Artificial intelligence-based line loss cause analysis and positioning method, system and medium

Abstract

The application provides a line loss cause analysis and positioning method, system and medium based on artificial intelligence, relating to the technical field of power system management, which comprises the following steps: when the total line loss rate of a target substation exceeds a preset line loss threshold, an excitation sequence is issued to the total meter, circuit breaker and user meter of the target substation, corresponding transient response data are obtained, and then a dynamic topology graph of the target substation is constructed; then the topology structure is optimized and the abnormal connection point is detected in combination with a graph neural network and an attention mechanism. Accordingly, the downstream associated user meters are determined, and the electricity abnormality score is calculated based on the load data. The steepest front wave signal is injected into the meter with the highest score to obtain the reflection characteristic parameters. The Double DQN reinforcement learning algorithm is used for iterative analysis to determine the fault point, line loss cause type and confidence, so that efficient and accurate line loss cause positioning and analysis are realized. The application realizes efficient and accurate investigation of line loss causes.

Description

Technical Field

[0001] This invention relates to the field of power system management technology, and more specifically, to a method, system, and medium for analyzing and locating the causes of line losses based on artificial intelligence. Background Technology

[0002] Line loss is a crucial economic indicator in power system operation. Excessive line loss not only leads to energy waste and economic losses but may also imply power equipment failures, safety hazards, and other problems, affecting the stable and reliable operation of the power system. Currently, once line loss is confirmed, it is usually necessary to rely on manual experience and simple statistical analysis to analyze and locate the causes of the line loss.

[0003] In related technologies, when facing complex and uncertain topological environments such as old residential areas and urban-rural fringe areas, when a transformer substation experiences high line loss, maintenance personnel can only locate the cause of the high line loss by checking each household and cutting off power in sections. This is not only time-consuming and labor-intensive, but also cannot pinpoint the fault point of a branch line or individual user in the case of complex wiring, thus affecting the effectiveness of troubleshooting the cause of line loss. Summary of the Invention

[0004] The problem addressed by this invention is how to improve the accuracy and efficiency of troubleshooting the causes of line loss.

[0005] To address the aforementioned problems, this invention provides a method, system, and medium for analyzing and locating the causes of line loss based on artificial intelligence.

[0006] In a first aspect, the present invention provides a method for analyzing and locating the causes of line loss based on artificial intelligence, comprising:

[0007] When the bus loss rate of the target transformer area exceeds the preset line loss threshold, an excitation sequence is sent to the main meter, circuit breaker and user meter of the target transformer area to obtain the transient response data corresponding to the main meter, the circuit breaker and the user meter respectively.

[0008] Based on the transient response data corresponding to the master meter, the circuit breaker, and the user meter, a dynamic topology map of the target distribution area is constructed.

[0009] Based on the dynamic topology graph, combined with graph neural networks and attention mechanisms, dynamic optimization of the topology structure and anomaly detection are performed to determine abnormal connection points;

[0010] Based on the location information of the abnormal connection point, determine the downstream associated user's electricity meter of the abnormal connection point;

[0011] Based on the load data of each associated user's electricity meter within a preset time period, determine the electricity consumption anomaly score of each associated user's electricity meter;

[0012] A steep-front traveling wave signal is injected into the electricity meter of the associated user with the highest electricity consumption anomaly score to obtain the reflection peak time difference, reflection coefficient, and attenuation gradient of the associated user's electricity meter.

[0013] Based on the reflection peak time difference, the reflection coefficient, and the attenuation gradient, the reflection path is iterated using the Double DQN reinforcement learning algorithm to obtain the fault point of the target transformer area, the line loss cause type of the fault point, and the confidence level corresponding to the line loss cause type.

[0014] Optionally, when the bus loss rate of the target distribution area exceeds a preset line loss threshold, an excitation sequence is sent to the main meter, circuit breaker, and user meter of the target distribution area to obtain transient response data corresponding to the main meter, the circuit breaker, and the user meter, respectively, including:

[0015] The bus loss rate of the target transformer area is determined based on the actual line loss value of each branch line in the target transformer area.

[0016] When the bus loss rate exceeds the preset line loss threshold, the excitation sequence is generated and synchronously sent to the main meter of the target transformer area, the circuit breaker of each branch line, and the user meter through a parallel communication protocol.

[0017] Simultaneously, the transient voltage and transient current of the main meter, the circuit breaker, and the user meter are acquired during a first preset acquisition time period; the first preset acquisition time period is set from a first preset time point before the excitation sequence is issued to a second preset time point after the excitation sequence is issued.

[0018] The transient response data is obtained based on the transient voltage and transient current of the master meter, the circuit breaker, and the user meter within the preset acquisition time period.

[0019] Optionally, constructing a dynamic topology map of the target distribution area based on the transient response data corresponding to the master meter, the circuit breaker, and the user meter includes:

[0020] Feature extraction is performed on the transient response data to obtain the key feature vector of the transient response data;

[0021] Construct a topology node system for the target transformer area, wherein the main meter is used as a first-level node, the circuit breaker of each branch line is used as a second-level node, and the user meter is used as a third-level node.

[0022] Based on the key feature vectors corresponding to the master meter, the circuit breaker, and the user meter, the electrical correlation strength between each node in the topology node system is determined.

[0023] The electrical association strength is input into a preset lightweight pulse-graph attention network. The key feature vector is used as the initial node embedding vector, and the electrical association strength is used as the initial edge weight. The node embedding vector is iteratively updated through a multi-head attention mechanism to obtain a dynamic adjacency matrix.

[0024] The dynamic adjacency matrix is ​​subjected to threshold judgment and sparsification processing to generate the dynamic topology graph containing the dynamic connection relationships between the first-level nodes, the second-level nodes and the third-level nodes.

[0025] Optionally, the step of dynamically optimizing the topology and detecting anomalies based on the dynamic topology graph, combined with graph neural networks and attention mechanisms, to determine abnormal connection points includes:

[0026] The dynamic topology map is compared with the initial preset topology of the target station area edge-to-edge to obtain the connection confidence and deviation of each edge in the dynamic topology map;

[0027] Based on the connection confidence and the deviation, abnormal candidate edges in the dynamic topology graph are determined and a set of abnormal candidate edges is formed.

[0028] For each of the abnormal candidate edges in the set of abnormal candidate edges, the abnormal probability of the abnormal candidate edge is determined based on the historical electrical parameter difference characteristics of the two end nodes of the abnormal candidate edge, the echo signal similarity, and the geographical coordinate offset.

[0029] The candidate edges of anomalies whose anomalies are greater than or equal to a preset anomaly threshold are designated as anomaly connection edges.

[0030] The abnormal connection point is determined based on the abnormal connection edge using the gated propagation mechanism of the graph neural network.

[0031] Optionally, determining the downstream associated user meter of the abnormal connection point based on its location information includes:

[0032] Starting from the abnormal connection point, perform a breadth-first search along the outgoing edges of the dynamic topology graph until the third-level node is reached, thus obtaining a set of third-level nodes.

[0033] The user meters corresponding to all the third-level nodes in the third-level node set are regarded as the associated user meters.

[0034] Optionally, determining the electricity consumption anomaly score for each associated user's electricity meter based on the load data of each associated user's electricity meter within a preset time period includes:

[0035] Based on the daily frozen electricity consumption, power curve, and transient response data of the associated user's electricity meter within the preset time period, a multi-dimensional load vector of the associated user's electricity meter is generated within the preset time period.

