Transfer learning-based cross-regional electric meter anomaly detection method and system, and electric energy meter

Abstract

The invention discloses a cross-regional electricity meter anomaly detection method and system based on transfer learning and an electric energy meter, and belongs to the technical field of electricity meter detection, and the method comprises the steps: constructing a layered feature extraction network, and inputting electricity meter data into the layered feature extraction network; performing cross-regional feature alignment on feature data extracted from the layered feature extraction network; after cross-regional feature alignment is carried out, electric meter anomalies in a target region are classified through a classifier, and time sequence analysis is carried out on abnormal data through a time sequence anomaly positioning network; and after abnormality positioning, identifying association among a plurality of abnormal events, and constructing an abnormality propagation graph. In the implementation process of the technical scheme of the invention, cross-regional feature alignment of the feature data is realized through the hierarchical feature extraction network, the target region electric meter anomaly is classified and positioned through the classifier, and then the association between different abnormal events is identified through the association analysis technology, so that the electric meter anomaly detection effect is improved.

Description

Technical Field

[0001] The present application relates to the field of electric meter detection technology, specifically to a cross-regional electric meter anomaly detection method, system and electric energy meter based on transfer learning. Background Art

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

[0003] Currently, methods such as periodic calibration and manual inspections are commonly used to check the accuracy of electricity meters. These methods have problems such as low efficiency, high cost, and limited coverage, and are now gradually being eliminated. Instead, new technologies based on big data and artificial intelligence are being used to achieve intelligent analysis of cross-regional electricity meter data through transfer learning algorithms. Transfer learning is the process of transferring learning capabilities for specific tasks to different fields. When applied to electricity meter detection, it can improve the accuracy and efficiency of meter anomaly detection, especially for different types of electricity meters. Through the application of transfer learning, anomalies can be effectively identified and predicted between different regions.

[0004] However, although transfer learning has significant advantages in electricity meter detection, there are still some problems in practical applications. For example, in the process of cross-regional electricity meter detection, there are significant distribution differences between the electricity meter data in different regions. These distribution differences are reflected in differences in data features, anomaly types, and data volume, etc. Directly using the model of the source region for detection will lead to a decline in model performance when generalized to other regions, and the increase in model training costs due to the lack of data in the target region, resulting in poor cross-regional use of transfer learning.

[0005] Therefore, it is necessary to provide a cross-regional electricity meter anomaly detection method, system and electricity 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 for understanding the background technology of the present application concept, and therefore, it may contain information that does not constitute prior art. Summary of the Invention

[0007] Based on the above-mentioned problems existing in the prior art, the problem to be solved by this application is: to provide a cross-regional electricity meter anomaly detection method, system and electricity meter based on transfer learning, to optimize anomaly detection based on the characteristics of cross-regional electricity meter data, and to improve the adaptability of transfer learning.

[0008] The technical solution adopted by this application to solve the technical problem is: a cross-regional electricity meter anomaly detection method based on transfer learning, including: Constructing a hierarchical feature extraction network and inputting the electricity meter data from the source domain and the target domain into the hierarchical feature extraction network, wherein the hierarchical feature extraction network includes at least three shared layers and at least two target domain-specific layers; Perform cross-region feature alignment on the feature data extracted from the hierarchical feature extraction network and perform domain adaptation to compensate for the data distribution differences between the source domain and the target domain; After cross-regional feature alignment, the classifier is used to classify the meter anomalies in the target area, and the time series anomaly location network is used to perform time series analysis on the anomaly data to locate the specific time period and location where the anomaly occurred. After locating the anomaly, correlation analysis technology is used to identify the correlation between multiple abnormal events, construct an abnormal propagation map, and predict the diffusion path of the abnormal event.

[0009] During the implementation of the technical solution of this application, a hierarchical feature extraction network is used to achieve cross-regional feature alignment of feature data, and a classifier is used to classify and locate meter anomalies in the target area. Then, through association analysis technology, the association between different abnormal events is identified to improve the meter anomaly detection effect.

[0010] Furthermore, each shared layer of the hierarchical feature extraction network adopts an LSTM structure, the three shared layers have different time windows and numbers of neurons that progress from small to large, and the output of each shared layer is connected to at least two target domain-specific layers.

[0011] Furthermore, the shared layer is 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 with embedded power quality standards, the second shared layer is a timing characteristic layer for capturing the timing local characteristics of meter data, and the third shared layer is a regional commonality layer for extracting common abnormal patterns across regions.

