A cross-region electric meter anomaly detection method and system based on transfer learning and an electric energy meter

Abstract

The application discloses a cross-region electric meter anomaly detection method and system based on transfer learning and an electric energy meter, and belongs to the technical field of electric meter detection, and comprises the following steps: a layered feature extraction network is constructed, and electric meter data is input into the layered feature extraction network; cross-region feature alignment is performed on feature data extracted from the layered feature extraction network; after the cross-region feature alignment, a classifier is used to classify target region electric meter anomalies, and a time sequence anomaly positioning network is used to perform time sequence analysis on abnormal data; after anomaly positioning, the correlation between multiple abnormal events is identified, and an abnormal propagation graph is constructed. In the implementation process of the technical scheme, the layered feature extraction network is used to realize cross-region feature alignment of feature data, a classifier is used to classify and position target region electric meter anomalies, then the correlation between different abnormal events is identified through correlation analysis technology, and the electric meter anomaly detection effect is improved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of electric meter detection, in particular to a cross-region electric meter anomaly detection method, system and electric energy meter based on transfer learning. BACKGROUND

[0002] The power system is an indispensable infrastructure in the operation process of modern society, and its stable operation is crucial to economic development. As a key metering device in the power system, the accuracy of the electric meter directly affects the balance of power supply and demand and the safety of user electricity consumption.

[0003] For the accuracy check of the electric meter, periodic verification and manual inspection methods are currently used, which have problems such as low efficiency, high cost, limited coverage, etc. These methods have been gradually eliminated, and new technical means based on big data and artificial intelligence are used to realize intelligent analysis of cross-region electric meter data through transfer learning algorithm. Transfer learning is to transfer the learning ability on a specific task to different fields. In electric meter detection, it can improve the accuracy and efficiency of electric meter anomaly detection, especially for different types of electric meter detection. Through the application of transfer learning, abnormal conditions can be effectively identified and predicted between different regions.

[0004] However, although transfer learning has significant advantages in electric meter detection, there are still some problems in actual application. For example, in the process of cross-region electric meter detection, there are significant distribution differences between electric meter data in different regions. These distribution differences are reflected in data feature differences, abnormal type differences and data volume differences, etc. Directly using the model of the source region for detection will cause the model performance to decline when generalized to other regions, and the model training cost will increase due to the lack of target region data, resulting in poor cross-region use effect of transfer learning.

[0005] Therefore, it is necessary to provide a cross-region electric meter anomaly detection method, system and electric energy meter based on transfer learning to solve the above problems.

[0006] It should be noted that the above information disclosed in this background section is only used to understand the background technology of the concept of the application, and therefore, it can contain information that does not constitute prior art. SUMMARY

[0007] Based on the above problems existing in the prior art, the application solves the problem of providing a cross-region electric meter anomaly detection method, system and electric energy meter based on transfer learning, which optimizes anomaly detection based on the characteristics of cross-region electric meter data and improves the adaptability of transfer learning.

[0008] The technical solution adopted by the application to solve the technical problems is: a cross-region electric meter anomaly detection method based on transfer learning, comprising:

[0009] a hierarchical feature extraction network is constructed, and meter data from the source domain and the target domain is input into the hierarchical feature extraction network, the hierarchical feature extraction network comprising at least three shared layers and at least two target domain-specific layers;

[0010] The feature data extracted in the hierarchical feature extraction network is cross-region feature aligned and domain adapted, and the data distribution difference between the source domain and the target domain is compensated;

[0011] After cross-region feature alignment, the target region meter anomaly is classified by a classifier, and the abnormal data is analyzed in time sequence by a time sequence anomaly positioning network to locate the specific time period and position of the anomaly occurrence;

[0012] After anomaly positioning, the correlation between multiple abnormal events is identified by correlation analysis technology, an abnormal propagation graph is constructed, and the diffusion path of the abnormal event is predicted.

[0013] In the implementation process of the technical scheme of the present application, the hierarchical feature extraction network is used to realize cross-region feature alignment of feature data, and the target region meter anomaly is classified and positioned by a classifier, and then the correlation between different abnormal events is identified by correlation analysis technology, thereby improving the meter anomaly detection effect.

[0014] Further, each shared layer of the hierarchical feature extraction network adopts an LSTM structure, the three shared layers have different time windows and neuron numbers that increase progressively from small to large, and the output of each shared layer is connected to at least two target domain-specific layers.