[0036] By using a temporal convolutional network, feature extraction is performed on the multidimensional load vector to obtain the load dynamic features of the multidimensional load vector;

[0037] Based on the load dynamic characteristics, the electricity consumption deviation of the associated user's electricity meter is determined;

[0038] The abnormal electricity consumption score is obtained based on the electricity consumption deviation and the preset line loss influence coefficient.

[0039] Optionally, injecting a steep-front traveling wave signal into the associated user's meter with the highest electricity consumption anomaly score to obtain the reflection peak time difference, reflection coefficient, and attenuation gradient of the associated user's meter includes:

[0040] A steep-front traveling wave signal is injected at the position in front of the electricity meter of the associated user with the highest electricity consumption anomaly score;

[0041] Acquire the traveling wave reflection waveform of the associated user's electricity meter during the second preset acquisition time period; the second preset acquisition time period is set from the third preset time point before the injection of the steep leading edge traveling wave signal to the fourth preset time point after the injection of the steep leading edge traveling wave signal.

[0042] Feature extraction is performed on the traveling wave reflection waveform to obtain the reflection peak time difference, the reflection coefficient, and the attenuation gradient.

[0043] Optionally, the step of iterating the reflection path using the Double DQN reinforcement learning algorithm based on the reflection peak time difference, the reflection coefficient, and the attenuation gradient to obtain the fault point of the target transformer area, the line loss cause type of the fault point, and the confidence level corresponding to the line loss cause type includes:

[0044] A multidimensional state vector is formed based on the reflection peak time difference, the reflection coefficient, and the attenuation gradient;

[0045] The multidimensional state vector is input into the Double DQN network. The fault location is taken as the action and the minimization of reflected energy loss is taken as the immediate reward. The Q-value network is iteratively updated to obtain the location coordinates of the fault point.

[0046] Based on the location coordinates, determine the dual-mode features corresponding to the location coordinates; the dual-mode features include the transient impedance-frequency fingerprint and the reflected energy attenuation gradient of the fault point;

[0047] The transient impedance-frequency fingerprint and the reflected energy attenuation gradient are input into a multimodal fusion deep learning model for prediction, thereby obtaining the line loss cause type of the fault point and the confidence level corresponding to the line loss cause type.

[0048] Secondly, the present invention provides an artificial intelligence-based system for analyzing and locating the causes of line loss, comprising:

[0049] The excitation sequence sending module is used to send an excitation sequence to the main meter, circuit breaker and user meter of the target transformer area when the bus loss rate of the target transformer area exceeds the preset line loss threshold, so as to obtain the transient response data corresponding to the main meter, the circuit breaker and the user meter respectively.

[0050] The topology construction module is used to construct a dynamic topology map of the target transformer area based on the transient response data corresponding to the master meter, the circuit breaker, and the user meter, respectively.

[0051] The anomaly detection module is used to dynamically optimize the topology and detect anomalies based on the dynamic topology graph, combined with graph neural networks and attention mechanisms, and to determine abnormal connection points.

[0052] The associated user location module is used to determine the associated user's electricity meter downstream of the abnormal connection point based on the location information of the abnormal connection point.

[0053] The score evaluation module is used to determine the electricity consumption anomaly score of each associated user's electricity meter based on the load data of each associated user's electricity meter within a preset time period;

[0054] The traveling wave injection module is used to inject a steep-front traveling wave signal into the associated user's electricity meter with the highest electricity consumption anomaly score, so as to obtain the reflection peak time difference, reflection coefficient and attenuation gradient of the associated user's electricity meter;

[0055] The analysis and localization module is used to obtain the fault point of the target transformer area, the line loss cause type of the fault point, and the confidence level corresponding to the line loss cause type by iterating the reflection path through the Double DQN reinforcement learning algorithm based on the reflection peak time difference, the reflection coefficient, and the attenuation gradient.

[0056] Thirdly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the artificial intelligence-based method for analyzing and locating line loss causes as described above.

[0057] This invention presents an AI-based method, system, and medium for analyzing and locating line loss causes. By triggering the transient response of equipment through an excitation sequence, it can rapidly acquire dynamic data of key equipment (main meter, circuit breaker, and user meters) within a target distribution area. This data reflects the immediate state of the equipment under the action of the excitation signal, providing a foundation for subsequent analysis. Simultaneous acquisition of response data from the main meter, circuit breaker, and user meters ensures data comprehensiveness. This invention not only covers the overall state of the distribution area but also includes detailed information at the branch and user levels, providing rich data support for subsequent topology construction and anomaly detection. By constructing a dynamic topology diagram, it can reflect the electrical connections and equipment status within the distribution area in real time. Compared to traditional static topology diagrams, dynamic topology diagrams can capture real-time interactions and changes between devices, more accurately reflecting the actual operating status. For example, in complex topology environments such as old residential areas and urban-rural fringe areas, dynamic topology diagrams can clearly display branch lines and user connection relationships, providing a clear network structure foundation for subsequent anomaly detection. Simultaneously, an artificial intelligence combination consisting of graph neural networks (GNNs) and attention mechanisms is used for dynamic optimization and anomaly detection. GNNs model nodes (devices) and edges (connections) in the topology graph, capturing complex dependencies between nodes. By learning patterns and rules in the topology, GNNs can effectively identify potential anomalies. Attention mechanisms focus on key nodes and connections in the topology graph, automatically identifying areas with significant impact on line loss. By assigning higher weights to these key components, attention mechanisms improve the accuracy and efficiency of anomaly detection. Combining GNNs and attention mechanisms allows for dynamic adjustment of the weights of nodes and edges in the topology graph, optimizing the topology and further enhancing the accuracy of anomaly detection. By identifying downstream user meters at anomalous connection points, the investigation scope is narrowed from the entire transformer area to specific branch lines or user groups, reducing the number of devices requiring further analysis and improving investigation efficiency. The identification of associated user meters is based on the topology, ensuring the targeted and accurate nature of the analysis. Subsequent analysis of these associated meters can more directly reflect the impact range of the anomalous connection points. By analyzing user meter load data and calculating electricity anomaly scores, the degree of anomaly for each user's meter can be quantified, providing data support for subsequent fault location. Load data analysis within a preset time period can capture dynamic changes in user electricity consumption behavior, helping to identify continuous or intermittent abnormal electricity consumption behaviors. Steep-front traveling wave signals can propagate and reflect on power lines; by analyzing the reflection peak time difference, reflection coefficient, and attenuation gradient, the location of the fault point can be accurately measured. The location method based on traveling wave signals features high accuracy and fast response. Parameters such as reflection coefficient and attenuation gradient reflect the electrical characteristics of the fault point, providing crucial information for subsequent fault cause analysis.The Double DQN algorithm iteratively optimizes reflection paths to dynamically adjust fault location strategies, improving accuracy and reliability. The algorithm not only determines the location of faults but also identifies the type of line loss based on reflection parameters and provides corresponding confidence levels, achieving AI-powered line loss cause analysis. This provides maintenance personnel with clear fault diagnosis results, facilitating subsequent repair and optimization measures.

[0058] In summary, the transient response data obtained from the excitation sequence of this invention forms the basis for constructing a dynamic topology graph, which in turn provides network structure support for subsequent anomaly detection. The dynamic topology graph provides the connection relationships between devices, upon which graph neural networks and attention mechanisms perform anomaly detection and dynamic optimization, further improving detection accuracy. The location information of the abnormal connection point is used to determine its downstream associated user meters, narrowing the investigation scope and improving investigation efficiency. The load data of associated user meters is used to calculate the electricity consumption anomaly score, providing data support for subsequent fault location. The user meter with the highest electricity consumption anomaly score is injected with a traveling wave signal, and the precise location and characteristics of the fault point are obtained through reflection parameters. The reflection parameters obtained from the traveling wave signal are iteratively analyzed using the Double DQN algorithm to ultimately determine the fault point, the type of line loss cause, and its confidence level. This invention, through the organic combination of the above technical features, achieves efficient and accurate investigation of line loss causes. It collectively solves the problems of low efficiency and insufficient accuracy of traditional methods in complex topology environments. This invention can not only quickly locate the fault point, but also clarify the specific cause and confidence level of line loss, providing strong technical support for the stable operation of the power system and energy saving and loss reduction. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating the artificial intelligence-based method for analyzing and locating the causes of line loss according to an embodiment of the present invention.

