A deep learning-based power consumption information collection terminal abnormal behavior prediction and detection method

Abstract

This invention discloses a method for predicting and detecting abnormal behavior of electricity information collection terminals based on deep learning, comprising the following steps: acquiring and preprocessing the terminal operation data of a distribution substation to construct a spatiotemporal dataset; constructing a graph structure based on topological wiring relationships and inputting it into a spatiotemporal graph convolutional network to extract spatiotemporal correlation features; introducing a power balance model based on Kirchhoff's laws and line loss conservation as a physical constraint to optimize and solve the spatiotemporal correlation features, obtaining topological residuals; inputting the spatiotemporal correlation features and topological residuals into a time-series memory network for trend evolution to generate prediction residuals; performing power balance verification on real-time transient electrical parameters to obtain detection residuals; combining the spatial statistical deviation between the prediction residuals and the detection residuals to achieve anomaly identification and output anomaly interception commands. This invention achieves collaborative analysis of prediction and real-time detection of abnormal terminal behavior, improving the accuracy and stability of anomaly identification.

Description

Technical Field

[0001] This invention relates to the field of power system monitoring and intelligent analysis technology, and in particular to a method for predicting and detecting abnormal behavior of electricity consumption information collection terminals based on deep learning. Background Technology

[0002] In the power consumption information collection system of a distribution substation, a large number of power consumption information collection terminals are distributed between the user side and the concentrator. These terminals continuously collect operational data such as three-phase voltage, current, active power, and reactive power, and upload them to the main station system via a communication network to support services such as power metering, load analysis, and operation monitoring. In actual operation, the large number of terminals and the complex operating environment result in data exhibiting obvious temporal fluctuations and spatial topological correlations. This makes it difficult for data from a single moment or a single terminal to fully reflect the overall operating status. Abnormal terminal behavior often exhibits characteristics of gradual evolution and cross-node correlation propagation.

[0003] Existing technologies often employ threshold-based judgment or single-point statistical feature analysis to detect abnormal terminal behavior, such as alarm judgment based on voltage over-limit, current surges, or power anomalies. However, these methods rely solely on observation data at local moments and lack the ability to model the historical time-series evolution of the terminal, making it difficult to characterize the dynamic evolution of abnormal behavior. Furthermore, while some detection methods based on machine learning or deep learning introduce time-series modeling mechanisms, they typically ignore the electrical topology constraints between concentrators and terminals within the distribution substation, leading to discrepancies between model outputs and actual power transmission patterns. In addition, existing technologies often fail to effectively integrate the consistency between predicted and real-time detection results during anomaly detection, resulting in insufficient stability of anomaly identification results and a tendency for misjudgments or missed detections under load fluctuations or communication disturbances.

[0004] Therefore, how to provide a method for predicting and detecting abnormal behavior of electricity information collection terminals based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for predicting and detecting abnormal behavior of electricity information collection terminals based on deep learning. This invention introduces a spatiotemporal graph convolution and physical conservation constraint optimization architecture, analyzes the topology of the distribution area and performs temporal memory network trend evolution, combines transient electrical parameter balance verification to inversely deduce the spatial statistical deviation of the double residuals, constructs an anomaly identification mechanism based on topological residuals and prediction residuals, and uses anomaly interception commands to perform cross-layer blocking and control at the distribution area terminal, realizing dynamic verification of electricity collection data and accurate interception of abnormal behavior. It has the advantages of high anomaly identification accuracy, strong physical constraint defense and low interception response latency.

[0006] A method for predicting and detecting abnormal behavior of an electricity consumption information collection terminal based on deep learning according to an embodiment of the present invention includes the following steps: S1. Read the raw data from each power consumption information collection terminal in the distribution area, perform preprocessing on the raw data, and group and map the preprocessed data according to the spatiotemporal feature dimension to construct a spatiotemporal dataset. S2. Obtain the physical wiring topology between the concentrator and each electricity information collection terminal in the distribution substation area and map it into a topology graph structure. Input the topology graph structure and the spatiotemporal dataset into the spatiotemporal graph convolutional network and output the spatiotemporal correlation feature matrix. S3. Construct a charge balance model based on Kirchhoff's laws and the line loss conservation equation. Embed the charge balance model as a physical constraint into the loss function of the spatiotemporal graph convolutional network, perform constraint solution on the spatiotemporal correlation feature matrix, and output the topological residual matrix. S4. Input the spatiotemporal correlation feature matrix and the topological residual matrix into the temporal memory network to perform trend evolution and output the prediction residual; S5. Real-time capture of data reported within the distribution area, extraction of transient electrical parameters from the data, input into the power balance model for algebraic sum verification, and output of detection residuals; S6. Calculate the spatial statistical deviation between the corresponding nodes of the detection residual and the prediction residual. When the spatial statistical deviation exceeds the preset threshold, it will determine that the corresponding electricity information collection terminal has triggered abnormal behavior and output an abnormal interception command.