[0015] Further, the shared layers are divided into a first shared layer, a second shared layer and a third shared layer, wherein the first shared layer is a physical constraint layer, which is embedded with power quality standards, the second shared layer is a time sequence characteristic layer, which is used to capture the time sequence locality features of the meter data, and the third shared layer is a regional commonality layer, which is used to extract cross-region general abnormal patterns.

[0016] Further, the third shared layer is provided with a power grid topology model as a constraint, adopts a graph neural network plus regional feature fusion gating mechanism, and takes a region as a node to learn the similarity of abnormal patterns between regions.

[0017] Further, the target domain-specific layer introduces a small sample fast adaptation mechanism, which establishes a regional pattern library and an adaptive routing network, the regional pattern library contains feature vectors of typical abnormal patterns, and based on the power operation and maintenance standards, the abnormal types are divided into fixed abnormal types or variable abnormal types.

[0018] Further, the cross-region feature alignment method further comprises: setting a mode-guided domain adversarial training, and embedding a mode-aware domain discriminator to output a mode-aware result, the mode-aware result comprising a probability of belonging to a source domain or a target domain; adjusting a weight of the domain adversarial training through the mode-aware result, and preferentially aligning features with significant regional characteristics; inputting a physical constraint in the adversarial process to make the domain adversarial process conform to electrical and mechanical physical characteristics, the physical constraint comprising a current amplitude limit, a voltage fluctuation range, and a harmonic energy proportion; and compensating for a data distribution difference between the source domain and the target domain according to a domain adversarial training result.

[0019] Further, the weight adjustment strategy of the domain adversarial training is as follows: the weights of the two parties in the confrontation are proportional to the mode-aware result, when the mode-aware result shows that a certain feature has a higher probability of belonging to the source domain, the weight of the source domain feature is increased by a proportional function value of the probability of the feature belonging to the source domain, and when the mode-aware result shows that a certain feature has a higher probability of belonging to the target domain, the weight of the target domain feature is increased by a proportional function value of the probability of the feature belonging to the target domain.

[0020] Further, the classifier comprises a shared classification layer and a dedicated classification layer, wherein the shared classification layer outputs a cross-domain general abnormal type based on shared features, the dedicated classification layer outputs an abnormal type specific to the target domain based on target domain features, and the output results of the two are fused as a final abnormal classification result, and a weighted fusion strategy is adopted in the fusion.

[0021] The application also provides a cross-region electric meter abnormality detection system based on transfer learning, which comprises:

[0022] a hierarchical extraction module for constructing a hierarchical feature extraction network and inputting electric meter data from a source domain and a target domain into the hierarchical feature extraction network, the hierarchical feature extraction network comprising at least three shared layers and at least two target domain dedicated layers;

[0023] a cross-region feature alignment module for performing cross-region feature alignment on feature data extracted by the hierarchical feature extraction network and performing domain adaptation to compensate for a data distribution difference between the source domain and the target domain;

[0024] a classification and positioning module for classifying target region electric meter abnormalities through a classifier after cross-region feature alignment, and performing time series analysis on abnormal data through a time series abnormality positioning network to locate specific time periods and positions of abnormality occurrence;

[0025] a correlation analysis module for identifying correlations between multiple abnormal events through correlation analysis technology after abnormality positioning, constructing an abnormal propagation graph, and predicting a diffusion path of the abnormal events.

[0026] The application also provides an electric energy meter, which comprises:

[0027] The hierarchical feature processing unit is configured to construct a hierarchical feature extraction network and input the meter data from the source domain and the target domain into the hierarchical feature extraction network.

[0028] The cross-region feature alignment unit is configured to perform cross-region feature alignment on the feature data extracted by the hierarchical feature extraction network, perform domain adaptation, and compensate for the data distribution difference between the source domain and the target domain.

[0029] The multi-core classification processor comprises a shared classification core and a target domain core.

[0030] The correlation analysis unit is configured to, after the abnormality is located, identify the correlation between the abnormal events by using a correlation analysis technique, construct an abnormality propagation graph, and predict the diffusion path of the abnormal events.