[0060] Figure 2 This is a schematic diagram of the structure of the artificial intelligence-based line loss cause analysis and location system according to an embodiment of the present invention. Detailed Implementation

[0061] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Although some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the accompanying drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0062] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0063] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to"; the term "based on" means "at least partially based on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; and the term "optionally" means "optional embodiments". Definitions of other terms will be given in the following description. It should be noted that the concepts of "first," "second," etc., mentioned in this invention are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0064] It should be noted that the terms "one" and "more" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0065] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties. The collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0066] Combination Figure 1 As shown in the figure, an embodiment of the present invention provides a method for analyzing and locating the causes of line loss based on artificial intelligence, including:

[0067] When the bus loss rate of the target transformer area exceeds the preset line loss threshold, an excitation sequence is sent to the main meter, circuit breaker and user meter of the target transformer area to obtain the transient response data corresponding to the main meter, the circuit breaker and the user meter respectively.

[0068] Specifically, when the bus loss rate of the target distribution area exceeds a preset line loss threshold, the system automatically triggers the excitation sequence distribution process. The excitation sequence is a specially designed signal that can stimulate transient responses in power equipment (such as main meters, circuit breakers, and user meters). The implementation is as follows: The excitation sequence uses a high-frequency pulse signal, the frequency and amplitude of which are optimized according to the characteristics of the power equipment to ensure effective equipment response while avoiding interference with normal power transmission. The excitation sequence is simultaneously sent to the main meters, circuit breakers, and user meters in the target distribution area via power line carrier communication (PLC) technology. PLC technology uses power lines as the communication medium, enabling fast and stable signal transmission. Upon receiving the excitation sequence, the main meters, circuit breakers, and user meters immediately generate transient responses. The built-in sensors in these devices collect the response signals in real time and transmit the response data back to the system via PLC technology.

[0069] In a preferred embodiment of the present invention, the line loss rate refers to the ratio of electrical energy lost during the transmission and distribution of electrical energy in a power system to the input electrical energy within a certain period of time. It is usually expressed as a percentage and its calculation formula is as follows:

[0070] Line loss rate = (Input energy - Output energy) / Input energy × 100%

[0071] Input energy refers to the total electrical energy input to the target distribution area from the upstream power grid, typically measured by the area's master meter (e.g., the incoming line master meter). Output energy refers to the total electrical energy allocated to each user within the target distribution area, usually calculated from the accumulated data of user meters. Input energy data is collected from the master meter (incoming line master meter) installed in the target distribution area. Output energy data is calculated by collecting energy data from all user meters within the target distribution area and summing the output energy. Master meter data acquisition is performed periodically using power line carrier communication (PLC) or wireless communication modules to read the master meter's energy data. User meter data acquisition is also performed periodically using PLC or wireless communication modules to read the energy data from each user's meter and aggregate the data into the system. Data synchronization ensures that the data acquisition time of the master meter and user meters is synchronized to avoid calculation errors caused by time differences.

[0072] The input energy is the energy data read from the master meter.

[0073] Input electrical energy = ;

[0074] The output energy is calculated by summing the energy data from all users' meters.

[0075] Output electrical energy = ;

[0076] in, This refers to the number of electricity meters used by users within the transformer area.

[0077] Calculate the line loss rate using the above formula:

[0078] Line loss rate = ×100%;

[0079] Based on historical data and actual operational experience, a reasonable line loss threshold is preset. For example, assuming the preset line loss threshold is 5%, the system periodically (e.g., hourly) calculates the line loss rate of the target area and compares it with the preset line loss threshold. If the calculated line loss rate is greater than the preset line loss threshold (e.g., the calculated result is 7%), it is determined that the total line loss rate of the target area exceeds the preset line loss threshold. If the line loss rate is less than or equal to the preset line loss threshold, it is considered that the line loss is within the normal range, and real-time monitoring continues.

[0080] Suppose that the total meter in the target distribution area records an input energy of 10,000 kWh during a certain period, and there are 100 users in the distribution area, whose meters record a total output energy of 9,500 kWh.

[0081] Line loss rate = ) × 100% = 5%;

[0082] If the preset line loss threshold is 3%, then if the calculated result of 5% is greater than the preset threshold of 3%, the line loss cause analysis and location process will be triggered; if the preset line loss threshold is 6%, then if the calculated result of 5% is less than the preset threshold of 6%, the line loss cause analysis and location process will not be triggered.

[0083] Using the above method, the system can accurately determine whether the bus loss rate of the target transformer area exceeds the preset line loss threshold, and accordingly initiate subsequent line loss cause analysis and location processes. This real-time monitoring and judgment mechanism can promptly detect abnormal line loss situations, providing support for the efficient operation and maintenance of the power system.

[0084] Based on the transient response data corresponding to the master meter, the circuit breaker, and the user meter, a dynamic topology map of the target transformer area is constructed.

[0085] Specifically, after receiving transient response data from the main meter, circuit breakers, and user meters, the system enters the dynamic topology map construction phase. The collected transient response data undergoes filtering and noise reduction to improve data quality. Algorithms such as wavelet transform are used to remove high-frequency noise and retain valid signals. Based on the time delay and amplitude changes of the transient response data, the locations of the main meter, circuit breakers, and user meters are identified. The response time differences of each device reflect its relative position in the topology. By analyzing the correlation of response signals between devices, the connection relationships between them are determined. Correlation analysis algorithms, such as the Pearson correlation coefficient, are used to determine the electrical connection paths between devices. The identified nodes and connection relationships are integrated into a dynamic topology map. This topology map graphically displays the electrical network structure within the distribution area, reflecting the connection status and interaction relationships between devices in real time.

[0086] Based on the dynamic topology graph, combined with graph neural networks and attention mechanisms, dynamic optimization of the topology structure and anomaly detection are performed to identify abnormal connection points.

[0087] Specifically, based on the dynamic topology graph, the system further combines a graph neural network (GNN) and an attention mechanism for dynamic optimization of the topology and anomaly detection. The dynamic topology graph is input into the graph neural network model. The GNN learns the features of nodes (devices) and edges (connections) to capture complex patterns and regularities in the topology. An attention mechanism is introduced on top of the GNN to assign higher weights to key nodes and connections in the topology graph. The attention mechanism automatically identifies areas with a significant impact on line loss by calculating the weight coefficients between nodes. The combination of GNN and the attention mechanism dynamically optimizes the topology, adjusting the weights of nodes and edges to improve the accuracy of anomaly detection. Through iterative optimization, the system can accurately identify abnormal connection points.

[0088] Based on the location information of the abnormal connection point, the downstream associated user's electricity meter of the abnormal connection point is determined.

[0089] Specifically, based on the location information of the abnormal connection point, the system further determines the downstream associated user meters, extracts the downstream path of the abnormal connection point from the dynamic topology map, and identifies all user meters directly or indirectly connected to the abnormal connection point. Based on the topology path, user meters related to the abnormal connection point are selected. The electricity consumption behavior of these meters may be affected by the abnormal connection point and is the focus of subsequent analysis.