[0007] Optionally, the raw data includes three-phase real-time voltage, three-phase real-time current, active power, reactive power, power factor, cumulative electricity consumption in the current period, terminal operating status word, topology hop count, and communication signal-to-noise ratio; the preprocessing includes performing linear interpolation of missing values, singular value removal and replacement, and global data normalization on the raw data.

[0008] Optionally, S1 specifically includes: S11. Obtain the raw data uploaded by each electricity information collection terminal in the distribution area, and group and map the raw data according to the terminal identifier to form a data set indexed by the terminal. S12. Parse the fields of various types of data in the dataset, extract the corresponding time identifiers, spatial node identifiers and business type identifiers, and construct the initial spatiotemporal mapping unit based on the time identifiers and spatial node identifiers; S13. Sort each initial spatiotemporal mapping unit in chronological order, and perform sliding segmentation on the sorted data based on a preset time window to generate multi-scale time series segments. S14. Map each multi-scale time series segment to its corresponding topological location according to the spatial node affiliation relationship, and construct a spatiotemporal data queue with dual indexes of time and space dimensions. S15. Perform standardized reconstruction processing on the spatiotemporal data queue, map different types of data to a unified feature vector space, and output them in batches according to time windows to form a spatiotemporal dataset.

[0009] Optionally, S2 specifically includes: S21. Obtain the physical wiring relationship between the concentrator and each power consumption information collection terminal in the distribution area, parse the node set and the connection relationship between the nodes, and construct the topology node set and topology edge set. S22. Establish the corresponding topological graph structure based on the set of topological nodes and the set of topological edges, and establish the mapping relationship between each topological node and the feature vector in the spatiotemporal dataset according to the node identifier; S23. Input the topology graph structure into the spatiotemporal graph convolutional network, perform neighborhood aggregation calculation on the feature vectors of adjacent nodes according to the node connection relationship in the topology graph structure, and complete the node feature update according to the preset number of propagation layers. S24. Based on the update time order, perform time dimension recursive calculation on the node features after neighborhood aggregation is completed, and continuously update the node features at each time step in time order. S25. After completing the time recursive calculation, rearrange the node features according to the topological node identifier, and output the corresponding spatiotemporal correlation feature matrix according to the topological association relationship between each node.

[0010] Optionally, the spatiotemporal graph convolutional network includes: Obtain the node identifiers of each electricity consumption information collection terminal and their corresponding physical wiring topology, and map the node identifiers with connection relationships to the adjacency matrix; Extract the operating data of each electricity consumption information collection terminal and map it to the corresponding node feature sequence according to the node identifier. At the same time, obtain the timestamp and feature dimension information of each node feature sequence. Based on the node feature sequence, sort the nodes in ascending order according to their timestamps to determine the update order of each node feature at different time steps; Based on the node connection relationships in the adjacency matrix, the sorted node feature sequence is written into the adjacency propagation sequence, and the adjacency propagation offset of the current node is calculated by accumulating the feature lengths of the adjacent nodes. Within each propagation round, the node feature sequence is copied node by node according to the adjacency propagation offset, and the features of adjacent nodes are propagated along the topological relationship; The node feature sequence after propagation and updating is sorted by time based on the timestamp, and the sorted node feature sequence is recursively updated. The recursively updated node feature sequence is remapped to the graph structure space according to the node identifier and then assembled and arranged in the order of the nodes.

[0011] Optionally, S3 specifically includes: S31. Extract the node identifiers corresponding to the spatiotemporal correlation feature matrix, obtain the voltage, current and electrical parameters of each node, and map them according to the node identifier and timestamp to form a node electrical parameter sequence. S32. Calculate the current inflow and outflow of each node based on the node electrical parameter sequence, and construct the power transmission path in combination with the distribution area topology connection relationship; S33. Based on Kirchhoff's laws, the current inflow and outflow at nodes are constrained by conservation, and the power transmission path is constrained by the line loss conservation equation. A power balance model is then jointly constructed. S34. The power balance model calculates the node current imbalance and the path power loss deviation, and performs fusion processing to generate physical constraint terms. S35. Embed physical constraints into the loss function of the spatiotemporal graph convolutional network, drive the loss function to optimize the spatiotemporal correlation feature matrix, and output the node power balance deviation value. S36. Rearrange the power balance deviation values ​​according to the node topology relationship to generate the topology residual matrix.