[0031] The application has the following beneficial effects: the cross-region meter abnormality detection method, system and electric energy meter provided by the application can realize cross-region feature alignment of the feature data by using a hierarchical feature extraction network, classify and locate the target region meter abnormality by using a classifier, identify the correlation between different abnormal events by using a correlation analysis technique, and improve the meter abnormality detection effect.

[0032] In addition to the purposes, features and advantages described above, the application has other purposes, features and advantages. The application will be described in detail below with reference to the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0033] The drawings accompanying the specification of the application form part of the application and serve to provide further understanding of the application. The schematic embodiments of the application and the description thereof serve to explain the application and do not constitute an improper limitation on the application. In the drawings:

[0034] Figure 1 FIG. 1 is a schematic diagram of the overall structure of the cross-region meter abnormality detection method based on the migration learning in the application;

[0035] Figure 2 FIG. 5 is a schematic diagram of the structure of the hierarchical feature extraction network;

[0036] Figure 3 FIG. 8 is a schematic diagram of the module structure of the cross-region meter abnormality detection system based on the migration learning in the application;

[0037] Figure 4 FIG. 11 is a schematic diagram of the structure of the electric energy meter in the application. DETAILED DESCRIPTION

[0038] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0039] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0040] Example 1: As Figure 1 As shown, this application provides a cross-regional electricity meter anomaly detection method based on transfer learning. This method is applied to the cross-regional electricity meter anomaly detection process, achieving cross-regional electricity meter detection through transfer learning, and optimizing for the data distribution differences between the source domain and the target domain, thereby achieving more accurate electricity meter anomaly detection. The method includes the following steps:

[0041] Step 101: Construct a hierarchical feature extraction network and input the meter data from the source domain and the target domain into the hierarchical feature extraction network. The hierarchical feature extraction network contains at least three shared layers and at least two target domain-specific layers.

[0042] In transfer learning, the source domain refers to the origin of the data and model, while the target domain refers to the place where the model is applied. When transfer learning is applied to cross-regional meter detection, there are differences in data distribution between the source and target domains. For example, the different meter usage habits in cities (as the source domain) and rural areas (as the target domain) lead to shifts in data features. Furthermore, the load characteristics of household electricity consumption differ significantly from those of industrial electricity consumption. When using transfer learning, problems such as differences in data features, anomaly types, and data volume can easily arise. Therefore, a hierarchical feature extraction network is constructed to extract meter data features from both the source and target domains. This hierarchical feature extraction network is designed based on a recurrent neural network structure, such as a Long Short-Term Memory (LSTM) network, which can process time-series data like meter data. This hierarchical feature extraction network includes multiple shared layers and target domain-specific layers. The shared layers extract general features, and the training of these shared layers uses only source domain data. The target domain-specific layers are used to extract target domain-specific features and are trained using target domain data. This approach retains general features while adapting to the data characteristics of the target domain.

[0043] like Figure 2As shown, the hierarchical feature extraction network includes three shared layers, and each shared layer adopts an LSTM structure, the three shared layers have different time windows and neuron numbers that are progressively larger, and capture short-term, medium-term and long-term meter data features, respectively, while the output of each shared layer is connected to at least two target domain specific layers, and when data features are extracted, a single shared layer or multiple shared layers in combination can work with the target domain specific layers to effectively process meter data of different quantities and types.

[0044] For example, when processing household electricity data, only the first shared layer can capture short-term features, and after the load pattern of household electricity is extracted through the shared layer, the target domain specific layer can further refine the features to identify abnormal electricity consumption behavior; when processing industrial electricity data, at least two shared layers are combined to capture multi-time scale features, and then the target domain specific layer can be used for in-depth analysis to identify complex abnormal patterns in industrial electricity consumption. Through this hierarchical combination, the comprehensiveness of feature extraction is ensured, and the data differences in cross-region detection are addressed, improving the accuracy of subsequent meter detection.

[0045] In addition, the three shared layers have different time windows and neuron numbers, and are optimized for the data characteristics of the meter. Specifically, the shared layers include a first shared layer, a second shared layer, and a third shared layer. The first shared layer is a physical constraint layer, which embeds power quality standards such as the GB / T 14549-1993 power quality standard, and inputs physical constraints in the structure design to ensure that the extracted features meet the power grid requirements, thereby preventing extraction errors due to data feature deviations and extracting bottom-level features that meet the power grid rules (such as power balance parameters and frequency stability).