[0090] Based on the load data of each associated user's electricity meter within a preset time period, determine the electricity consumption anomaly score for each associated user's electricity meter.

[0091] Specifically, based on the load data of each associated user's electricity meter within a preset time period, an electricity consumption anomaly score is calculated. Load data for the associated user's electricity meter within the preset time period, including parameters such as current, voltage, and power, is obtained from the power system database. Statistical analysis methods, such as Z-Score standardization, are used to calculate the electricity consumption anomaly score for each user's electricity meter. The anomaly score reflects the degree of deviation between the user's electricity meter's load data and normal electricity consumption behavior. The associated user's electricity meters are then ranked according to their electricity consumption anomaly scores, with the meter having the highest score considered the most likely source of the anomaly.

[0092] A steep-front traveling wave signal is injected into the electricity meter of the associated user with the highest electricity consumption anomaly score to obtain the reflection peak time difference, reflection coefficient, and attenuation gradient of the associated user's electricity meter.

[0093] Specifically, a steep-front traveling wave signal is injected into the electricity meter of the associated user with the highest power consumption anomaly score, and its reflection parameters are obtained. Using a dedicated traveling wave injection device, the steep-front traveling wave signal is injected into the target user's electricity meter. The traveling wave signal has a fast rise edge, enabling it to propagate and reflect on the power line. During its propagation on the power line, the traveling wave signal will be reflected when it encounters a fault point. The system collects the reflected signal using a high-precision sensor and calculates the reflection peak time difference, reflection coefficient, and attenuation gradient. The reflection peak time difference reflects the distance between the fault point and the injection point; the reflection coefficient and attenuation gradient reflect the electrical characteristics of the fault point, such as impedance changes and insulation status.

[0094] Based on the reflection peak time difference, the reflection coefficient, and the attenuation gradient, the reflection path is iterated using the Double DQN reinforcement learning algorithm to obtain the fault point of the target transformer area, the line loss cause type of the fault point, and the confidence level corresponding to the line loss cause type.

[0095] Specifically, based on reflection parameters, the reflection path is iterated using the Double DQN reinforcement learning algorithm to ultimately determine the fault point, line loss cause type, and confidence level of the target transformer area. The reflection peak time difference, reflection coefficient, and attenuation gradient are used as input parameters to initialize the Double DQN algorithm. The algorithm learns the relationship between these parameters and the fault point location to build a prediction model. Through the iterative process of reinforcement learning, the algorithm continuously adjusts the reflection path to optimize the fault point location accuracy. In each iteration, the algorithm selects the optimal path based on the current reflection parameters and updates the model parameters. After multiple iterations, the algorithm finally determines the specific location of the fault point and identifies the line loss cause type based on the characteristics of the reflection parameters, such as equipment aging, insulation damage, or electricity theft. Through multiple simulations and verifications, the confidence level of fault point location and cause classification is evaluated, providing reliable diagnostic results for maintenance personnel.

[0096] This invention presents an AI-based method, system, and medium for analyzing and locating line loss causes. By triggering the transient response of equipment through an excitation sequence, it can rapidly acquire dynamic data of key equipment (main meter, circuit breaker, and user meters) within a target distribution area. This data reflects the immediate state of the equipment under the action of the excitation signal, providing a foundation for subsequent analysis. Simultaneous acquisition of response data from the main meter, circuit breaker, and user meters ensures data comprehensiveness. This invention not only covers the overall state of the distribution area but also includes detailed information at the branch and user levels, providing rich data support for subsequent topology construction and anomaly detection. By constructing a dynamic topology diagram, it can reflect the electrical connections and equipment status within the distribution area in real time. Compared to traditional static topology diagrams, dynamic topology diagrams can capture real-time interactions and changes between devices, more accurately reflecting the actual operating status. For example, in complex topology environments such as old residential areas and urban-rural fringe areas, dynamic topology diagrams can clearly display branch lines and user connection relationships, providing a clear network structure foundation for subsequent anomaly detection. Simultaneously, an artificial intelligence combination consisting of graph neural networks (GNNs) and attention mechanisms is used for dynamic optimization and anomaly detection. GNNs model nodes (devices) and edges (connections) in the topology graph, capturing complex dependencies between nodes. By learning patterns and rules in the topology, GNNs can effectively identify potential anomalies. Attention mechanisms focus on key nodes and connections in the topology graph, automatically identifying areas with significant impact on line loss. By assigning higher weights to these key components, attention mechanisms improve the accuracy and efficiency of anomaly detection. Combining GNNs and attention mechanisms allows for dynamic adjustment of the weights of nodes and edges in the topology graph, optimizing the topology and further enhancing the accuracy of anomaly detection. By identifying downstream user meters at anomalous connection points, the investigation scope is narrowed from the entire transformer area to specific branch lines or user groups, reducing the number of devices requiring further analysis and improving investigation efficiency. The identification of associated user meters is based on the topology, ensuring the targeted and accurate nature of the analysis. Subsequent analysis of these associated meters can more directly reflect the impact range of the anomalous connection points. By analyzing user meter load data and calculating electricity anomaly scores, the degree of anomaly for each user's meter can be quantified, providing data support for subsequent fault location. Load data analysis within a preset time period can capture dynamic changes in user electricity consumption behavior, helping to identify continuous or intermittent abnormal electricity consumption behaviors. Steep-front traveling wave signals can propagate and reflect on power lines; by analyzing the reflection peak time difference, reflection coefficient, and attenuation gradient, the location of the fault point can be accurately measured. The location method based on traveling wave signals features high accuracy and fast response. Parameters such as reflection coefficient and attenuation gradient reflect the electrical characteristics of the fault point, providing crucial information for subsequent fault cause analysis.The Double DQN algorithm iteratively optimizes reflection paths to dynamically adjust fault location strategies, improving accuracy and reliability. The algorithm not only determines the location of faults but also identifies the type of line loss based on reflection parameters and provides corresponding confidence levels, achieving AI-powered line loss cause analysis. This provides maintenance personnel with clear fault diagnosis results, facilitating subsequent repair and optimization measures.

[0097] In summary, the transient response data obtained from the excitation sequence of this invention forms the basis for constructing a dynamic topology graph, which in turn provides network structure support for subsequent anomaly detection. The dynamic topology graph provides the connection relationships between devices, upon which graph neural networks and attention mechanisms perform anomaly detection and dynamic optimization, further improving detection accuracy. The location information of the abnormal connection point is used to determine its downstream associated user meters, narrowing the investigation scope and improving investigation efficiency. The load data of associated user meters is used to calculate the electricity consumption anomaly score, providing data support for subsequent fault location. The user meter with the highest electricity consumption anomaly score is injected with a traveling wave signal, and the precise location and characteristics of the fault point are obtained through reflection parameters. The reflection parameters obtained from the traveling wave signal are iteratively analyzed using the Double DQN algorithm to ultimately determine the fault point, the type of line loss cause, and its confidence level. This invention, through the organic combination of the above technical features, achieves efficient and accurate investigation of line loss causes. It collectively solves the problems of low efficiency and insufficient accuracy of traditional methods in complex topology environments. This invention can not only quickly locate the fault point, but also clarify the specific cause and confidence level of line loss, providing strong technical support for the stable operation of the power system and energy saving and loss reduction.

[0098] Optionally, when the bus loss rate of the target distribution area exceeds a preset line loss threshold, an excitation sequence is sent to the main meter, circuit breaker, and user meter of the target distribution area to obtain transient response data corresponding to the main meter, the circuit breaker, and the user meter, respectively, including:

[0099] The bus loss rate of the target transformer area is determined based on the actual line loss value of each branch line in the target transformer area.

