Deep learning-based multifunctional electric energy meter anomaly detection method and device

Abstract

The invention relates to the field of electric energy meter anomaly detection, in particular to a multifunctional electric energy meter anomaly detection method and device based on deep learning. The method comprises the steps that a power grid topological structure is modeled into a graph neural network, nodes of the graph neural network represent electric energy meters, attributes of the nodes are electric energy meter parameters and electric energy meter functions, and edges represent connection relations of the electric energy meters; the input of the graph neural network is feature vectors representing parameters and function character strings of the electric energy meter, and the output is the abnormal degree of the electric energy meter for each function; calculating the importance score of each electric energy meter for each function; setting an initial abnormal threshold for each function, and adjusting the initial abnormal threshold according to the importance score to obtain an optimal abnormal threshold corresponding to each function; and comparing the output of the graph neural network with the optimal anomaly threshold to obtain an anomaly detection result. The method has the effects of adaptively adjusting the abnormal threshold and accurately detecting the abnormal state of the multifunctional electric energy meter.

Description

Technical Field

[0001] This application relates to the field of electricity meter anomaly detection, and in particular to a method and device for anomaly detection of multifunctional electricity meters based on deep learning. Background Art

[0002] With the rapid development of smart grids and the improvement of information technology in power systems, multi-functional energy meters, as core devices for power metering and data acquisition, play an important role in energy management, load monitoring, and user behavior analysis.

[0003] However, due to external environmental interference, equipment aging, line faults, or human factors, multi-functional energy meters may experience abnormal operating conditions, such as energy metering errors, data transmission interruptions, or abnormal load detection. These problems not only affect the meter's own metering accuracy but may also lead to inaccurate energy consumption assessments of the power system, increased electricity safety risks, and economic losses.

[0004] Traditional methods for detecting anomalies in electricity meters mainly rely on statistical analysis or rule-based threshold judgments, such as identifying abnormal data by setting fixed thresholds. However, these methods have limitations when facing complex and ever-changing real-world scenarios, such as difficulty in effectively identifying highly concealed anomalies and weak adaptability to changes in data distribution. Summary of the Invention

[0005] To address the technical challenge of adaptively setting anomaly thresholds, this application provides a deep learning-based method for anomaly detection in multifunctional energy meters.

[0006] Firstly, this application provides a deep learning-based method for detecting anomalies in multi-functional energy meters, employing the following technical solution: The deep learning-based method for detecting anomalies in multi-functional energy meters includes the following steps: acquiring power grid topology data, energy meter parameters, and energy meter functions; encoding the energy meter functions to obtain function strings; modeling the power grid topology as a graph neural network, where nodes represent energy meters, node attributes are energy meter parameters and functions, and edges represent the connection relationships between energy meters; the input to the graph neural network is a feature vector representing the energy meter parameters and function strings, and the output is the degree of anomaly of the energy meter for each function; calculating the importance score of each energy meter for each function; the calculation method is as follows: under each function, calculating the network density and betweenness centrality of nodes, and obtaining the energy meter contribution based on the network density and betweenness centrality; calculating the matching degree of real-time data of the energy meter in terms of function; calculating the feature similarity of the energy meter and its neighboring energy meters in terms of function; calculating the importance of nodes for each function based on the energy meter contribution, matching degree, and feature similarity; and obtaining the importance score of the node after normalizing the importance score. For each function, an initial anomaly threshold is set, and the initial anomaly threshold is adjusted according to the importance score to obtain the optimal anomaly threshold for each function; the anomaly detection result is obtained by comparing the output of the graph neural network with the optimal anomaly threshold.

[0007] Optionally, the network density can be defined as the ratio of the number of edges connecting each node to the maximum number of edges connecting a node in the graph neural network.

[0008] Optionally, the number of actual connecting edges can be obtained directly from the power grid topology data or extracted from the edge set in the graph neural network; the maximum possible number of edges for a node can be obtained by dividing the product of the total number of nodes and the total number of nodes minus 1 by 2; the ratio of the actual number of edges to the maximum possible number of edges can be used as the average network density; and the ratio of the number of connecting edges for each node to the average network density can be normalized and used as the network density.