[0012] Furthermore, the third shared layer is provided with a power grid topology model as a constraint, adopts a graph neural network plus a regional feature fusion gating mechanism, and uses regions as nodes to learn the similarity of abnormal patterns between regions.

[0013] Furthermore, the target domain dedicated layer introduces a small sample rapid adaptation mechanism. The small sample rapid adaptation mechanism establishes a regional pattern library and an adaptive routing network. The regional pattern library contains characteristic vectors of typical abnormal patterns and divides abnormal types into fixed abnormal types or variable abnormal types based on power operation and maintenance standards.

[0014] Furthermore, the method for cross-regional feature alignment further includes: setting up pattern-guided domain adversarial training, and embedding a pattern-aware domain discriminator to output pattern perception results, which include the probability of belonging to the source domain or the target domain; adjusting the weights of domain adversarial training through the pattern perception results, and prioritizing the alignment of features with significant regional characteristics; inputting physical constraints during the adversarial process to make the domain adversarial process conform to the physical characteristics of electrical engineering, and the physical constraints include current amplitude limit, voltage fluctuation range, and harmonic energy ratio; and compensating for the data distribution differences between the source domain and the target domain based on the domain adversarial training results.

[0015] Furthermore, the weight adjustment strategy for domain adversarial training is as follows: the weights of the two adversarial parties are proportional to the pattern perception results. When the pattern perception results show that the probability of a certain feature belonging to the source domain is higher, the weight of the source domain feature is increased by a proportional function value of the probability that the feature belongs to the source domain. When the pattern perception results show that the probability of a certain feature belonging to the target domain is higher, the weight of the target domain feature is increased by a proportional function value of the probability that the feature belongs to the target domain.

[0016] Furthermore, the classifier includes a shared classification layer and a dedicated classification layer, wherein the shared classification layer outputs a cross-domain common anomaly type based on shared features, and the dedicated classification layer outputs anomaly types specific to the target domain based on target domain features, and the output results of the two are fused as the final anomaly classification result. During the fusion, a weighted fusion strategy is adopted.

[0017] This application also provides a cross-regional electricity meter anomaly detection system based on transfer learning, which includes: a hierarchical extraction module 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, wherein the hierarchical feature extraction network comprises at least three shared layers and at least two target domain-specific layers; The cross-region feature alignment module is used to align the feature data extracted from the hierarchical feature extraction network across regions and perform domain adaptation to compensate for the data distribution differences between the source domain and the target domain. The classification and positioning module is used to classify meter anomalies in the target area through a classifier after cross-region feature alignment, and perform time series analysis on the anomaly data through a time series anomaly positioning network to locate the specific time period and location of the anomaly; The association analysis module is used to identify the associations between multiple abnormal events through association analysis technology after the abnormality is located, build an abnormal propagation map, and predict the diffusion path of the abnormal event.

[0018] The present application also provides an electric energy meter, which includes: a hierarchical feature processing unit, configured to construct a hierarchical feature extraction network and input the electric meter data from the source domain and the target domain into the hierarchical feature extraction network; The cross-region feature alignment unit is used to align the feature data extracted from the hierarchical feature extraction network across regions and perform domain adaptation to compensate for the data distribution differences between the source domain and the target domain. a multi-core classification processor, the multi-core classification processor comprising a shared classification core and a target domain core; The correlation analysis unit is used to identify the correlation between multiple abnormal events through correlation analysis technology after the abnormality is located, construct an abnormal propagation map, and predict the diffusion path of the abnormal event.

[0019] The beneficial effects of the present application are as follows: the present application provides a cross-regional electricity meter anomaly detection method, system and electricity meter based on transfer learning, which realizes cross-regional feature alignment of feature data through a hierarchical feature extraction network, and classifies and locates the anomalies of electricity meters in the target area through a classifier, and then uses association analysis technology to identify the association between different abnormal events, thereby improving the effect of electricity meter anomaly detection.