[0100] When the bus loss rate exceeds the preset line loss threshold, the excitation sequence is generated and synchronously sent to the main meter of the target transformer area, the circuit breaker of each branch line, and the user meter through a parallel communication protocol.

[0101] Simultaneously, the transient voltage and transient current of the main meter, the circuit breaker, and the user meter are acquired during a first preset acquisition time period; the first preset acquisition time period is set from a first preset time point before the excitation sequence is issued to a second preset time point after the excitation sequence is issued.

[0102] The transient response data is obtained based on the transient voltage and transient current of the master meter, the circuit breaker, and the user meter within the preset acquisition time period.

[0103] Specifically, the overall line loss rate of the target distribution area is first determined based on the actual line loss value of each branch line within the target area. Specifically, monitoring equipment installed on each branch line collects input and output energy data in real time. Based on this data, the line loss value of each branch line, i.e., the branch line loss rate, is calculated. Then, the line loss values ​​of all branch lines are weighted and averaged to obtain the overall line loss rate of the target distribution area. The weighted average is calculated based on the proportion of input energy of each branch line to the total input energy, ensuring that the overall line loss rate accurately reflects the line loss situation of the entire distribution area. When the calculated overall line loss rate exceeds a preset line loss threshold, the system triggers the subsequent excitation sequence generation and distribution process. The preset line loss threshold is set based on historical data and actual operating experience; for example, it can be set to 5%. The system periodically calculates the overall line loss rate of the target distribution area and compares it with the preset line loss threshold. If the overall line loss rate exceeds this threshold, it indicates that there may be abnormal line loss in the distribution area, requiring further analysis and location.

[0104] At this point, the system generates an excitation sequence, which is a specially designed high-frequency pulse signal. Its frequency and amplitude are optimized according to the characteristics of the power equipment to ensure effective equipment response while avoiding interference with normal power transmission. The duration of the excitation sequence is typically in the millisecond range, such as 10 milliseconds. Then, through a parallel communication protocol, such as power line carrier communication (PLC) or a wireless communication module, the excitation sequence is synchronously transmitted to the main meter of the target distribution area, the circuit breakers of each branch line, and the user's meter. The parallel communication protocol ensures that all devices receive the excitation signal at the same time, thereby guaranteeing the synchronization and accuracy of data acquisition.

[0105] Before and after the excitation sequence is issued, the system simultaneously acquires the transient voltage and transient current of the main meter, circuit breaker, and user meters during a first preset acquisition time period. The first preset acquisition time period begins at a first preset time point before the excitation sequence is issued and continues until a second preset time point after the excitation sequence is issued. For example, the first preset time point can be set to 110ms before issuance, and the second preset time point can be set to 200ms before issuance. This acquisition time period covers the critical moments before and after the excitation sequence is issued, enabling the capture of the dynamic response of the equipment under excitation.

[0106] During the data acquisition process, high-precision sensors installed on the main meter, circuit breakers, and user meters collect transient voltage and current data in real time. The acquired data includes the amplitude, phase, and time series information of voltage and current. This data undergoes preprocessing after acquisition, such as filtering and noise reduction, to remove high-frequency noise and interference signals, retaining only valid signals. The preprocessed data is then used to generate transient response data. Specifically, the response characteristics of each device under the excitation sequence are calculated, such as the peak value, rate of change, and time delay of voltage and current. These characteristics are ultimately integrated into transient response data, providing accurate foundational data for subsequent dynamic topology construction and anomaly detection.

[0107] This invention achieves rapid response and accurate monitoring of abnormal line loss conditions by accurately calculating the bus loss rate of the target transformer area and triggering the synchronous distribution of excitation sequences and transient response data acquisition when the rate exceeds a preset threshold. First, the bus loss rate is calculated based on the actual line loss value of the branch lines, accurately reflecting the overall line loss status of the transformer area and avoiding errors that may arise from a single monitoring point, thereby improving the reliability of line loss assessment. When the bus loss rate exceeds the threshold, the system immediately generates and synchronously distributes the excitation sequence to the main meter, circuit breakers, and user meters. This parallel communication mechanism ensures that all devices receive signals at the same time, greatly improving the synchronicity and accuracy of data acquisition. Simultaneously, within a preset acquisition time period before and after the excitation sequence distribution, the system acquires transient voltage and current data. This process not only captures the dynamic changes of the equipment under excitation but also removes noise interference through preprocessing, retaining key features and providing a high-quality data foundation for subsequent analysis. Ultimately, by integrating these transient response data, the system can provide accurate information for the construction of dynamic topology maps and anomaly detection, thereby enabling efficient analysis and precise location of the causes of line losses, significantly improving the efficiency and reliability of power system operation and maintenance.

[0108] Optionally, constructing a dynamic topology map of the target distribution area based on the transient response data corresponding to the master meter, the circuit breaker, and the user meter includes:

[0109] Feature extraction is performed on the transient response data to obtain the key feature vector of the transient response data;

[0110] Construct a topology node system for the target transformer area, wherein the main meter is used as a first-level node, the circuit breaker of each branch line is used as a second-level node, and the user meter is used as a third-level node.

[0111] Based on the key feature vectors corresponding to the master meter, the circuit breaker, and the user meter, the electrical correlation strength between each node in the topology node system is determined.

[0112] The electrical association strength is input into a preset lightweight pulse-graph attention network. The key feature vector is used as the initial node embedding vector, and the electrical association strength is used as the initial edge weight. The node embedding vector is iteratively updated through a multi-head attention mechanism to obtain a dynamic adjacency matrix.

[0113] The dynamic adjacency matrix is ​​subjected to threshold judgment and sparsification processing to generate the dynamic topology graph containing the dynamic connection relationships between the first-level nodes, the second-level nodes and the third-level nodes.

[0114] Specifically, when constructing the dynamic topology map of the target distribution area, features are first extracted from the transient response data corresponding to the main meter, circuit breakers, and user meters to obtain key feature vectors. By analyzing the voltage and current change characteristics in the transient response data, key information reflecting the equipment status and connection relationships is extracted, forming a key feature vector for each device. Subsequently, a topology node system for the target distribution area is constructed, with the main meter as a first-level node, the circuit breakers of each branch line as second-level nodes, and the user meters as third-level nodes, forming a hierarchical node structure. Based on these key feature vectors, the electrical correlation strength between each node in the topology node system is further determined, i.e., by analyzing the similarity or correlation between feature vectors, the tightness of the electrical connection between nodes is quantified. Specifically, a similarity measurement method is used to calculate the similarity between two nodes. By calculating the cosine of the angle between two vectors, the closer the value is to 1, the higher the similarity. Next, these electrical correlation strengths are input into a pre-defined lightweight pulse-graph attention network, using the key feature vectors as initial node embedding vectors and the electrical correlation strength as initial edge weights. Through a multi-head attention mechanism, the network can dynamically update node embedding vectors, thereby more accurately capturing complex relationships between nodes and generating a dynamic adjacency matrix. This dynamic adjacency matrix reflects the dynamic connections between nodes, but may contain some redundant information. Therefore, thresholding and sparsification are applied to the dynamic adjacency matrix to remove connections below a set threshold, resulting in a more concise and accurate dynamic topology graph. Specifically, a threshold θ is first set to filter out significant electrical connections. For each element A in the dynamic adjacency matrix... ij If A ij If the connection value is greater than a threshold θ, the connection is retained; otherwise, it is set to zero. Sparsity processing, by preserving the strongest connections and highlighting critical electrical paths, generates a dynamic topology graph containing the dynamic connections between first-level, second-level, and third-level nodes. The resulting dynamic topology graph clearly shows the dynamic connections between the main meter, circuit breakers, and user meters, providing an accurate network structure foundation for subsequent line loss cause analysis and anomaly detection.