[0012] Optionally, the line loss conservation equation includes: Each transmission branch is defined as a power transmission path based on the topological connection relationship, and the input and output nodes of the corresponding path are obtained. The injected electrical power of the input node and the consumed electrical power of the output node in each power transmission path are obtained, and the corresponding power transmission amount is calculated based on voltage, current and power factor. Based on the law of conservation of electrical energy, a power balance expression is constructed, such that the difference between the input and output electrical energy in each power transmission path is equal to the line loss. Based on line resistance parameters and current transmission characteristics, line loss is parameterized and mapped to topology path length and node connection relationships. The power balance expressions for each power transmission path are aggregated and calculated at the regional level across the entire station to form a system of line loss conservation equations that satisfy topological constraints.

[0013] Optionally, S4 specifically includes: S41. Group and sort the spatiotemporal correlation feature matrix according to the node identifier to form a node temporal feature sequence, and perform index consistency alignment with the topological residual matrix. S42. Rearrange the topological residual matrix according to the node identifier and establish a correspondence with the node time series feature sequence in the node dimension and time step dimension to generate an aligned residual sequence. S43. Align and concatenate the spatiotemporal correlation feature sequence and the topological residual sequence according to time steps to form a joint time-series input sequence; S44. Based on the temporal memory network, the joint temporal input sequence is recursively updated, and a time gating mechanism is introduced during the time step propagation process to selectively retain and update historical state information, forming a trend evolution sequence. S45. Perform dimensional expansion and mapping reconstruction on the trend evolution sequence to generate the prediction residuals of the corresponding nodes.

[0014] Optionally, S5 specifically includes: S51. Capture the real-time operation data reported within the distribution radio area, and parse the real-time operation data to extract the transient voltage, current and power parameters of the corresponding acquisition terminal. S52. Based on the node identifier, the transient electrical parameters are mapped to form a sequence of transient electrical parameters that corresponds one-to-one with the node; S53. Input the transient electrical parameter sequence into the charge balance model, and perform algebraic sum and verification calculations based on Kirchhoff current relations and line loss conservation constraints; S54. Calculate the power balance deviation value of each node based on the algebraic sum and verification results, and use the power balance deviation value as the node error amount; S55. Output the node error in a structured manner to generate the detection residual corresponding to each electricity information collection terminal.

[0015] Optionally, S6 specifically includes: S61. Match the predicted residuals and the detected residuals according to the node identifiers to form a node-level residual sequence; S62. Align the node-level residual sequence with time windows in the time dimension so that the prediction residual and the detection residual can establish a correspondence at the same time step. S63. Calculate the difference between the prediction residual and the detection residual at each time step, and use the difference to characterize the node spatial statistical bias value. S64. Combining the node topology connection relationship, the node spatial statistical deviation value is corrected for neighborhood consistency to obtain the node spatial deviation index. S65. When the spatial deviation index of any node exceeds the preset threshold, it is determined that an abnormal state has occurred in the application electricity information collection terminal, and an abnormal interception command is output.

[0016] The beneficial effects of this invention are: (1) By constructing a spatiotemporal grouping mapping mechanism for the operation data of the distribution area terminal, and combining the physical wiring topology relationship between the concentrator and the acquisition terminal to form a graph structure input spatiotemporal graph convolutional network, the node features are jointly expressed in the spatial topology and time dimension, thereby enhancing the structured representation capability of the terminal status features and improving the separability and stability of abnormal behavior features.

[0017] (2) By introducing a power balance constraint mechanism based on Kirchhoff's laws and the relationship between line loss conservation, physical consistency verification of spatiotemporal correlation features is performed and topological residuals are generated, so that data-driven feature learning is integrated with the physical laws of the power system, effectively suppressing error accumulation and false correlation in pure data modeling, and improving the robustness and interpretability of the model to abnormal states.