[0020] In addition to the above-described purposes, features and advantages, the present application also has other purposes, features and advantages. The present application will be further described in detail below with reference to the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The drawings that constitute part of this application are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation on this application. In the drawings: Figure 1 This is an overall schematic diagram of a cross-regional electricity meter anomaly detection method based on transfer learning in this application; Figure 2 Schematic diagram of the structure of the hierarchical feature extraction network; Figure 3 This is a schematic diagram of the module structure of a cross-regional electricity meter anomaly detection system based on transfer learning in this application; Figure 4 This is a schematic diagram of the structure of an electric energy meter in this application. DETAILED DESCRIPTION

[0022] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0023] In order to enable those skilled in the art to better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0024] Example 1: Figure 1 As shown, the present application provides a cross-regional meter anomaly detection method based on transfer learning. The method is applied to the cross-regional meter anomaly detection process, realizes cross-regional meter detection through transfer learning, and optimizes the data distribution difference between the source domain and the target domain, thereby achieving more accurate meter anomaly detection. The method includes the following steps: Step 101: constructing a hierarchical feature extraction network and inputting the electricity meter data from the source domain and the target domain into the hierarchical feature extraction network, wherein the hierarchical feature extraction network includes at least three shared layers and at least two target domain-specific layers; In transfer learning, the source domain refers to the source of data and models, and the target domain refers to the place where the model is applied. In the process of applying transfer learning to cross-regional electricity meter detection, there will be differences in data distribution between the source domain and the target domain. For example, the usage habits of electricity meters in cities (as the source domain) and rural areas (as the target domain) are different, resulting in data feature offsets, and the load characteristics of household electricity consumption and industrial electricity consumption are significantly different. When using transfer learning, it is easy to have problems such as differences in data features, differences in anomaly types, and differences in data volume. Therefore, a hierarchical feature extraction network is constructed to realize the feature extraction of electricity meter data in the source domain and the target domain. The hierarchical feature extraction network is based on the recurrent neural network structure design, such as the long short-term memory network (LSTM). It can process data with time series characteristics such as electricity meter data. The hierarchical feature extraction network includes multiple shared layers and target domain-specific layers. Common features are extracted through the shared layers. The shared layers are trained only with source domain data. The target domain-specific layers are used to extract target domain-specific features and are trained with target domain data. This way, while retaining common features, it adapts to the data characteristics of the target domain. like Figure 2As shown in the figure, the hierarchical feature extraction network includes three shared layers, each of which uses an LSTM structure. The three shared layers have different time windows and progressively larger numbers of neurons, respectively capturing short-term, medium-term, and long-term meter data features. At the same time, the output of each shared layer is connected to at least two target domain-specific layers. When extracting data features, using a single shared layer or combining multiple shared layers can work in conjunction with the target domain-specific layers, effectively processing different amounts and types of meter data. For example, when processing household electricity data, only the first shared layer is used to capture short-term features. After extracting the household electricity load pattern through this shared layer, the target domain-specific layer further refines the features and identifies abnormal electricity usage behavior. When processing industrial electricity data, at least two shared layers are combined to capture multi-timescale features. Then, the target domain-specific layer conducts in-depth analysis to identify complex abnormal patterns in industrial electricity usage. This layered combination approach ensures comprehensive feature extraction while addressing data differences in cross-regional detection, improving the accuracy of subsequent meter detection. In addition to having different time windows and numbers of neurons, the three shared layers are also optimized for the data characteristics of electricity meters. Specifically, the shared layers are divided into the first, second, and third shared layers. The first shared layer is the physical constraint layer, which embeds power quality standards such as the "GB / T 14549-1993" power quality standard. Physical constraints are input during structural design to ensure that the extracted features meet power quality requirements. This prevents misjudgments in the shared layer due to data feature deviations and extracts underlying features that conform to power grid laws (such as power balance parameters and frequency stability). The second shared layer is a time series characteristic layer, which is used to capture the time series local characteristics of the meter data, such as the transition characteristics between transient events and steady-state events, adapt to the short-term strong disturbances and long-term smooth changes of meter anomalies, and avoid the problem of detail loss when the general model processes meter data with time series characteristics. Among them, the second shared layer is structurally provided with multi-scale convolution kernels and a time attention mechanism, which can automatically mark the start and end time of transient events, such as voltage surge and dip events in a certain area. The multi-scale convolution kernel can adopt the existing improved inception structure, replacing the traditional large convolution kernel with a smaller asymmetric convolution combination, thereby achieving finer time resolution capture. In this embodiment, after adopting the improved inception structure, the second shared layer can achieve multi-scale time resolution capture of one millisecond, ten milliseconds, and one hundred milliseconds when processing voltage surge and dip events, ensuring that the complete meter data changes are captured at different time scales; The third shared layer is the regional commonality layer, which is used to extract common cross-regional anomaly patterns, such as the correlation between the amplitude and frequency of load mutations, and the common characteristics caused by aging urban equipment and aging rural lines. The third shared layer is set with a power grid topology model as a constraint to ensure that the extracted common characteristics conform to the power grid topology and avoid misjudgments due to regional differences. In terms of structural design, the third shared layer adopts a graph neural network plus a regional feature fusion gating mechanism. It has self-learning capabilities and can adapt to changes in power grid topology in different regions. It uses regions as nodes to learn the similarity of anomaly patterns between regions. Graph neural network (GNN) is a model that uses neural networks to learn graph-structured data. Through the connection relationship between nodes and edges, it captures the complex interactions in the power grid topology, thereby achieving high-precision anomaly recognition in cross-regional detection. The