[0115] In this embodiment of the invention, by extracting features from transient response data, key feature vectors can be filtered out from a large amount of raw data. This process not only effectively reduces data dimensionality but also retains the core information that best reflects the device status and connection relationships, providing a high-quality data foundation for subsequent topology map construction and significantly improving data processing efficiency and analysis accuracy. Secondly, a hierarchical topology node system is constructed, with the main meter, circuit breaker, and user meters serving as first-level, second-level, and third-level nodes, respectively. This hierarchical structure clearly reflects the hierarchical relationships and information flow paths within the transformer area. This hierarchical design not only simplifies the analysis of complex networks but also allows the topology map to more intuitively display the electrical connection relationships within the transformer area, providing a clear framework for subsequent anomaly detection and line loss analysis. Furthermore, determining the electrical correlation strength between nodes based on key feature vectors allows for precise quantification of the tightness of connections between devices. This quantification process provides a scientific basis for the construction of dynamic topology maps, enabling the topology map to accurately reflect the electrical connection status within the transformer area and providing an accurate network structure foundation for subsequent analysis. Electrical association strength is input into a lightweight pulse-graph attention network, with key feature vectors as initial node embedding vectors and electrical association strength as initial edge weights. A multi-head attention mechanism iteratively updates the node embedding vectors to generate a dynamic adjacency matrix. This process not only dynamically captures complex relationships between nodes but also focuses on key connections through the attention mechanism, further improving the accuracy and reliability of the topology graph. The lightweight network design also ensures computational efficiency, making it suitable for real-time or near-real-time power system monitoring scenarios. Finally, the dynamic adjacency matrix undergoes thresholding and sparsification to generate a concise and accurate dynamic topology graph. This processing step effectively removes redundant connections, avoiding analysis difficulties caused by an overly complex topology graph, while preserving key connections to ensure the topology graph accurately reflects the dynamic electrical connection status within the transformer area. The sparsified topology graph is not only easier to analyze but also reduces computational resource consumption, improving the overall system performance.

[0116] By employing feature extraction, hierarchical node architecture construction, electrical correlation strength quantification, and dynamic optimization of graph attention networks, efficient and accurate construction of dynamic topology maps for target transformer areas was achieved. This not only provides an accurate network structure foundation for line loss cause analysis and anomaly detection but also significantly improves the intelligence level of power system operation and maintenance, enhances the diagnostic capability for line loss problems in complex topology environments, and provides strong technical support for the stable operation and energy conservation and loss reduction of power systems.

[0117] Optionally, the step of dynamically optimizing the topology and detecting anomalies based on the dynamic topology graph, combined with graph neural networks and attention mechanisms, to determine abnormal connection points includes:

[0118] The dynamic topology map is compared with the initial preset topology of the target station area edge-to-edge to obtain the connection confidence and deviation of each edge in the dynamic topology map;

[0119] Based on the connection confidence and the deviation, abnormal candidate edges in the dynamic topology graph are determined and a set of abnormal candidate edges is formed.

[0120] For each of the abnormal candidate edges in the set of abnormal candidate edges, the abnormal probability of the abnormal candidate edge is determined based on the historical electrical parameter difference characteristics of the two end nodes of the abnormal candidate edge, the echo signal similarity, and the geographical coordinate offset.

[0121] The candidate edges of anomalies whose anomalies are greater than or equal to a preset anomaly threshold are designated as anomaly connection edges.

[0122] The abnormal connection point is determined based on the abnormal connection edge using the gated propagation mechanism of the graph neural network.

[0123] Specifically, during dynamic optimization and anomaly detection of the topology, the dynamic topology map is first compared edge-to-edge with the initial preset topology of the target area. This comparison involves analyzing the differences between each edge in the dynamic topology map and its corresponding edge in the initial preset topology, calculating the connection confidence and deviation of each edge. The connection confidence reflects the reliability of the edge maintaining its initial connection state in the dynamic topology map, while the deviation indicates the degree of deviation between the edge and the initial topology. Based on the connection confidence and deviation, the system further identifies candidate anomaly edges in the dynamic topology map and forms a set of candidate anomaly edges. Candidate anomaly edges are those with high deviation or low connection confidence, resulting from abnormal connections caused by equipment failure, connection errors, or topology changes. For each candidate anomaly edge in the set, the probability of anomaly is determined by comprehensively considering the historical electrical parameter differential characteristics of its two endpoints, echo signal similarity, and geographical coordinate offset. Historical electrical parameter difference features reflect the changes in electrical parameters between the two endpoints in historical data; echo signal similarity determines the presence of abnormal reflections or transmission problems by comparing the echo signals from the two endpoints; geographic coordinate offset considers the deviation between the actual geographical locations of the two endpoints and the preset locations. By integrating these factors, the system can more comprehensively assess the anomaly probability of each candidate edge. In this embodiment, firstly, the historical electrical parameter difference features, echo signal similarity, and geographic coordinate offset of the two endpoints of each candidate edge are extracted. These features reflect the changes in electrical parameters, abnormal signal transmission, and deviations in geographical location, respectively. Specifically, the historical electrical parameter difference features calculate the difference values ​​and trends of electrical parameters (such as voltage, current, power factor, etc.) between the two endpoints in historical data. The echo signal similarity is calculated by comparing the echo signals from the two endpoints; a lower similarity indicates a more abnormal signal transmission. The geographic coordinate offset calculates the deviation between the actual geographical locations of the two endpoints and the preset locations. These features are normalized to the [0, 1] interval for subsequent fusion calculations. The normalized features are then weighted and summed according to the weight value corresponding to each feature to obtain the anomaly probability.

[0124] Finally, the system identifies candidate edges with anomaly probabilities greater than or equal to a preset anomaly threshold as anomalous connections. This preset threshold, set based on historical data and practical operational experience, distinguishes between normal and anomalous connections. Through a graph neural network's gating propagation mechanism, the system further determines anomalous connection points based on these edges. This gating propagation mechanism dynamically updates node information, accurately locating anomalous connection points by analyzing the information of the two endpoints and adjacent nodes of the anomalous connection edge, thereby achieving dynamic optimization of the topology and anomaly detection.

[0125] This invention compares the dynamic topology graph with the initial preset topology edge-to-edge differences. The system can calculate the connection confidence and deviation of each edge. This difference analysis method can accurately identify edges in the dynamic topology graph that deviate significantly from the initial topology, thus determining candidate abnormal edges. This method not only considers changes in the topology but also quantifies the degree of change through confidence and deviation, providing a scientific basis for subsequent anomaly detection. After determining candidate abnormal edges, the system further considers the historical electrical parameter differential characteristics of the two endpoints, echo signal similarity, and geographic coordinate offset to calculate the anomaly probability of the candidate edges. This multi-dimensional analysis method can comprehensively evaluate the characteristics of abnormal edges, avoiding misjudgments that may arise from a single indicator. Historical electrical parameter differential characteristics reflect changes in the electrical characteristics between nodes, echo signal similarity can detect abnormal reflections in signal transmission, and geographic coordinate offset verifies the rationality of the connection from a physical location perspective. By integrating these factors, the system can more accurately determine the probability of abnormal edges, thereby improving the accuracy of anomaly detection. By leveraging the gated propagation mechanism of graph neural networks, the system can dynamically update node information based on abnormal connections, accurately locating anomalous connection points. The gated propagation mechanism, through dynamically updating node embedding vectors, adapts to changes in topology and optimizes the topology graph representation in real time. This method not only responds quickly to topology changes but also maintains high-efficiency anomaly detection capabilities in complex network environments, ensuring dynamic optimization of the topology.