[0018] (3) By constructing a dual-channel collaborative comparison mechanism of prediction residual and detection residual, a spatial statistical deviation criterion is established between the time evolution prediction result and the real-time transient electrical parameter verification result, so as to realize the cross-time scale consistency identification of abnormal terminal behavior, thereby improving the sensitivity of abnormal detection and early warning capability. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the overall process of a deep learning-based method for predicting and detecting abnormal behavior of electricity consumption information collection terminals proposed in this invention. Figure 2 This is a spatiotemporal graph convolutional modeling structure diagram of the abnormal behavior prediction and detection method for electricity information collection terminals based on deep learning proposed in this invention; Figure 3 This is a flowchart illustrating the anomaly detection and judgment process of an abnormal behavior prediction and detection method for an electricity information collection terminal based on deep learning, as proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figures 1-3 A method for predicting and detecting abnormal behavior of electricity consumption information collection terminals based on deep learning, comprising the following steps: S1. Read the raw data from each power consumption information collection terminal in the distribution area, perform preprocessing on the raw data, and group and map the preprocessed data according to the spatiotemporal feature dimension to construct a spatiotemporal dataset. S2. Obtain the physical wiring topology between the concentrator and each electricity information collection terminal in the distribution substation area and map it into a topology graph structure. Input the topology graph structure and the spatiotemporal dataset into the spatiotemporal graph convolutional network and output the spatiotemporal correlation feature matrix. S3. Construct a charge balance model based on Kirchhoff's laws and the line loss conservation equation. Embed the charge balance model as a physical constraint into the loss function of the spatiotemporal graph convolutional network, perform constraint solution on the spatiotemporal correlation feature matrix, and output the topological residual matrix. S4. Input the spatiotemporal correlation feature matrix and the topological residual matrix into the temporal memory network to perform trend evolution and output the prediction residual; S5. Real-time capture of data reported within the distribution area, extraction of transient electrical parameters from the data, input into the power balance model for algebraic sum verification, and output of detection residuals; S6. Calculate the spatial statistical deviation between the corresponding nodes of the detection residual and the prediction residual. When the spatial statistical deviation exceeds the preset threshold, it will determine that the corresponding electricity information collection terminal has triggered abnormal behavior and output an abnormal interception command.

[0022] In this embodiment, the raw data includes three-phase real-time voltage, three-phase real-time current, active power, reactive power, power factor, cumulative electricity consumption in the current period, terminal operating status word, topology hop count, and communication signal-to-noise ratio; the preprocessing includes performing linear interpolation of missing values, singular value removal and replacement, and global data normalization on the raw data.

[0023] In this embodiment, S1 specifically includes: S11. Obtain the raw operating data periodically reported by each power consumption information collection terminal in the distribution area, and collect and process the data according to the unique identifier of the terminal to construct a data set indexed by the terminal identifier. S12. Perform field parsing on the data set, extract the time identifier and spatial node identifier from each data record, and establish an initial spatiotemporal mapping unit based on the time identifier and spatial node identifier, so that each terminal data is bound to its corresponding station topology location. S13. Sort all initial spatiotemporal mapping units in ascending order according to timestamps, and perform sliding segmentation on the continuous data sequence according to a preset time window to generate time series segments covering different time scales. The length of the preset time window is preferably 24 to 96 consecutive sampling steps, corresponding to a 6-hour to 24-hour data observation window under a 15-minute sampling interval. The sliding step size is preferably set to 1 or 4 sampling steps. S14. Map each time series segment to the corresponding node position in the distribution area topology according to the spatial node affiliation relationship, and construct a spatiotemporal data queue containing time dimension index and spatial dimension index. S15. Apply a uniform scale to the spatiotemporal data queue, map electrical parameters of different dimensions to the same feature vector space, and output them in batches according to time windows to form a spatiotemporal dataset.

[0024] In this embodiment, S2 specifically includes: S21. Obtain the physical wiring relationship between the concentrator and each power consumption information collection terminal in the distribution area, parse and form a node set and the connection relationship between nodes, and construct a topology node set and a topology edge set based on the connection relationship. S22. Generate the corresponding topological graph structure based on the set of topological nodes and the set of topological edges, and establish a mapping relationship between topological nodes and feature vectors in the spatiotemporal dataset based on node identifiers, so that the topological position of the node and the running features are bound one-to-one. S23. Input the topological graph structure into the spatiotemporal graph convolutional network, and perform neighborhood information aggregation processing on the feature vectors of adjacent nodes according to the topological connection relationship. The node feature is updated through multi-layer propagation iteration to characterize the electrical correlation influence between nodes. S24. Based on the timestamp order, perform time-dimensional recursive calculation on the updated node features, and continuously evolve and update the node features at each time step according to the time sequence to depict the temporal change trend of the node's running state. S25. The recursively updated node features are rearranged according to the topological node identifiers and structured in combination with the topological association relationships between nodes to generate the corresponding spatiotemporal association feature matrix.