feature fusion gating mechanism has an updateable database that can update the regional feature library in real time and encode it into the hierarchy based on pre-labeled typical anomaly patterns. Combined with real-time data from the target domain, the anomaly data is dynamically analyzed to ensure that the model's generalization ability and adaptability are improved in cross-regional anomaly detection. The target domain dedicated layer also needs to meet the power quality requirements. Therefore, both target domain dedicated layers are provided with physical constraints. The principle is consistent with the physical constraints of the first shared layer and will not be repeated in this embodiment. In addition, the target domain dedicated layer also introduces a small sample fast adaptation mechanism, which only requires a small number of target domain samples to achieve rapid adaptation of the model in the target domain, avoiding retraining of a large amount of data. Specifically, the small sample fast adaptation mechanism achieves rapid matching and feature extraction of target domain samples by establishing a regional pattern library and an adaptive routing network. The regional pattern library contains feature vectors of typical abnormal patterns and divides abnormal types into fixed abnormal types or variable abnormal types based on power operation and maintenance standards. For example, traditional abnormalities such as equipment aging, line overload, and harmonic interference are fixed abnormal types, and intelligent device interference and new power interference are variable abnormal types. The abnormal types are dynamically classified based on the adaptive routing network. For example, if a meter in the target domain shows persistent high current and low voltage, traditional pattern recognition methods would directly identify this as equipment aging (a common anomaly in urban areas) due to the difference between urban and rural areas. However, in rural areas, due to the different load distribution of the power grid, it is more likely to be a line overload. However, the use of a small sample rapid adaptation mechanism, through the dynamic classification of known anomaly feature vectors in the regional pattern library and the adaptive routing network, can accurately identify the rural area as a line overload, avoiding misjudgment. Step 102: Perform cross-region feature alignment on the feature data extracted from the hierarchical feature extraction network and perform domain adaptation to compensate for the data distribution differences between the source domain and the target domain; After extracting features from the source and target domain data using a hierarchical feature extraction network, cross-region feature alignment is required. This is because different regions have different load characteristics and data deviations, resulting in inter-domain feature differences. Therefore, cross-region feature alignment is necessary. Specifically, the cross-region feature alignment method further includes the following steps: Step 201: Setting up pattern-guided domain adversarial training and embedding a pattern-aware domain discriminator to output a pattern-aware result, which includes the probability of belonging to the source domain or the target domain; In traditional feature alignment methods, generative adversarial networks are usually used for domain adaptation to distinguish source domain and target domain features. However, this method will cause the problem of blind alignment, resulting in some features being over-aligned, ignoring regional characteristics, and affecting model accuracy. Therefore, based on the pattern characteristics of meter anomalies, pattern-guided domain adversarial training is set up. By embedding a pattern-aware domain discriminator, it not only outputs the source domain or target domain, but also can identify the probability of the source domain or target domain based on pattern perception, thereby facilitating subsequent domain adversarial training, improving the retention of regional characteristics, and reducing the error caused by blind alignment. The domain discriminator uses a deep learning network structure consisting of multiple convolutional and fully connected layers. It accesses the output results of shared layers and target domain-specific layers, aligning the common features of the source and target domains rather than simply aligning the data distribution. It outputs the probability of belonging to the source or target domain by inputting shared features (Fshared) and target domain features (Fdomain). This accurately identifies regional characteristics and reduces misjudgments. Shared features include the amplitude-frequency correlation of load mutations, while target domain features include voltage sag characteristics of rural electricity meters, harmonic distortion characteristics of industrial meters, and frequency flicker characteristics in remote areas. Step 202: Based on the pattern perception results, adjust the weights of the domain adversarial training to prioritize the alignment of features with significant regional characteristics; During domain adversarial training, it is necessary to control the weight distribution of the two adversarial parties so that features with obvious regional characteristics are aligned first. In traditional adversarial training, weight distribution often ignores regional characteristics, resulting in inaccurate alignment. In the aforementioned process, the probability results output by the pattern-aware domain discriminator can be used to dynamically adjust the weights based on the probability results without manual setting, ensuring that features with significant regional characteristics are aligned first. Specifically, the weight adjustment strategy of domain adversarial training is as follows: the weights of the two adversarial parties are proportional to the pattern perception results. For example, when the pattern perception results show that the probability of a feature belonging to the source domain is high, the weight of the source domain feature is increased accordingly, and the increase is a proportional function value of the probability of the feature belonging to the source domain. Conversely, when the pattern perception results show that the probability of a feature belonging to the target domain is high, the weight of the target domain feature is increased accordingly, and the increase is also a proportional function value of the probability of the feature belonging to the target domain, thereby achieving accurate alignment of regional characteristics and dynamically adjusting the weights based on this probability result to ensure that features with significant regional characteristics are aligned first, thereby improving the adaptability of the model in specific areas and reducing misjudgments. Step 203: Input physical constraints during the confrontation process to ensure that the domain confrontation process complies with the physical characteristics of electricity. The physical constraints include current amplitude limit, voltage fluctuation range, and harmonic energy ratio. By introducing physical constraints, the model ensures that it not only considers data distribution during domain adversarial training but also adheres to the physical laws of the power system. This further reduces errors caused by data deviations from actual physical characteristics, improves the model's stability and reliability in practical applications, and enables more accurate identification and processing of various meter anomalies. Physical constraints include, but are not limited to, current amplitude limits, voltage fluctuation ranges, and harmonic energy percentages. The harmonic energy percentage constraint ensures that the model accurately reflects the harmonic distribution of the actual power system when identifying harmonic distortion. This is achieved using constraint equations, and the specific equations can be referenced in existing technologies. Step 204: Compensate for the data distribution difference between the source domain and the target domain based on the domain adversarial training results.