[0046] The second shared layer is a time sequence characteristic layer for capturing the time sequence locality features of the meter data, such as the transition features of transient events and steady-state events, and is suitable for short-time strong disturbances and long-time smooth changes of meter abnormalities, avoiding the loss of details when a general model processes meter data with time sequence characteristics. The second shared layer is provided with a multi-scale convolution kernel and a time attention mechanism in structure, which can automatically mark the start and end time of transient events, such as voltage surge events in a certain area. The multi-scale convolution kernel can use an improved inception structure to replace the traditional large convolution kernel with smaller asymmetric convolution combinations, thereby achieving finer time resolution capture. In this embodiment, after using the improved inception structure, the second shared layer can capture multi-scale time resolution of one millisecond, ten milliseconds and one hundred milliseconds when processing voltage surge events, ensuring that complete meter data changes are captured at different time scales.

[0047] The third shared layer is a regional common layer, which is used to extract common abnormal patterns across regions, such as the amplitude and frequency correlation of load mutation, and common characteristics caused by urban device aging and rural line aging. The third shared layer is provided with a power grid topology model as a constraint to ensure that the extracted common characteristics conform to the power grid topology structure, avoiding misjudgment caused by regional differences. In terms of structure design, the third shared layer adopts a graph neural network plus regional feature fusion gating mechanism, which has a self-learning function and can adapt to changes in the power grid topology of different regions. It learns the similarity of abnormal patterns between regions with the region as a node. The graph neural network (GNN) is a model that uses neural networks to learn graph structure data. It captures complex interactions in the power grid topology through the connection relationship between nodes and edges, thereby achieving high-precision abnormality recognition in cross-regional detection. The feature fusion gating mechanism has an updateable database that can update the regional feature library in real time. It encodes typical abnormal patterns into the hierarchy based on pre-labeled data, and dynamically analyzes abnormal data based on real-time data in the target domain, ensuring that the generalization ability and adaptability of the model are improved in cross-regional anomaly detection.

[0048] The target domain specific layer also needs to meet the power quality requirements, so both target domain specific layers are provided with physical constraints, which have the same principle as the physical constraints of the first shared layer and will not be described again in this embodiment. In addition, the target domain specific layer also introduces a small sample fast adaptation mechanism, which can quickly adapt the model to the target domain with only a small amount of target domain samples, avoiding the need to retrain a large amount of data. Specifically, the small sample fast adaptation mechanism achieves fast matching and feature extraction of target domain samples through the establishment of a regional pattern library and an adaptive routing network. The regional pattern library contains feature vectors of typical abnormal patterns, and based on power operation standards, abnormal types are divided into fixed abnormal types or variable abnormal types, such as traditional abnormalities such as device aging, line overload, and harmonic interference as fixed abnormal types, and smart device interference and new power interference as variable abnormal types. The adaptive routing network dynamically classifies abnormal types.

[0049] For example, the data of a certain meter in the target domain shows sustained high current and low voltage. Using traditional pattern judgment methods, it will be directly judged as device aging (a common abnormality in urban areas), while in rural areas, due to differences in power load distribution, it will be more prone to line overload. Thanks to the small sample fast adaptation mechanism, through the known abnormal feature vectors in the regional pattern library and the dynamic classification of the adaptive routing network, it can accurately identify it as line overload in rural areas, avoiding misjudgment.

[0050] Step 102: Align the features extracted by the hierarchical feature extraction network across regions, and adapt to the domain to compensate for the differences in data distribution between the source domain and the target domain.

[0051] After the source domain and target domain data are respectively extracted by the hierarchical feature extraction network, cross-area feature alignment is still needed because the load characteristics of different areas are different, there is a certain data deviation, and thus domain feature difference is generated, so cross-area feature alignment is needed. Specifically, the method of cross-area feature alignment further includes the following steps:

[0052] Step 201: Set mode-guided domain adversarial training, and embed a mode-aware domain discriminator to output a mode-aware result, which includes the probability of belonging to the source domain or the target domain;