[0009] Optionally, the method for obtaining the contribution of the electricity meter based on network density and betweenness centrality is to use the product of network density and betweenness centrality as the contribution of the electricity meter.

[0010] Optionally, the degree of functional matching of the real-time data of the electricity meter can be calculated. The calculation method is as follows: the degree of matching of the electricity meter in a specific function is obtained by calculating the cosine similarity between the feature vector of the electricity meter and the preset reference vector.

[0011] Optionally, the method for calculating the functional similarity between an energy meter and its neighboring energy meters is as follows: construct an energy meter parameter sequence, calculate the mean Pearson correlation coefficient between the energy meter parameter sequence corresponding to any node and the energy meter parameter sequence corresponding to the neighboring nodes connected to that node, and use the mean Pearson correlation coefficient as the functional similarity between the energy meter and its neighboring energy meters.

[0012] Optionally, the optimal abnormal threshold for each function can be obtained by adjusting the initial abnormal threshold based on the importance score. The method is as follows: traverse all electricity meters in the power grid, and sum the importance scores of all electricity meters for a certain function as the total importance weight of that function; calculate the average importance weight of all functions; take the difference between the total importance weight and the average importance weight as the deviation of the importance weight; use the product of a preset adjustment coefficient and the deviation as the threshold adjustment amount; sum the threshold adjustment amount and the initial abnormal threshold as the optimal abnormal threshold for that function, and traverse all functions to obtain the optimal abnormal threshold for each function.

[0013] Optionally, the initial outlier threshold can be set by obtaining historical data statistics and adding a certain number of standard deviations to the average value as the initial outlier threshold.

[0014] Optionally, after normalizing the importance, the importance score can be obtained by using the ratio of the importance of each node to the sum of the importance of its neighbors to the sum of the importance of all nodes in the graph and their neighbors as the importance score of a node to its neighbors.

[0015] Secondly, this application provides a multi-functional energy meter anomaly detection device based on deep learning, which adopts the following technical solution: A deep learning-based multi-functional energy meter anomaly detection device includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the deep learning-based multi-functional energy meter anomaly detection method described above.

[0016] The beneficial effect is that the above-mentioned deep learning-based multi-functional energy meter anomaly detection method is generated into a computer program and stored in a memory so that it can be loaded and executed by a processor. Thus, a device can be made based on the memory and processor, which is convenient to use.

[0017] This application has the following technical effects: 1. By calculating the importance score of the electricity meter, this technical solution can dynamically adjust the abnormal thresholds of each function to adapt to complex and ever-changing real-world scenarios. In multi-functional electricity meters, different functions may have different importance and abnormal characteristics, so static thresholds are often insufficient to meet the requirements. This solution generates optimal abnormal thresholds based on multi-dimensional information such as the electricity meter's contribution, matching degree, and feature similarity, ensuring the accuracy and robustness of the detection results and further optimizing the electricity meter's functional monitoring and anomaly handling capabilities.

[0018] 2. The importance of a function varies for each electricity meter under different tasks (such as billing). By calculating network density and node betweenness centrality, combined with the meter's contribution, the functional matching degree of real-time data, and the similarity of functional characteristics with neighboring meters, the importance of an electricity meter in a specific function can be comprehensively evaluated. In the scenario of multi-functional electricity meters, not only the parameters and functional performance of an individual electricity meter are considered, but also its position in the network and the influence of neighboring nodes. This enables accurate detection of abnormal meter states, more effectively identifying highly concealed anomalies and improving the operational reliability of the power system. Attached Figure Description

[0019] Figure 1 This is a flowchart of the anomaly detection method for a multi-functional energy meter based on deep learning, according to an embodiment of this application.

[0020] Figure 2 This is a flowchart illustrating the method for calculating important scores in the deep learning-based anomaly detection method for multifunctional energy meters according to embodiments of this application. Detailed Implementation

[0021] This application discloses a deep learning-based anomaly detection method for multi-functional energy meters, referring to... Figure 1 The steps include: S1: Obtain power grid topology data, electricity meter parameters, and electricity meter functions, and encode the electricity meter functions to obtain function strings.