[0025] After completing domain adversarial training, due to differences in data distribution between the source domain and the target domain, in addition to aligning significant features, it is also necessary to narrow the differences through data compensation technology, such as using a generative adversarial network to generate target domain samples to supplement the source domain data to ensure the model's generalization ability in cross-domain applications, or using existing transfer learning technology to fine-tune the source domain model to the target domain, further optimizing the model's adaptability between different domains and improving the overall recognition accuracy.

[0026] Step 103: After performing cross-region feature alignment, the classifier is used to classify the meter anomalies in the target area, and the time series anomaly location network is used to perform time series analysis on the anomaly data to locate the specific time period and location where the anomaly occurred. The meter anomaly data includes anomaly information of different categories and different locations. After cross-regional feature alignment, it can be aligned for classification and positioning. In this embodiment, the target area meter anomaly is classified by a classifier. The classifier includes a shared classification layer and a dedicated classification layer. The shared classification layer outputs cross-domain common anomaly types based on shared features, such as the probability of equipment aging and harmonic distortion. The dedicated classification layer outputs target domain-specific anomaly types based on target domain features, such as regional power grid load mutations and instantaneous voltage drops. The output results of the two are fused as the final anomaly classification result. During fusion, a weighted fusion strategy is adopted, in which the weights are consistent with the weights dynamically adjusted in the aforementioned process. The specific weighted fusion method can be weighted average or weighted sum. The output classification result contains both cross-domain common anomaly information and target domain-specific anomaly details, so that the classification result can better reflect the actual power system. A time series anomaly location network is used to locate abnormal data. The time series anomaly location network inputs the time series of raw meter data and uses a bidirectional LSTM network structure to output the time interval of abnormal events. The time series anomaly location network uses bidirectional LSTM to capture the front-end and back-end dependencies in the time series, thereby accurately identifying the specific time period and location of the anomaly. Step 104: After the anomaly is located, correlation analysis technology is used to identify the correlation between multiple abnormal events, construct an abnormal propagation map, and predict the diffusion path of the abnormal event; After determining the specific time period and location of the abnormal event, it is also necessary to determine whether there is a correlation between multiple abnormal events, so as to optimize the meter anomaly detection process and predict the diffusion path of abnormal events. Specifically, correlation analysis technology is used, such as methods based on the Pearson correlation coefficient or mutual information, to calculate the correlation between different abnormal events and construct an abnormal propagation graph. 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 algorithms to predict potential affected areas. For example, if the power grid load in a certain area suddenly increases, it may cause voltage fluctuations in the surrounding areas. In the abnormal propagation graph, strong correlation edges are established between the load sudden increase nodes and the voltage fluctuation nodes. Based on this, the diffusion path is predicted and preventive measures are taken in advance to ensure the stable operation of the power system.