[0025] In this embodiment, the spatiotemporal graph convolutional network includes: Obtain the node identifiers and physical wiring topology of each terminal, map the node identifiers with electrical connections to the corresponding addresses in the matrix, construct an initial adjacency matrix, and determine the network topology of the transformer area. Extract the running data of each terminal, map it to the corresponding node feature sequence according to the node identifier, and obtain its timestamp and feature dimension information to complete the spatiotemporal alignment of the data. The number of time slots of the node feature sequence stored continuously in the buffer is preferably 5 to 12 consecutive time slots. The node feature sequences are globally sorted in ascending order based on timestamps to determine the update order at different time steps and establish the pipeline parallel update timing. Based on the initial adjacency matrix, the sorted node feature sequence is linearly written into the dynamic buffer and reconstructed into an adjacency propagation sequence. Address accumulation is performed according to the number of neighboring nodes and feature length of the current node, and the adjacency propagation offset of the current node is output. Within a propagation round, the computing unit performs parallel copying of node features on the adjacent propagation sequence according to the adjacent propagation offset, spatially aggregates the features of adjacent nodes along the physical wiring topology, and outputs the propagated updated node feature sequence. The number of propagation iteration layers in the spatial dimension is preferably 2, 3, or 4 integer layers. Based on the timestamp, the node feature sequence after propagation and update is sorted by time, and the sorted sequence is recursively updated along the time axis to extract the load evolution characteristics of a single terminal. The recursively updated node feature sequences are remapped back to the graph structure space according to the node identifiers, and linearly arranged according to the node hierarchy. The spatiotemporal correlation feature matrix is ​​then output in a structured manner in the output register.

[0026] In this embodiment, S3 specifically includes: S31. Extract the node identifiers corresponding to each terminal in the spatiotemporal correlation feature matrix, read the voltage, current and electrical parameters of each terminal, and map them according to the node identifier and timestamp to form a node electrical parameter sequence in the dynamic storage area. S32. Read the node electrical parameter sequence, calculate the current inflow and outflow of each terminal at each sampling point through algebraic summation, and construct the power transmission path in the network space in combination with the physical wiring topology. S33. Based on Kirchhoff's laws, the current inflow and outflow are subject to algebraic sum-zero conservation constraints, and the power transmission path is subject to loss constraints in conjunction with the line loss conservation equation, thus jointly constructing a power balance model. S34. Calculate the current imbalance at each node using the power balance model, extract the path power loss deviation of the power transmission path, and perform vector summation and fusion of the current imbalance and path power loss deviation to output the physical constraint terms. S35. Embed the physical constraint term as a penalty factor into the loss function of the spatiotemporal graph convolutional network, and optimize the spatiotemporal correlation feature matrix by backpropagation error update to output the node power balance deviation value. S36. Based on the topology hierarchy order in the physical wiring topology, the memory addresses of the node power balance deviation values ​​are rearranged, and the topology residual matrix is ​​generated in the output register in a structured manner.

[0027] In this embodiment, the line loss conservation equation includes: Based on the physical wiring topology, each transmission branch in the distribution substation is defined as a power transmission path. The connection endpoints of the transmission branch are analyzed to obtain the input and output nodes of the corresponding power transmission path. Extract the injected electrical power of the input node and the consumed electrical power of the output node in each power transmission path, read the voltage, current and power factor in the node electrical parameter sequence, and calculate the corresponding power transmission amount by multiplication and addition. Based on the law of conservation of electrical energy, a power balance expression is constructed. The power transmission amount is then incorporated into the power balance expression so that the difference between the input and output electrical energy in each power transmission path is equal to the line loss. Read the line resistance parameters and current transmission characteristics of the transmission branch, express the line loss in a parameterized manner, and perform addressing association mapping between the line loss and the path length and node connection relationship of the power transmission path. The power balance expressions for each power transmission path are summed and calculated at the regional level across the entire network. The topological constraints of the network space are locked, and a system of line loss conservation equations that satisfy the topological constraints is generated. The power balance model is then subjected to loss constraints.

[0028] In this embodiment, S4 specifically includes: S41. Group and sort the spatiotemporal correlation feature matrix according to the node identifier, output the node temporal feature sequence, and read the topological residual matrix to ensure that the node indexes of the two are aligned in a consistent manner. S42. Rearrange the topological residual matrix according to the node identifier, and establish an addressing correspondence between the rearranged matrix and the node time-series feature sequence in the node dimension and time step dimension to generate an aligned residual sequence. S43. Perform memory address alignment and splicing on the node time-series feature sequence and the aligned residual sequence at the same time step to construct a joint time-series input sequence in the merging buffer; S44. Input the joint time-series input sequence into the time-series memory network to perform recursive updates, and introduce a time gating mechanism during the time step propagation process of the time-series memory network to selectively retain and clear the historical state register information, and output the trend evolution sequence. The feature dimension of the hidden layer register inside the time-series memory network is preferably set to an integer multiple of 64, 128 or 256. S45. Perform dimensional expansion and memory mapping reconstruction on the trend evolution sequence to eliminate the time axis dimension, and generate the prediction residuals corresponding to each electricity information acquisition terminal in the output register in a structured manner.