[0053] In the traditional feature alignment method, a generative adversarial network is usually used for domain adaptation to distinguish the source domain and target domain features. However, this method will cause the problem of blind alignment, resulting in over-alignment of part of the features, ignoring the regional characteristics, and affecting the model accuracy. Therefore, based on the patternized characteristics of the electric meter anomaly, mode-guided domain adversarial training is set, and a mode-aware domain discriminator is embedded to not only output the source domain or the target domain, but also identify the probability of the source domain or the target domain based on mode awareness, thereby facilitating subsequent domain adversarial training, improving the preservation of regional characteristics, and reducing errors caused by blind alignment. The mode-aware domain discriminator adopts a deep learning network structure, which includes multiple convolutional layers and fully connected layers, and accesses the output results of the shared layer and the target domain dedicated layer to align the general features of the source domain and the target domain, rather than simply aligning the data distribution. The probability of belonging to the source domain or the target domain is output by inputting the shared feature (Fshared) and the target domain feature (Fdomain), accurately identifying the regional characteristics, reducing misjudgment, and the shared feature is, for example, the amplitude frequency correlation of load mutation, and the target domain feature is, for example, the voltage drop feature of rural electric meters, the harmonic distortion feature of industrial area electric meters, and the frequency flicker feature of remote areas, etc.

[0054] Step 202: Adjust the weight of the domain adversarial training through the mode-aware result, and preferentially align the features with significant regional characteristics;

[0055] In the process of domain adversarial training, the weight distribution of the two opposing parties needs to be controlled to preferentially align the features with obvious regional characteristics. In traditional adversarial training, the weight distribution often ignores regional characteristics, leading to inaccurate alignment. In the foregoing process, the probability result output by the pattern perception domain discriminator can be used to dynamically adjust the weight based on the probability result without manual setting, ensuring that features with obvious regional characteristics are preferentially aligned. Specifically, the weight adjustment strategy of domain adversarial training is as follows: the weights of the two opposing parties are proportional to the pattern perception result. For example, when the pattern perception result shows that a feature has a high probability of belonging to the source domain, the weight of the source domain feature is increased accordingly, and the increase is proportional to the function value of the probability that the feature belongs to the source domain. Conversely, when the pattern perception result shows that a feature has a high probability of belonging to the target domain, the weight of the target domain feature is increased accordingly, and the increase is also proportional to the function value of the probability that the feature belongs to the target domain. Thus, accurate alignment of regional characteristics is achieved, and the weight is dynamically adjusted based on the probability result to ensure that features with obvious regional characteristics are preferentially aligned, improve the adaptability of the model in a specific region, and reduce misjudgment.

[0056] Step 203: Input physical constraints in the adversarial process to make the domain adversarial process comply with the physical characteristics of electromagnetism. The physical constraints include current amplitude limit, voltage fluctuation range, and harmonic energy proportion.

[0057] By introducing physical constraints, it is ensured that the model not only considers data distribution in domain adversarial training, but also follows the physical laws of power systems, further reducing errors caused by data deviating from actual physical characteristics, and improving the stability and reliability of the model in actual application, so as to more accurately identify and handle various meter abnormal situations. The physical constraints include but are not limited to current amplitude limit, voltage fluctuation range, and harmonic energy proportion. The constraint of harmonic energy proportion ensures that the model can accurately reflect the harmonic distribution of the actual power system when identifying harmonic distortion. It is implemented by using a constraint equation, and the specific equation can refer to the prior art.

[0058] Step 204: According to the domain adversarial training result, compensate for the difference in data distribution between the source domain and the target domain.

[0059] After completing the domain adversarial training, since there is a difference in data distribution between the source domain and the target domain, in addition to aligning the significant features, data compensation techniques are needed to reduce the difference, such as using a generative adversarial network to generate target domain samples to supplement source domain data, ensuring the generalization ability of the model in cross-domain application, or using existing transfer learning techniques to fine-tune the source domain model to the target domain, further optimizing the adaptability of the model between different domains, and improving the overall recognition accuracy.

