in order from small to large. 9.A power consumption anomaly identification device based on multi-model fusion, characterized in that, Comprise: An acquisition module is configured to acquire electricity charge data of each user; An identification module is configured to respectively utilize an LSTM model, a graph neural network model and an isolation forest algorithm model to perform abnormality identification on the electricity charge data, and obtain abnormality identification results corresponding to each user under different models; A weighting module is configured to: For each user, respectively determine abnormality weights corresponding to the abnormality identification results of the user under different models based on user attributes of the user; Based on the abnormality weights, perform weighted fusion on the abnormality identification results of the user under different models to obtain an electricity charge abnormality identification result corresponding to the user.

10. An electronic device, comprising: A memory and a processor are included, the memory stores a computer program, and the processor implements the method of any one of claims 1 to 8 when executing the computer program.

Images