[0029] In this embodiment, S5 specifically includes: S51. Capture and parse the real-time operating data reported by each power information collection terminal in the distribution area, and extract the transient voltage, current and power parameters of the power information collection terminal. S52. Based on the node identifier, the transient voltage, current and power electrical parameters are mapped to form a transient electrical parameter sequence that corresponds one-to-one with each node in the dynamic storage area. S53. Input the transient electrical parameter sequence into the charge balance model, and perform algebraic sum and verification calculations based on Kirchhoff current relations and line loss conservation equations. S54. Output the unbalanced electricity of each node based on the verification results calculated by algebra and verification, and use the unbalanced electricity as the node error quantity. S55. The node error is structured and the detection residuals corresponding to each power consumption information acquisition terminal are structured and output in the output register.

[0030] In this embodiment, S6 specifically includes: S61. Match the predicted residuals and the detected residuals according to the node identifiers to form a node-level residual sequence in the video memory; S62. Align the node-level residual sequence with time windows in the time dimension so that the prediction residual and the detection residual can establish a correspondence at the same time step. S63. The algebraic absolute difference between the prediction residual and the detection residual at each time step is calculated by subtraction, and the algebraic absolute difference is used to characterize the node spatial statistical deviation value. S64. Based on the physical wiring topology, perform neighborhood consistency mean sliding correction on the node spatial statistical deviation value and output the node spatial deviation index. S65. When any node's spatial deviation index exceeds the preset threshold 3, 5, or 10 times consecutively, the electricity information collection terminal is determined to be in an abnormal state, and an abnormal interception command is hard-triggered in the output register.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to a power consumption information collection system for a heavily loaded and complex grid distribution substation within the jurisdiction of a power supply bureau. This substation includes one 400kVA centralized distribution transformer, one dedicated transformer acquisition terminal, 12 public transformer concentrators, and 350 smart meters. In this scenario, each terminal reports data every 15 minutes, with a three-phase voltage resolution of 16 bits, a data packet size ranging from 512 bits to 2048 bits, and a total available feature space dimension of 4096 dimensions. Due to interference from a complex electromagnetic environment, the communication carrier signal-to-noise ratio varies drastically between 10dB and 45dB, resulting in sudden data loss and transient anomalies in the raw data acquisition. Traditional methods relying on fixed electricity theft thresholds are insufficient to effectively intercept these anomalies.

[0032] The network's underlying pipeline directly reads the 350×96-dimensional spatiotemporal correlation feature matrix and topological residual matrix. The temporal memory network configures the hidden layer dimension to 128 dimensions, performs the same 96-time-step sliding segmentation and alignment concatenation on the matrix, and outputs the prediction residual. Subsequently, the system captures real-time transient electrical parameter sequences and inputs them into the power balance model to perform algebraic sum verification, outputting a detection residual containing 350 terminal error points. When the spatial deviation index of any terminal exceeds the threshold of 75 minutes for 5 consecutive times, the system triggers an anomaly interception command to accurately locate the abnormal terminal.

[0033] Experimental data shows that within the detection period using this method, the success rate of intercepting electricity theft from dedicated transformers reached 98.5%, the fault identification rate of single-phase reverse current acquisition in distributed photovoltaic inverters increased to 95.2%, the transient error in daily line loss calculation for the entire distribution area decreased from 4.2% to 0.6%, and the overall system's abnormal behavior false negative rate was diluted from 12.4% to 0.8%. As shown in the table below, the abnormal behavior detection method based on physical constraints of power balance proposed in this invention, when dealing with complex flow directions such as distributed photovoltaic reverse grid connection and commercial peak loads, achieves high-precision, zero-malfunction convergence of the state residual matrix compared to traditional methods, enhancing the synergistic effect of refined and compliant power consumption management in the distribution network. This demonstrates the advanced nature, reliability, and industrial practical value of this technical solution in the fields of power acquisition information security and abnormal detection.

[0034] Table 1. Comparison of Substation-Level Performance of Different Anomaly Detection Methods