[0022] Specifically, the power grid topology represents the connection relationships between electricity meters in the form of a graph. Data sources can include: wiring diagrams of the actual power grid, network topology files, or topology information in a smart grid management system. The data format can be an adjacency matrix, an edge list, or records in a graph database.

[0023] The functions of an electricity meter refer to the specific functional modules or services supported by the meter, such as: electricity metering: recording the active power (kWh) and reactive power (kvarh) consumed by the user; demand detection: monitoring the user's maximum power consumption (demand) for time-of-use billing or demand-side management; voltage / current measurement: measuring the voltage and current values ​​in the power grid in real time; power factor calculation: calculating the power factor of electrical equipment to reflect power efficiency; frequency monitoring: monitoring frequency fluctuations in the power grid to ensure power quality, etc.

[0024] To facilitate subsequent calculations and modeling, the functions of the electricity meter need to be converted into a machine-readable form, namely a function string. Each function is assigned a bit; if the function is supported, the corresponding bit is 1, otherwise it is 0. For example, the function string is [1,1,0,0,1] (bits 1, 2, and 5 are 1, indicating support for the corresponding function). Alternatively, other existing encoding methods applicable to the scenario of this application can be used to encode the electricity meter functions into a function string.

[0025] S2: Model the power grid topology as a graph neural network. The nodes of the graph neural network represent electricity meters, and the attributes of the nodes are the electricity meter parameters and functions. The edges represent the connection relationships of the electricity meters. The input of the graph neural network is a feature vector representing the electricity meter parameters and function strings, and the output is the degree of abnormality of the electricity meter for each function.

[0026] Specifically, a Graph Neural Network (GNN) is a deep learning model specifically designed for processing graph-structured data. In this application, the power grid topology is modeled as a graph G=(V,E), where each node in the set V represents an electricity meter, and each edge in the set E represents the connection between electricity meters.

[0027] The parameter vector is standardized (e.g., Z-score normalization or Min-Max normalization) to ensure all features have the same numerical range. Real-time parameters (e.g., voltage, current, power factor) and basic parameters (e.g., rated voltage, rated current) of the electricity meter are combined into a feature vector. This parameter vector is then concatenated with the function string from step S1 to obtain the node-comprehensive feature vector. For example, for an electricity meter... Given a parameter vector of [230, 8, 0.95] (representing voltage, current, and power factor respectively) and a function string of [1, 1, 0, 0, 1], the feature vector is: .

[0028] The architecture of graph neural networks is existing technology and will not be elaborated upon here. Edge weights reflect the strength or importance of the connection between electricity meters. They can be defined as follows: based on physical distance, the closer the electricity meters are physically, the larger the edge weight. Based on functional relevance, the larger the edge weight if two electricity meters support similar functions. Based on historical data, relevance is calculated based on historical electricity consumption data and used as the edge weight. For example, assume the electricity meters... and The physical distance between them is Then the edge weight can be defined as: ; This is a hyperparameter used to avoid a denominator of 0. It can be set to 0.01, but in practical applications, it needs to be adjusted based on the data distribution and experimental results. Multiple factors can be considered to define edge weights, introducing a multi-dimensional edge weight definition method, and integrating the influence of different dimensions through normalization processing. Existing technologies will not be elaborated further.

[0029] In the training of the graph neural network, historical data of electricity meters are collected, and the status of each electricity meter under various functions (such as normal or abnormal) is labeled as label data. Cross-entropy loss or mean squared error loss is used to measure the difference between the predicted results and the true labels. Gradient descent method (such as Adam) is used to optimize the model parameters.

[0030] The input to the graph neural network is the feature vectors of all nodes and the edge weight matrix. For each node, the output is its degree of anomalousness under each function. For example, if an electricity meter supports three functions, the output of the graph neural network would be: [0.1,0.8,0.3] indicates that the degree of abnormality of the electricity meter is 0.1 for the first function, 0.8 for the second function, and 0.3 for the third function.

[0031] S3: Calculate the importance score of each energy meter for each function.

[0032] Reference Figure 2 The calculation method for important scores includes steps S30-S33, as detailed below: S30: Under each function, calculate the network density and betweenness centrality of the nodes, and obtain the contribution of the electricity meter based on the network density and betweenness centrality.