[0126] By accurately identifying abnormal connection points, the system can quickly locate fault points or abnormal connections that may lead to line losses, thus providing power system operation and maintenance personnel with a clear direction for troubleshooting. This method significantly reduces the time and workload of manual troubleshooting, improving operation and maintenance efficiency. Simultaneously, through multi-dimensional anomaly detection and dynamic optimization, the system can more reliably identify and handle abnormal situations, enhancing the stability and reliability of the power system. For example, in complex topology environments such as old residential areas and urban-rural fringe areas, traditional methods struggle to effectively identify abnormal connection points. This invention, by combining graph neural networks and attention mechanisms, can dynamically adapt to complex network structure changes and accurately detect abnormal connections. This method not only improves the adaptability of anomaly detection but also provides strong technical support for power system operation and maintenance in complex environments. By setting preset anomaly thresholds, the system can effectively distinguish between normal and abnormal connections, avoiding false alarms caused by minor deviations. This method, combined with multi-dimensional feature analysis, can more robustly judge abnormal situations, thereby reducing the false alarm rate and improving the reliability of anomaly detection. Through comparison of the dynamic topology graph with the initial topology, multi-dimensional anomaly feature analysis, and dynamic optimization of the graph neural network, efficient optimization of the topology structure and anomaly detection are achieved. This method not only improves the accuracy of anomaly detection, but also enhances the efficiency and reliability of power system operation and maintenance, providing strong technical support for line loss cause analysis and fault location in complex environments.

[0127] Optionally, determining the downstream associated user meter of the abnormal connection point based on its location information includes:

[0128] Starting from the abnormal connection point, perform a breadth-first search along the outgoing edges of the dynamic topology graph until the third-level node is reached, thus obtaining a set of third-level nodes.

[0129] The user meters corresponding to all the third-level nodes in the third-level node set are regarded as the associated user meters.

[0130] Specifically, after identifying the abnormal connection point, a breadth-first search (BFS) algorithm is used to determine the associated user meters downstream of the abnormal connection point. First, using the abnormal connection point as the starting vertex, a BFS is performed along the outgoing edges in the dynamic topology graph. The BFS algorithm traverses the graph structure layer by layer, visiting all reachable nodes in ascending order. During the search, the system visits nodes connected to the abnormal connection point layer by layer until it reaches the tertiary nodes, i.e., the nodes where the user meters are located. These visited tertiary nodes are collected into a set, forming a tertiary node set. During the search, the information of each visited node and its corresponding user meter is recorded. When the search is complete, the tertiary node set contains all user meters directly or indirectly connected to the abnormal connection point. These user meters, being downstream of the abnormal connection point, may be affected by it and are therefore identified as associated user meters. In this way, the system can quickly and accurately identify user meters related to the abnormal connection point, providing a clear scope for further line loss analysis and fault location.

[0131] This invention employs a breadth-first search algorithm, starting from the anomalous connection point and proceeding along the outgoing edges of the dynamic topology graph. This allows for rapid traversal of all downstream nodes until reaching the third-level nodes (user meters). This method quickly identifies all user meters directly or indirectly connected to the anomalous connection point, avoiding the inefficient method of checking each meter individually and significantly improving efficiency. The breadth-first search algorithm visits nodes layer by layer in ascending order, ensuring that all visited third-level nodes are actual downstream user meters of the anomalous connection point. This layer-by-layer traversal not only accurately locates associated users but also avoids omissions or misjudgments, ensuring the accuracy and completeness of the investigation.

[0132] For example, in complex topological environments such as old residential areas and urban-rural fringe areas, dynamic topology maps may contain numerous branches and connections. Breadth-first search (BFS) algorithms can effectively handle this complexity by dynamically adapting to changes in the topology through layer-by-layer traversal, ensuring accurate identification of associated user meters even in complex network environments. The set of associated user meters identified through BFS provides a clear scope for subsequent line loss cause analysis and fault location. Maintenance personnel can focus their efforts on analyzing the data from these associated user meters, avoiding wasting time and resources on numerous irrelevant users, thus improving the focus and efficiency of subsequent analysis. Dynamic topology maps can reflect the electrical connections within a transformer area in real time, and BFS algorithms can quickly respond to topology changes in this dynamic environment. Even when the topology changes, the system can still accurately identify associated user meters through BFS, ensuring the accuracy and timeliness of troubleshooting. By accurately locating associated user meters, maintenance personnel can quickly narrow down the investigation scope, reducing unnecessary inspections. This method not only saves time and manpower costs but also reduces the impact of troubleshooting on users' normal electricity consumption, improving the overall efficiency and economy of power system operation and maintenance.

[0133] Optionally, determining the electricity consumption anomaly score for each associated user's electricity meter based on the load data of each associated user's electricity meter within a preset time period includes:

[0134] Based on the daily frozen electricity consumption, power curve, and transient response data of the associated user's electricity meter within the preset time period, a multi-dimensional load vector of the associated user's electricity meter is generated within the preset time period.

[0135] By using a temporal convolutional network, feature extraction is performed on the multidimensional load vector to obtain the load dynamic features of the multidimensional load vector;

[0136] Based on the load dynamic characteristics, the electricity consumption deviation of the associated user's electricity meter is determined;

[0137] The abnormal electricity consumption score is obtained based on the electricity consumption deviation and the preset line loss influence coefficient.

[0138] Specifically, when determining the electricity consumption anomaly score for each associated user's electricity meter, a multi-dimensional load vector for that meter is first generated based on the daily frozen electricity consumption, power curve, and transient response data of the associated user's electricity meter within a preset time period. This process integrates different types of data into a single vector to comprehensively reflect the meter's load characteristics during that time period. The daily frozen electricity consumption provides information on the meter's cumulative electricity consumption at a specific point in time, the power curve provides information on the meter's power changes at different points in time, and the transient response data provides information on the meter's dynamic response characteristics under the action of excitation signals. By fusing these data, the resulting multi-dimensional load vector can more comprehensively describe the meter's load status. Next, a temporal convolutional network is used to extract features from the multi-dimensional load vector. The temporal convolutional network is a neural network structure specifically designed for processing time-series data, effectively capturing the time dependence and dynamic change features in the multi-dimensional load vector. By performing convolution operations on the multi-dimensional load vector, the network can extract key features reflecting the dynamic changes in the meter's load; these features are called load dynamic features. These features can reveal the changing patterns and anomalies of the meter's load within the preset time period. Based on the extracted load dynamic characteristics, the system further determines the electricity consumption deviation of associated user meters. Electricity consumption deviation is an indicator that measures the difference between the actual load of a meter and its normal load pattern. By analyzing load dynamic characteristics, the system can identify the degree of abnormal fluctuations or deviations from the normal load pattern of the meters, thereby quantifying the electricity consumption deviation. This indicator reflects the degree of deviation between the meter's electricity consumption behavior and normal electricity consumption behavior within a preset time period, providing a basis for subsequent anomaly score calculation. Finally, combining the electricity consumption deviation and a preset line loss influence coefficient, the electricity consumption anomaly score of the associated user meters is calculated. The preset line loss influence coefficient is set based on historical data and experience and is used as a weighting factor to convert the electricity consumption deviation into anomaly score. By multiplying the electricity consumption deviation by the line loss influence coefficient, the system can obtain the electricity consumption anomaly score for each associated user meter. The higher this score, the more abnormal the meter's electricity consumption behavior, which may be a potential source of line loss problems. In this way, the system can quantitatively assess the degree of anomaly for each associated user meter, providing an important basis for subsequent fault location and line loss cause analysis.