[0060] Step 103: After cross-region feature alignment, the target region electric meter abnormality is classified by a classifier, and the abnormal data is analyzed by a time sequence anomaly positioning network to locate the specific time period and position of the abnormality;

[0061] The electric meter abnormality data includes abnormal information of different categories and different positions. After cross-region feature alignment, the classification and positioning can be performed. In this embodiment, the target region electric meter abnormality is classified by a classifier. The classifier includes a shared classification layer and a special classification layer. The shared classification layer outputs the cross-domain general abnormal type based on the shared features, such as the probability of device aging and harmonic distortion. The special classification layer outputs the abnormal type specific to the target domain based on the target domain features, such as regional power grid load mutation and voltage transient drop. The outputs of the two are fused as the final abnormal classification result. In the fusion, a weighted fusion strategy is adopted, and the weight is consistent with the dynamically adjusted weight in the foregoing process. The specific weighted fusion method can be weighted average or weighted summation. The output classification result contains not only the cross-domain general abnormal information but also the details of the abnormality specific to the target domain, so that the classification result can better reflect the actual power system.

[0062] The positioning of the abnormal data adopts a time sequence anomaly positioning network. The time sequence anomaly positioning network inputs the time sequence of the original electric meter data and adopts a bidirectional LSTM network structure to output the time interval of the abnormal event. The time sequence anomaly positioning network captures the forward and backward dependency in the time sequence through the bidirectional LSTM, so as to accurately identify the specific time period and position of the abnormality.

[0063] Step 104: After the abnormality is positioned, the correlation between multiple abnormal events is identified through correlation analysis technology, an abnormal propagation graph is constructed, and the diffusion path of the abnormal event is predicted.

[0064] After the specific time period and position of the abnormal event are determined, it is also necessary to determine whether there is correlation between multiple abnormal events, so as to optimize the electric meter abnormality detection process and predict the diffusion path of the abnormal event. Specifically, the correlation analysis technology is adopted, such as the Pearson correlation coefficient or mutual information method, to calculate the correlation between different abnormal events, and an abnormal propagation graph is constructed. The nodes in the graph represent abnormal events, and the edges represent the correlation strength between events. Then, the abnormal propagation path is analyzed through graph theory algorithm to predict the potential impact area. For example, if the load of a regional power grid suddenly increases, it may cause voltage fluctuation in the surrounding area. In the abnormal propagation graph, a strong correlation edge is established between the load sudden increase node and the voltage fluctuation node, and the diffusion path is predicted in this way to take preventive measures in advance to ensure the stable operation of the power system.

[0065] Embodiment two: as Figure 3As shown, the present application proposes a cross-regional electric meter anomaly detection system based on transfer learning, which runs the cross-regional electric meter anomaly detection method based on transfer learning in embodiment one. Through the transfer learning technology, the source domain electric meter data features are transferred to the target domain, and the target domain anomaly detection accuracy is improved. The system comprises:

[0066] The hierarchical extraction module is used to construct a hierarchical feature extraction network, and electric meter data from the source domain and the target domain are input into the hierarchical feature extraction network. The hierarchical feature extraction network contains at least three shared layers and at least two target domain specific layers.

[0067] The cross-regional feature alignment module is used to perform cross-regional feature alignment on the feature data extracted in the hierarchical feature extraction network, and perform domain adaptation to compensate for the data distribution difference between the source domain and the target domain.

[0068] The classification and positioning module is used to classify the target regional electric meter anomaly through the classifier after cross-regional feature alignment, and perform time series analysis on the abnormal data through the time series anomaly positioning network to locate the specific time period and position of the anomaly occurrence.

[0069] The correlation analysis module is used to identify the correlation between multiple abnormal events through correlation analysis technology after anomaly positioning, construct an abnormal propagation graph, and predict the diffusion path of the abnormal event.

[0070] Embodiment three: the present application also proposes an electric energy meter, as shown, Figure 4 The hierarchical feature processing unit is a dedicated chip, which is used to construct a hierarchical feature extraction network, and input electric meter data from the source domain and the target domain into the hierarchical feature extraction network. Through a three-layer stacked neural network engine, feature extraction of electric power data is realized. The three-layer stacked neural network engine includes a first shared layer, a second shared layer, and a third shared layer. The first shared layer is a physical constraint layer, which is embedded with power quality standards. The second shared layer is a time series characteristic layer, which is used to capture the time series locality features of the electric meter data. The third shared layer is a regional common layer, which is used to extract cross-regional general abnormal patterns.

[0071] The cross-regional feature alignment unit is used to perform cross-regional feature alignment on the feature data extracted in the hierarchical feature extraction network, and perform domain adaptation to compensate for the data distribution difference between the source domain and the target domain.