[0033] In one embodiment, the network density is defined as the ratio of the number of edges connecting each node to the maximum number of edges connected to a single node in the graph neural network. This embodiment is suitable for the analysis of small power grids or localized areas, where the network size is small and the calculation is simple.

[0034] In other embodiments, the number of actual connecting edges is obtained directly from the power grid topology data or extracted from the edge set in the graph neural network; the maximum possible number of edges for each node is obtained by dividing the product of the total number of nodes and the total number of nodes minus 1 by 2; the ratio of the actual number of edges to the maximum possible number of edges is used as the average network density; the network density is obtained by normalizing the ratio of the number of connecting edges for each node to the average network density, using existing normalization methods such as standard normalization or maximum / minimum value normalization. This embodiment is suitable for large-scale power grids or global analysis and can more accurately reflect the overall connectivity of the network.

[0035] Betweenness centrality is an existing technique used to measure a node's mediating ability in a network (i.e., how many shortest paths pass through that node). It is calculated directly using graph neural network toolkits (such as NetworkX), and the existing techniques will not be elaborated further. The product of network density and betweenness centrality is used as the contribution of the electricity meter.

[0036] S31: Calculate the degree of functional matching between the real-time data of the electricity meter and the actual data.

[0037] In one embodiment, the degree of matching of the energy meter in a specific function is obtained by calculating the cosine similarity between the energy meter's feature vector and a preset reference vector. The cosine similarity value ranges from [-1, 1], with a higher value indicating a higher degree of matching.

[0038] By setting a benchmark vector (generated based on historical data or expert experience) that reflects the ideal functional performance, the cosine similarity is used to calculate the degree of matching between the feature vector of the electricity meter and the benchmark vector. The larger the value, the closer the performance of the electricity meter in a specific function is to the ideal state, thereby realizing a quantitative evaluation of the functional performance of the multi-functional electricity meter.

[0039] A benchmark vector is a standard used to measure the ideal performance of an energy meter in a specific function. To align with the real-world scenarios of multi-functional energy meters, the benchmark vector needs to be set by combining specific functional requirements with historical data or expert experience. The optimal feature combination can be fitted using machine learning models (such as regression analysis) as the benchmark vector. To ensure comparability between different features, the benchmark vector is normalized.

[0040] S32: Calculate the functional similarity between an electricity meter and its neighboring electricity meters.

[0041] The method for calculating the functional similarity between an electricity meter and its neighboring electricity meters is as follows: Construct a sequence of electricity meter parameters, and calculate the mean Pearson correlation coefficient between the parameter sequence of any node and the parameter sequences of its neighboring nodes. The calculation of the Pearson correlation coefficient is existing technology and will not be elaborated further. The mean Pearson correlation coefficient is used as the functional similarity between the electricity meter and its neighboring electricity meters. If the similarity is high (close to 1), it indicates that the electricity meter... If a meter functions identically to its neighboring meters, it is likely a normally functioning node. If the characteristic similarity is low (close to 0 or negative), it indicates that the meter... The energy meter may exhibit significant functional differences from its neighboring meters, potentially indicating abnormal or isolated behavior. By calculating the functional similarity between the energy meter and its neighboring meters, the functional consistency of energy meters within a local area can be effectively assessed.

[0042] S33: Calculate the importance of nodes to each function based on the contribution of the electricity meter, the degree of matching, and the feature similarity; obtain the importance score of the node after normalizing the importance score.

[0043] Specifically, weights 1, 2, and 3 are set. The sum of the products of weight 1 and the contribution of the electricity meter, weight 2 and the matching degree, and weight 3 and the feature similarity is used as the importance level. Weights 1, 2, and 3 can be adjusted through experiments or expert experience; for example, the default values ​​are 0.4 for weight 1, 0.3 for weight 2, and 0.3 for weight 3. The sum of weights 1, 2, and 3 is 1.

[0044] By comprehensively evaluating the contribution, functional matching degree, and feature similarity of electricity meters, their importance in specific functions is calculated, and a final importance score is obtained through normalization. This approach not only considers the characteristics of the electricity meter itself but also its location and neighbor relationships in the network, thus comprehensively reflecting the relative importance of the electricity meter in multifunctional scenarios.