[0035] The results in Table 1 demonstrate that the method of this invention embeds physical topological constraints such as Kirchhoff's law and line loss conservation into the spatiotemporal features. This not only significantly improves the electricity theft interception rate and reverse flow fault identification rate to 98.5% and 95.2% respectively, and reduces the false alarm rate to 0.8%, but also breaks the limitations of the traditional black-box model, endowing the system with topological physical traceability capabilities. Furthermore, under the parallel optimization of the underlying pipeline, the average detection latency is compressed to 3.1 seconds, achieving high determinism and high interpretability of anomaly interception in massive concurrent scenarios.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting and detecting abnormal behavior of electricity consumption information collection terminals based on deep learning, characterized in that, Includes the following steps: S1. Read the raw data from each power consumption information collection terminal in the distribution area, perform preprocessing on the raw data, and group and map the preprocessed data according to the spatiotemporal feature dimension to construct a spatiotemporal dataset. S2. Obtain the physical wiring topology between the concentrator and each electricity information collection terminal in the distribution substation area and map it into a topology graph structure. Input the topology graph structure and the spatiotemporal dataset into the spatiotemporal graph convolutional network and output the spatiotemporal correlation feature matrix. S3. Construct a charge balance model based on Kirchhoff's laws and the line loss conservation equation. Embed the charge balance model as a physical constraint into the loss function of the spatiotemporal graph convolutional network, perform constraint solution on the spatiotemporal correlation feature matrix, and output the topological residual matrix. S4. Input the spatiotemporal correlation feature matrix and the topological residual matrix into the temporal memory network to perform trend evolution and output the prediction residual; S5. Real-time capture of data reported within the distribution area, extraction of transient electrical parameters from the data, input into the power balance model for algebraic sum verification, and output of detection residuals; S6. Calculate the spatial statistical deviation between the corresponding nodes of the detection residual and the prediction residual. When the spatial statistical deviation exceeds the preset threshold, it will determine that the corresponding electricity information collection terminal has triggered abnormal behavior and output an abnormal interception command.

2. The method for predicting and detecting abnormal behavior of electricity information collection terminals based on deep learning according to claim 1, characterized in that, The raw data includes three-phase real-time voltage, three-phase real-time current, active power, reactive power, power factor, cumulative electricity consumption in the current period, terminal operating status word, topology hop count, and communication signal-to-noise ratio; the preprocessing includes performing linear interpolation of missing values, singular value removal and replacement, and global data normalization on the raw data.

3. The method for predicting and detecting abnormal behavior of electricity information collection terminals based on deep learning according to claim 1, characterized in that, S1 specifically includes: S11. Obtain the raw data uploaded by each electricity information collection terminal in the distribution area, and group and map the raw data according to the terminal identifier to form a data set indexed by the terminal. S12. Parse the fields of various types of data in the dataset, extract the corresponding time identifiers, spatial node identifiers and business type identifiers, and construct the initial spatiotemporal mapping unit based on the time identifiers and spatial node identifiers; S13. Sort each initial spatiotemporal mapping unit in chronological order, and perform sliding segmentation on the sorted data based on a preset time window to generate multi-scale time series segments. S14. Map each multi-scale time series segment to its corresponding topological location according to the spatial node affiliation relationship, and construct a spatiotemporal data queue with dual indexes of time and space dimensions. S15. Perform standardized reconstruction processing on the spatiotemporal data queue, map different types of data to a unified feature vector space, and output them in batches according to time windows to form a spatiotemporal dataset.

4. The method for predicting and detecting abnormal behavior of electricity information collection terminals based on deep learning according to claim 1, characterized in that, S2 specifically includes: S21. Obtain the physical wiring relationship between the concentrator and each power consumption information collection terminal in the distribution area, parse the node set and the connection relationship between the nodes, and construct the topology node set and topology edge set. S22. Establish the corresponding topological graph structure based on the set of topological nodes and the set of topological edges, and establish the mapping relationship between each topological node and the feature vector in the spatiotemporal dataset according to the node identifier; S23. Input the topology graph structure into the spatiotemporal graph convolutional network, perform neighborhood aggregation calculation on the feature vectors of adjacent nodes according to the node connection relationship in the topology graph structure, and complete the node feature update according to the preset number of propagation layers. S24. Based on the update time order, perform time dimension recursive calculation on the node features after neighborhood aggregation is completed, and continuously update the node features at each time step in time order. S25. After completing the time recursive calculation, rearrange the node features according to the topological node identifier, and output the corresponding spatiotemporal correlation feature matrix according to the topological association relationship between each node.

5. The method for predicting and detecting abnormal behavior of a power consumption information collection terminal based on deep learning according to claim 4, characterized in that, The spatiotemporal graph convolutional network includes: Obtain the node identifiers of each electricity consumption information collection terminal and their corresponding physical wiring topology, and map the node identifiers with connection relationships to the adjacency matrix; Extract the operating data of each electricity consumption information collection terminal and map it to the corresponding node feature sequence according to the node identifier. At the same time, obtain the timestamp and feature dimension information of each node feature sequence. Based on the node feature sequence, sort the nodes in ascending order according to their timestamps to determine the update order of each node feature at different time steps; Based on the node connection relationships in the adjacency matrix, the sorted node feature sequence is written into the adjacency propagation sequence, and the adjacency propagation offset of the current node is calculated by accumulating the feature lengths of the adjacent nodes. Within each propagation round, the node feature sequence is copied node by node according to the adjacency propagation offset, and the features of adjacent nodes are propagated along the topological relationship; The node feature sequence after propagation and updating is sorted by time based on the timestamp, and the sorted node feature sequence is recursively updated. The recursively updated node feature sequence is remapped to the graph structure space according to the node identifier and then assembled and arranged in the order of the nodes.