[0072] A multi-core classification processor, comprising a shared classification core and a target domain core, wherein the shared classification core can run a shared feature classification model using an ARM Cortex-M7 core, and the target domain core can load a region-specific anomaly detection model (such as a rural line grounding model) using a RISC-V coprocessor, for classifying target region meter anomalies through a classifier after cross-region feature alignment, and performing time sequence analysis on abnormal data through a time sequence anomaly positioning network to locate the specific time period and position of the anomaly occurrence;

[0073] An association analysis unit for identifying the association between multiple abnormal events after anomaly positioning through association analysis techniques, constructing an abnormal propagation graph, and predicting the diffusion path of abnormal events.

[0074] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Those skilled in the art can make various modifications and changes to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for cross-region electric meter anomaly detection based on transfer learning, characterized in that: The application relates to a method for power meter anomaly detection based on cross-region feature alignment and domain adaptation. The method comprises the following steps: a layered feature extraction network is constructed, and power meter data from a source domain and a target domain are input into the layered feature extraction network, the layered feature extraction network comprising at least three shared layers and at least two target domain-specific layers; cross-region feature alignment is performed on the feature data extracted from the layered feature extraction network, and domain adaptation is performed to compensate for the data distribution difference between the source domain and the target domain; after the cross-region feature alignment, a classifier is used to classify target region power meter anomalies, and a time sequence anomaly positioning network is used to perform time sequence analysis on the abnormal data to locate the specific time period and position of the anomaly; after the anomaly positioning, correlation analysis technology is used to identify the correlation between multiple anomaly events, an anomaly propagation graph is constructed, and the diffusion path of the anomaly event is predicted; each shared layer of the layered feature extraction network adopts an LSTM structure, the three shared layers have different time windows and neuron numbers that increase progressively from small to large, and the output of each shared layer is connected to at least two target domain-specific layers; the shared layers are divided into a first shared layer, a second shared layer and a third shared layer, wherein the first shared layer is a physical constraint layer, the power quality standard is embedded in the first shared layer, and a physical constraint is input in the structure design to ensure that the extracted features meet the power quality requirements, the second shared layer is a time sequence characteristic layer, which is used to capture the time sequence locality features of the power meter data, the second shared layer is provided with a multi-scale convolution kernel and a time attention mechanism in the structure, which is used to automatically label the start and end time of a transient event, and the third shared layer is a regional commonality layer, which is used to extract cross-region general abnormal patterns; 2. The method of claim 1, wherein the method is based on transfer learning. the third shared layer is provided with a power grid topology model as a constraint, adopts a graph neural network plus regional feature fusion gating mechanism, and learns the similarity of abnormal patterns between regions with regions as nodes.

3. The method of claim 1, wherein the method is based on transfer learning. The target domain-specific layer introduces a small sample fast adaptation mechanism, which establishes a regional pattern library and a self-adaptive routing network, the regional pattern library contains feature vectors of typical abnormal patterns, and based on the power operation and maintenance standard, the anomaly types are divided into fixed abnormal types or variable abnormal types. The method further comprises the following steps: a mode-guided domain adversarial training is set, and a mode perception domain discriminator is embedded to output a mode perception result, the mode perception result including the probability of belonging to the source domain or the target domain; the weight of the domain adversarial training is adjusted through the mode perception result to preferentially align features with significant regional characteristics; physical constraints are input in the adversarial process to make the domain adversarial process conform to the electrical physics characteristics, the physical constraints including current amplitude limitation, voltage fluctuation range and harmonic energy proportion; and the data distribution difference between the source domain and the target domain is compensated according to the domain adversarial training result.

4. The method of claim 3, wherein the method is based on transfer learning. The weight adjustment strategy of the domain adversarial training is as follows: the weights of the two parties are proportional to the mode perception result, when the mode perception result shows that the probability of a certain feature belonging to the source domain is higher, the weight of the source domain feature is increased, and the increase is proportional to the probability of the feature belonging to the source domain, when the mode perception result shows that the probability of a certain feature belonging to the target domain is higher, the weight of the target domain feature is increased, and the increase is proportional to the probability of the feature belonging to the target domain.

5. The method of claim 1, wherein the method is based on transfer learning. The classifier comprises a shared classification layer and a dedicated classification layer, wherein the shared classification layer outputs a cross-domain general abnormal type based on shared features, the dedicated classification layer outputs an abnormal type specific to the target domain based on target domain features, and the outputs of the two are fused as a final abnormal classification result, and a weighted fusion strategy is adopted in the fusion.