[0045] In one embodiment, the method for normalizing importance is as follows: the ratio of the importance of each node to the sum of the importance of its neighbors to the sum of the importance of all nodes in the graph and their neighbors is used as the node's importance score relative to its neighbors. The aim is to place the node's own importance within the framework of the relative proportion between its importance and that of its neighbors to the total importance of all neighbors in the entire network, thereby reflecting the relative influence of a node on its local neighborhood and even the entire network.

[0046] The importance of a function varies for each electricity meter depending on the task (such as billing). By calculating network density and node betweenness centrality, and combining this with the meter's contribution, the functional matching degree of real-time data, and the similarity of functional characteristics with neighboring meters, the importance of a meter in a specific function can be comprehensively evaluated. For example, in a billing task, a particular meter may be particularly important for its electricity metering function due to its location in the network (high betweenness centrality and network density) or the proximity of its real-time data to a preset benchmark (high functional matching degree). Furthermore, if this meter's functional performance is highly consistent with its neighboring meters, its characteristic similarity is high, further enhancing its importance in billing-related tasks.

[0047] S4: Set an initial anomaly threshold for each function, adjust the initial anomaly threshold according to the importance score to obtain the optimal anomaly threshold for each function; obtain the anomaly detection result by comparing the output of the graph neural network with the optimal anomaly threshold.

[0048] Specifically, historical data statistics are obtained, and the average value plus a certain number of standard deviations is used as the initial anomaly threshold. For example, it could be 2 or 3 standard deviations. All electricity meters in the power grid are traversed, and the sum of the importance scores of all meters for a certain function is used as the total importance weight for that function. The average importance weight of all functions is calculated; the difference between the total importance weight and the average importance weight is used as the deviation of the importance weight; a preset adjustment coefficient is used as the product of the deviation as the threshold adjustment amount; the adjustment coefficient can be, for example, 0.1 or 0.05, to control the adjustment magnitude. The sum of the threshold adjustment amount and the initial anomaly threshold is used as the optimal anomaly threshold for that function. This process is repeated for all functions to obtain the optimal anomaly threshold for each function.

[0049] For each electricity meter, under each function, the anomaly level output by its graph neural network is compared with the optimal anomaly threshold for that function. If the anomaly level is greater than the threshold, it is determined to be abnormal; otherwise, it is determined to be normal. For example, in the electricity metering function, suppose an electricity meter records abnormal fluctuations in active and reactive power. Through graph neural network modeling, the feature vector of this electricity meter contains parameters such as real-time voltage, current, and power factor. The system calculates its network density and betweenness centrality, finding that the electricity meter is located in the core area of ​​the power grid and has a high contribution. Simultaneously, through cosine similarity calculation, it is found that the matching degree between its electricity metering data and the benchmark vector is low, and the feature similarity with its neighboring electricity meters also decreases significantly. The electricity meter has a low importance score, indicating a possible electricity metering error. Based on the adjusted optimal anomaly threshold, the system determines that the electricity meter has an anomaly in its electricity metering function.

[0050] This application also discloses a deep learning-based multi-functional energy meter anomaly detection device, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the deep learning-based multi-functional energy meter anomaly detection method according to this application.

[0051] The aforementioned device also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and therefore will not be described in detail here.

[0052] In this application, the aforementioned memory can be any tangible medium that contains or stores a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as resistive random access memory (DRAM), dynamic random access memory (DRAM), static random access memory (SRAM), etc., or any other medium that can be used to store desired information and can be accessed by an application program, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device.

[0053] The above are all preferred embodiments of the present application, and are not intended to limit the scope of protection of the present application. Therefore, any equivalent changes made based on the structure, shape, and principle of the present application should be included in the scope of protection of the present application.