[0027] Example 2: Figure 3 As shown, this application proposes a cross-regional electricity meter anomaly detection system based on transfer learning. The system runs the cross-regional electricity meter anomaly detection method based on transfer learning in Example 1. Through transfer learning technology, the source domain electricity meter data features are transferred to the target domain to improve the accuracy of target domain anomaly detection. The system includes: a hierarchical extraction module 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, wherein the hierarchical feature extraction network comprises at least three shared layers and at least two target domain-specific layers; The cross-region feature alignment module is used to align the feature data extracted from the hierarchical feature extraction network across regions and perform domain adaptation to compensate for the data distribution differences between the source domain and the target domain. The classification and positioning module is used to classify meter anomalies in the target area through a classifier after cross-region feature alignment, and perform time series analysis on the anomaly data through a time series anomaly positioning network to locate the specific time period and location of the anomaly; The association analysis module is used to identify the associations between multiple abnormal events through association analysis technology after the abnormality is located, build an abnormal propagation map, and predict the diffusion path of the abnormal event.

[0028] Example 3: This application also proposes an electric energy meter, such as Figure 4 As shown, the electric energy meter is equipped with a hierarchical feature processing unit, which is a dedicated chip for constructing a hierarchical feature extraction network. The electric energy meter data from the source domain and the target domain are input into the hierarchical feature extraction network, and the feature extraction of the electric energy data is realized through a three-layer stacked neural network engine. 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 with an embedded power quality standard. The second shared layer is a timing characteristic layer for capturing the timing local characteristics of the electric energy meter data. The third shared layer is a regional commonality layer for extracting common abnormal patterns across regions. The cross-region feature alignment unit is used to align the feature data extracted from the hierarchical feature extraction network across regions and perform domain adaptation to compensate for the data distribution differences between the source domain and the target domain. A multi-core classification processor, which includes a shared classification core and a target domain core. The shared classification core can use the ARM Cortex-M7 core to run a shared feature classification model, and the target domain core can use a RISC-V coprocessor to load a region-specific anomaly detection model (such as a rural line grounding model). After performing cross-regional feature alignment, the target domain core is used to classify meter anomalies in the target area using a classifier, and perform time series analysis on the anomaly data using a time series anomaly location network to locate the specific time period and location of the anomaly. The correlation analysis unit is used to identify the correlation between multiple abnormal events through correlation analysis technology after the abnormality is located, construct an abnormal propagation map, and predict the diffusion path of the abnormal event.

[0029] The above description is merely a preferred embodiment of the present application and is not intended to limit the present application. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present application shall be included within the scope of protection of the present application.

Claims

1. A cross-regional electricity meter anomaly detection method based on transfer learning, characterized by: include: Constructing a hierarchical feature extraction network and inputting the electricity meter data from the source domain and the target domain into the hierarchical feature extraction network, wherein the hierarchical feature extraction network includes at least three shared layers and at least two target domain-specific layers; Perform cross-region feature alignment on the feature data extracted from the hierarchical feature extraction network and perform domain adaptation to compensate for the data distribution differences between the source domain and the target domain; After cross-regional feature alignment, the classifier is used to classify the meter anomalies in the target area, and the time series anomaly location network is used to perform time series analysis on the anomaly data to locate the specific time period and location where the anomaly occurred. After locating the anomaly, correlation analysis technology is used to identify the correlation between multiple abnormal events, construct an abnormal propagation map, and predict the diffusion path of the abnormal event.

2. The cross-regional electricity meter anomaly detection method based on transfer learning according to claim 1 is characterized in that: Each shared layer of the hierarchical feature extraction network adopts an LSTM structure. The three shared layers have different time windows and numbers of neurons that progress from small to large. At the same time, the output of each shared layer is connected to at least two target domain-specific layers.