6. A cross-region power meter anomaly detection system based on transfer learning, used to implement a cross-region power meter anomaly detection method based on transfer learning according to any one of claims 1 to 5. The system comprises: The hierarchical extraction module is configured to construct a hierarchical feature extraction network and input meter data from the source domain and the target domain into the hierarchical feature extraction network, the hierarchical feature extraction network comprising at least three shared layers and at least two target domain dedicated layers; The cross-region feature alignment module is configured to perform cross-region feature alignment on the feature data extracted by the hierarchical feature extraction network and perform domain adaptation to compensate for the difference in data distribution between the source domain and the target domain; The classification and positioning module is configured to classify target region meter abnormalities through a classifier after cross-region feature alignment and perform time series analysis on abnormal data through a time series anomaly positioning network to locate specific time periods and positions of abnormal occurrences; The correlation analysis module is configured to identify correlations between multiple abnormal events through correlation analysis technology after abnormal positioning, construct an abnormal propagation graph, and predict a diffusion path of abnormal events; Each shared layer of the hierarchical feature extraction network adopts an LSTM structure, the three shared layers have different time windows and a number of neurons that progressively increases from small to large, and the output of each shared layer is connected to at least two target domain dedicated layers; The shared layers are divided into a first shared layer, a second shared layer, and a third shared layer, the first shared layer is a physical constraint layer, which is embedded with power quality standards and has physical constraints in the structure design to ensure that the extracted features meet the requirements of power quality, the second shared layer is a time series characteristic layer, which is used to capture the time series locality features of meter data, the second shared layer is provided with multi-scale convolution kernels and a time attention mechanism in the structure, which is used to automatically label the start and end times of transient events, and the third shared layer is a regional commonality layer, which is used to extract cross-region general abnormal patterns; The third shared layer is provided with a power grid topology model as a constraint, adopts a graph neural network plus regional feature fusion gating mechanism, and learns the similarity of abnormal patterns between regions with regions as nodes.

7. An electric energy meter, configured to implement the method of claim 1 to 5. The electric energy meter comprises: The hierarchical feature processing unit is configured to construct a hierarchical feature extraction network and input meter data from the source domain and the target domain into the hierarchical feature extraction network, the hierarchical feature extraction network comprising at least three shared layers and at least two target domain dedicated layers; The hierarchical feature processing unit is configured to construct a hierarchical feature extraction network and input meter data from the source domain and the target domain into the hierarchical feature extraction network, the hierarchical feature extraction network comprising at least three shared layers and at least two target domain dedicated layers; The cross-region feature alignment unit is configured to perform cross-region feature alignment on the feature data extracted by the hierarchical feature extraction network, perform domain adaptation, and compensate for the difference in data distribution between the source domain and the target domain. The multi-core classification processor includes a shared classification core and a target domain core. After cross-region feature alignment, the target region electric meter abnormality is classified by a classifier, and the abnormal data is subjected to time sequence analysis by a time sequence anomaly positioning network to locate the specific time period and position of the abnormality. The correlation analysis unit is configured to identify the correlation between multiple abnormal events and construct an abnormal propagation graph to predict the diffusion path of the abnormal event by using correlation analysis technology after anomaly positioning. Each shared layer of the hierarchical feature extraction network adopts an LSTM structure, and the three shared layers have different time windows and neuron numbers that increase progressively from small to large. In addition, the output of each shared layer is connected to at least two target domain-specific layers. The shared layers include a first shared layer, a second shared layer, and a third shared layer. The first shared layer is a physical constraint layer, which is embedded with power quality standards and is designed to ensure that the extracted features meet the physical constraints of power quality requirements. The second shared layer is a time sequence characteristic layer, which is configured to capture the time sequence locality features of the electric meter data. The second shared layer is provided with multi-scale convolution kernels and a time attention mechanism in the structure, which are used to automatically label the start and end times of transient events. The third shared layer is a regional commonality layer, which is configured to extract cross-region general abnormal patterns. The third shared layer is provided with a power grid topology model as a constraint, adopts a graph neural network plus regional feature fusion gating mechanism, and takes regions as nodes to learn the similarity of abnormal patterns between regions.

Images