Claims

1. A method for anomaly detection in a multi-functional energy meter based on deep learning, characterized in that, Including the following steps: Acquire power grid topology data, electricity meter parameters, and electricity meter functions, and encode the electricity meter functions to obtain function strings; The power grid topology is modeled as a graph neural network. The nodes of the graph neural network represent electricity meters, and the attributes of the nodes are the electricity meter parameters and functions. The edges represent the connection relationships of the electricity meters. The input of the graph neural network is a feature vector representing the electricity meter parameters and function strings, and the output is the degree of abnormality of the electricity meter for each function. Calculate the importance score of each energy meter for each function. The calculation method is as follows: for each function, calculate the network density and betweenness centrality of the nodes, and obtain the energy meter contribution based on the network density and betweenness centrality; calculate the matching degree of the real-time data of the energy meter in terms of function; calculate the feature similarity of the energy meter and its neighboring energy meters in terms of function; calculate the importance of the node to each function based on the energy meter contribution, matching degree, and feature similarity; and obtain the importance score of the node after normalizing the importance score. For each function, an initial anomaly threshold is set, and the initial anomaly threshold is adjusted according to the importance score to obtain the optimal anomaly threshold for each function; the anomaly detection result is obtained by comparing the output of the graph neural network with the optimal anomaly threshold.

2. The anomaly detection method for multifunctional energy meters based on deep learning according to claim 1, characterized in that, The network density is the ratio of the number of edges connecting each node to the maximum number of edges connecting a node in a graph neural network.

3. The anomaly detection method for multifunctional energy meters based on deep learning according to claim 1, characterized in that, The number of actual connecting edges can be obtained directly from power grid topology data or extracted from the edge set in a graph neural network. Divide the product of the total number of nodes and the total number of nodes minus 1 by 2 to get the maximum possible number of edges for a node. The ratio of the actual number of edges to the maximum possible number of edges is used as the average network density; The network density is calculated by normalizing the ratio of the number of edges connecting each node to the average network density.

4. The anomaly detection method for a multi-functional energy meter based on deep learning according to claim 2 or 3, characterized in that, The method for obtaining the contribution of the electricity meter based on network density and betweenness centrality is as follows: the product of network density and betweenness centrality is taken as the contribution of the electricity meter.

5. The anomaly detection method for a multi-functional energy meter based on deep learning according to claim 1, characterized in that, The degree of functional matching of the real-time data of the electricity meter is calculated by calculating the cosine similarity between the feature vector of the electricity meter and the preset reference vector to obtain the degree of matching of the electricity meter in a specific function.

6. The anomaly detection method for a multi-functional energy meter based on deep learning according to claim 1, characterized in that, The method for calculating the functional similarity between an electricity meter and its neighboring electricity meters is as follows: construct an electricity meter parameter sequence, calculate the mean Pearson correlation coefficient between the electricity meter parameter sequence corresponding to any node and the electricity meter parameter sequence corresponding to the neighboring nodes connected to that node, and use the mean Pearson correlation coefficient as the functional similarity between the electricity meter and its neighboring electricity meters.

7. The anomaly detection method for a multi-functional energy meter based on deep learning according to claim 1, characterized in that, The optimal abnormal threshold for each function is obtained by adjusting the initial abnormal threshold based on the importance score. The method is as follows: Iterate through all the electricity meters in the power grid and sum the importance scores of all the electricity meters for a certain function as the total importance weight of that function. Calculate the average importance weight of all functions; The difference between the total importance weight and the average importance weight is taken as the deviation of the importance weight; Use the product of the preset adjustment coefficient and the deviation as the threshold adjustment amount; The sum of the threshold adjustment amount and the initial abnormal threshold is used as the optimal abnormal threshold for this function. This process is repeated for all functions to obtain the optimal abnormal threshold for each function.

8. The anomaly detection method for a multi-functional energy meter based on deep learning according to claim 1, characterized in that, The initial outlier threshold is set by obtaining historical data statistics and adding a certain number of standard deviations to the average value.

9. The anomaly detection method for a multi-functional energy meter based on deep learning according to claim 1, characterized in that, After normalizing the importance, the importance score is obtained by using the ratio of the importance of each node to the sum of the importance of its neighbors to the sum of the importance of all nodes in the graph and their neighbors as the importance score of a node to its neighbors.

10. A multifunctional energy meter anomaly detection device based on deep learning, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the deep learning-based multi-functional energy meter anomaly detection method according to any one of claims 1-9.

Images