3. The cross-regional electricity meter anomaly detection method based on transfer learning according to claim 1 is characterized in that: The shared layer is divided into a first shared layer, a second shared layer and a third shared layer. The first shared layer is a physical constraint layer with embedded power quality standards. The second shared layer is a timing characteristic layer used to capture the timing local characteristics of meter data. The third shared layer is a regional commonality layer used to extract common abnormal patterns across regions.

4. The cross-regional electricity meter anomaly detection method based on transfer learning according to claim 3 is characterized by: The third shared layer is set with a power grid topology model as a constraint, adopts a graph neural network plus regional feature fusion gating mechanism, and uses regions as nodes to learn the similarity of abnormal patterns between regions.

5. The cross-regional electricity meter anomaly detection method based on transfer learning according to claim 1 is characterized in that: The target domain dedicated layer introduces a small sample rapid 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 divides abnormal types into fixed abnormal types or variable abnormal types based on power operation and maintenance standards.

6. The cross-regional electricity meter anomaly detection method based on transfer learning according to claim 1 is characterized in that: The method for cross-region feature alignment further includes: setting up pattern-guided domain adversarial training and embedding a pattern-aware domain discriminator to output pattern perception results, which include the probability of belonging to the source domain or the target domain; adjusting the weights of domain adversarial training based on the pattern perception results to prioritize alignment of features with significant regional characteristics; inputting physical constraints during the adversarial process to make the domain adversarial process conform to the physical characteristics of electrical engineering, which physical constraints include current amplitude limit, voltage fluctuation range, and harmonic energy ratio; and compensating for data distribution differences between the source domain and the target domain based on the domain adversarial training results.

7. The cross-regional electricity meter anomaly detection method based on transfer learning according to claim 6 is characterized in that: The weight adjustment strategy for domain adversarial training is as follows: the weights of the adversarial parties are proportional to the pattern perception results. When the pattern perception results show that a feature has a high probability of belonging to the source domain, the weight of the source domain feature is increased by a proportional function value of the probability that the feature belongs to the source domain. When the pattern perception results show that a feature has a high probability of belonging to the target domain, the weight of the target domain feature is increased by a proportional function value of the probability that the feature belongs to the target domain.

8. The cross-regional electricity meter anomaly detection method based on transfer learning according to claim 1 is characterized in that: The classifier includes a shared classification layer and a dedicated classification layer, wherein the shared classification layer outputs a cross-domain general anomaly type based on shared features, and the dedicated classification layer outputs anomaly types specific to the target domain based on target domain features. The output results of the two are fused as the final anomaly classification result, and a weighted fusion strategy is adopted during the fusion.

9. A cross-regional electricity meter anomaly detection system based on transfer learning, used to implement the cross-regional electricity meter anomaly detection method based on transfer learning according to any one of claims 1 to 8, characterized in that: The system includes: a hierarchical extraction module 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, wherein the hierarchical feature extraction network comprises at least three shared layers and at least two target domain-specific layers; The cross-region feature alignment module is used to align the feature data extracted from the hierarchical feature extraction network across regions and perform domain adaptation to compensate for the data distribution differences between the source domain and the target domain. The classification and positioning module is used to classify meter anomalies in the target area through a classifier after cross-region feature alignment, and perform time series analysis on the anomaly data through a time series anomaly positioning network to locate the specific time period and location of the anomaly; The association analysis module is used to identify the associations between multiple abnormal events through association analysis technology after the abnormality is located, build an abnormal propagation map, and predict the diffusion path of the abnormal event.

10. An electric energy meter, used to implement the cross-regional electric energy meter anomaly detection method based on transfer learning according to any one of claims 1 to 8, characterized in that: The energy meter includes: a hierarchical feature processing unit, configured to construct a hierarchical feature extraction network and input the electric meter data from the source domain and the target domain into the hierarchical feature extraction network; The cross-region feature alignment unit is used to align the feature data extracted from the hierarchical feature extraction network across regions and perform domain adaptation to compensate for the data distribution differences between the source domain and the target domain. a multi-core classification processor, the multi-core classification processor comprising a shared classification core and a target domain core; The correlation analysis unit is used to identify the correlation between multiple abnormal events through correlation analysis technology after the abnormality is located, construct an abnormal propagation map, and predict the diffusion path of the abnormal event.

Images