6. The method for predicting and detecting abnormal behavior of a power consumption information collection terminal based on deep learning according to claim 1, characterized in that, S3 specifically includes: S31. Extract the node identifiers corresponding to the spatiotemporal correlation feature matrix, obtain the voltage, current and electrical parameters of each node, and map them according to the node identifier and timestamp to form a node electrical parameter sequence. S32. Calculate the current inflow and outflow of each node based on the node electrical parameter sequence, and construct the power transmission path in combination with the distribution area topology connection relationship; S33. Based on Kirchhoff's laws, the current inflow and outflow at the nodes are constrained by conservation, and the power transmission path is constrained by the line loss conservation equation. A power balance model is then jointly constructed. S34. The power balance model calculates the node current imbalance and the path power loss deviation, and performs fusion processing to generate physical constraint terms. S35. Embed physical constraints into the loss function of the spatiotemporal graph convolutional network, drive the loss function to optimize the spatiotemporal correlation feature matrix, and output the node power balance deviation value. S36. Rearrange the power balance deviation values ​​according to the node topology relationship to generate the topology residual matrix.

7. The method for predicting and detecting abnormal behavior of a power consumption information collection terminal based on deep learning according to claim 6, characterized in that, The line loss conservation equation includes: Each transmission branch is defined as a power transmission path based on the topological connection relationship, and the input and output nodes of the corresponding path are obtained. The injected electrical power of the input node and the consumed electrical power of the output node in each power transmission path are obtained, and the corresponding power transmission amount is calculated based on voltage, current and power factor. Based on the law of conservation of electrical energy, a power balance expression is constructed, such that the difference between the input and output electrical energy in each power transmission path is equal to the line loss. Based on line resistance parameters and current transmission characteristics, line loss is parameterized and mapped to topology path length and node connection relationships. The power balance expressions for each power transmission path are aggregated and calculated at the regional level across the entire station to form a system of line loss conservation equations that satisfy topological constraints.

8. The method for predicting and detecting abnormal behavior of a power consumption information collection terminal based on deep learning according to claim 1, characterized in that, S4 specifically includes: S41. Group and sort the spatiotemporal correlation feature matrix according to the node identifier to form a node temporal feature sequence, and perform index consistency alignment with the topological residual matrix. S42. Rearrange the topological residual matrix according to the node identifier and establish a correspondence with the node time series feature sequence in the node dimension and time step dimension to generate an aligned residual sequence. S43. Align and concatenate the spatiotemporal correlation feature sequence and the topological residual sequence according to time steps to form a joint time-series input sequence; S44. Based on the temporal memory network, the joint temporal input sequence is recursively updated, and a time gating mechanism is introduced during the time step propagation process to selectively retain and update historical state information, forming a trend evolution sequence. S45. Perform dimensional expansion and mapping reconstruction on the trend evolution sequence to generate the prediction residuals of the corresponding nodes.

9. The method for predicting and detecting abnormal behavior of a power consumption information collection terminal based on deep learning according to claim 1, characterized in that, S5 specifically includes: S51. Capture the real-time operation data reported within the distribution radio area, and parse the real-time operation data to extract the transient voltage, current and power parameters of the corresponding acquisition terminal. S52. Based on the node identifier, the transient electrical parameters are mapped to form a sequence of transient electrical parameters that corresponds one-to-one with the node; S53. Input the transient electrical parameter sequence into the charge balance model, and perform algebraic sum and verification calculations based on Kirchhoff current relations and line loss conservation constraints; S54. Calculate the power balance deviation value of each node based on the algebraic sum and verification results, and use the power balance deviation value as the node error amount; S55. Output the node error in a structured manner to generate the detection residual corresponding to each electricity information collection terminal.

10. The method for predicting and detecting abnormal behavior of a power consumption information collection terminal based on deep learning according to claim 1, characterized in that, S6 specifically includes: S61. Match the predicted residuals and the detected residuals according to the node identifiers to form a node-level residual sequence; S62. Align the node-level residual sequence with time windows in the time dimension so that the prediction residual and the detection residual can establish a correspondence at the same time step. S63. Calculate the difference between the prediction residual and the detection residual at each time step, and use the difference to characterize the node spatial statistical bias value. S64. Combining the node topology connection relationship, the node spatial statistical deviation value is corrected for neighborhood consistency to obtain the node spatial deviation index. S65. When the spatial deviation index of any node exceeds the preset threshold, it is determined that an abnormal state has occurred in the application electricity information collection terminal, and an abnormal interception command is output.

Images