[0139] In a preferred embodiment of the present invention, a Temporal Convolutional Network (TCN) is used as a neural network structure to process time-series data. TCN processes time-series data through convolutional layers, effectively capturing temporal dependencies and local patterns in the data. In this embodiment, the temporal convolutional network is used to extract features from the multidimensional load vectors of associated user meters to generate key features reflecting dynamic changes in meter load.

[0140] Temporal convolutional networks consist of multiple convolutional layers, each containing multiple convolutional kernels (filters). These kernels slide across the time series, extracting local features through convolutional operations. Convolutional operations can capture local patterns in the time series, such as load fluctuations, abrupt changes, or periodic variations. In this embodiment, the input to the temporal convolutional network is a multidimensional load vector associated with the user's electricity meter, which includes information such as daily frozen electricity consumption, power curves, and transient response data.

[0141] Specifically, a multi-dimensional load vector within a preset time period is input into a temporal convolutional network. These vectors contain load information from the electricity meter at different points in time. The input data is processed through multiple convolutional layers, each containing multiple convolutional kernels. These kernels slide across the time series to extract local features. The convolutional operation captures the temporal dependencies and local patterns in the time series. After processing by multiple convolutional layers, the temporal convolutional network extracts key features reflecting the dynamic changes in the electricity meter load. These features reveal the patterns and anomalies in the meter load within the preset time period. Finally, the temporal convolutional network outputs the extracted load dynamic features, which are used in subsequent electricity consumption deviation calculations.

[0142] The multi-dimensional load vectors associated with user electricity meters are input into a temporal convolutional network. These vectors contain information such as daily frozen electricity consumption, power curves, and transient response data. The temporal convolutional network can simultaneously process this multi-dimensional data, extracting features that comprehensively reflect the meter load status. Through convolution operations, the temporal convolutional network captures dynamic changes in the time series; for example, it can identify short-term fluctuations, long-term trends, and periodic patterns in the meter load. These dynamic features can reveal abnormal behavior of the meter load. The sliding operation of the convolution kernel on the time series allows the temporal convolutional network to identify local patterns; for example, it can detect abrupt changes or abnormal fluctuations in the load data, which may be potential sources of line loss problems. The load dynamic features extracted by the temporal convolutional network can be directly used for subsequent electricity consumption deviation calculations. These features provide a high-quality data foundation for subsequent anomaly detection and line loss cause analysis, ensuring the consistency and accuracy of the entire system's operation. By applying a temporal convolutional network, this embodiment can efficiently extract key features reflecting the dynamic changes of meter load from a multidimensional load vector. This not only improves the efficiency of feature extraction but also enhances the processing capability of time series data, enabling the system to more accurately identify and locate line loss problems.

[0143] In this embodiment of the invention, a multidimensional load vector of associated user meters is generated, and its load dynamic features are extracted using a temporal convolutional network. This allows for the determination of electricity consumption deviation and anomaly scores, achieving a precise quantitative assessment of abnormal user meter behavior. First, daily frozen electricity consumption, power curves, and transient response data are integrated into a multidimensional load vector, comprehensively reflecting the meter's load status within a preset time period, providing a rich data foundation for subsequent analysis. The temporal convolutional network efficiently extracts key features reflecting dynamic load changes from the multidimensional load vector, capturing local patterns and time dependencies in the time series, thereby accurately identifying abnormal fluctuations and potential line loss problems. The electricity consumption deviation calculated based on these dynamic features further quantifies the degree of deviation between the meter load and the normal pattern. The anomaly score, obtained by combining the preset line loss impact coefficient, provides maintenance personnel with a clear priority for investigation, helping to quickly locate high-risk user meters and significantly improving the efficiency and accuracy of line loss cause analysis.

[0144] Optionally, injecting a steep-front traveling wave signal into the associated user's meter with the highest electricity consumption anomaly score to obtain the reflection peak time difference, reflection coefficient, and attenuation gradient of the associated user's meter includes:

[0145] A steep-front traveling wave signal is injected at the position in front of the electricity meter of the associated user with the highest electricity consumption anomaly score;

[0146] Acquire the traveling wave reflection waveform of the associated user's electricity meter during the second preset acquisition time period; the second preset acquisition time period is set from the third preset time point before the injection of the steep leading edge traveling wave signal to the fourth preset time point after the injection of the steep leading edge traveling wave signal.

[0147] Feature extraction is performed on the traveling wave reflection waveform to obtain the reflection peak time difference, the reflection coefficient, and the attenuation gradient.

[0148] Specifically, a steep-front traveling wave signal is first injected before the meter of the associated user with the highest electricity anomaly score. This signal is a pulse signal with a fast rising edge, capable of propagating in the power line and being reflected when it encounters a fault point or impedance change point. To ensure accurate capture of the reflected signal, this embodiment sets a second preset acquisition time period, starting from a third preset time point before the injection of the steep-front traveling wave signal and ending at a fourth preset time point after the injection. For example, the third preset time point can be set to 100 milliseconds, and the fourth preset time point can be set to 200 milliseconds, ensuring complete acquisition of the reflected waveform within the time range before and after signal injection. During the second preset acquisition time period, the traveling wave reflected waveform of the associated user's meter is acquired in real time using high-precision sensors. These sensors can record the voltage and current changes of the reflected signal with high time resolution, thereby generating a complete traveling wave reflected waveform. The acquired traveling wave reflected waveform contains rich information, such as the amplitude, phase, and time delay of the reflected signal, which is crucial for subsequent feature extraction. Next, feature extraction is performed on the acquired traveling wave reflected waveform. This process primarily focuses on three key parameters: reflection peak time difference, reflection coefficient, and attenuation gradient. The reflection peak time difference refers to the time interval from the injection of the traveling wave signal at the steep leading edge to the receipt of the reflected signal; this parameter reflects the distance between the fault point and the injection point. The reflection coefficient is the amplitude ratio of the reflected signal to the incident signal, reflecting the degree of impedance change at the fault point. The attenuation gradient is the attenuation rate of the reflected signal during propagation; this parameter provides information about the line insulation condition and the severity of the fault.

[0149] By extracting these feature parameters, the system can perform quantitative analysis of the electrical status of associated user meters. These parameters not only help determine the specific location of the fault but also provide clues about the nature of the fault, such as whether it is insulation damage, poor contact, or equipment aging. This feature extraction method based on traveling wave signals is characterized by high accuracy and fast response, providing accurate data support for subsequent fault diagnosis and line loss cause analysis.

[0150] This invention achieves high-precision fault location and preliminary fault assessment by injecting a steep-front traveling wave signal at the meter of the associated user with the highest electricity anomaly score and acquiring the reflected waveform of the traveling wave within a preset time period. By extracting three key features—reflection peak time difference, reflection coefficient, and attenuation gradient—the system can quickly determine the distance between the fault point and the injection point, assess the impedance change at the fault point, and determine the line insulation status and fault severity. This method not only improves the accuracy of fault location but also provides direct electrical characteristic data for line loss cause analysis, significantly enhancing the efficiency and reliability of power system fault diagnosis and providing strong support for rapid repair and optimized operation.

[0151] Optionally, the step of iterating the reflection path using the Double DQN reinforcement learning algorithm based on the reflection peak time difference, the reflection coefficient, and the attenuation gradient to obtain the fault point of the target transformer area, the line loss cause type of the fault point, and the confidence level corresponding to the line loss cause type includes:

[0152] A multidimensional state vector is formed based on the reflection peak time difference, the reflection coefficient, and the attenuation gradient;

[0153] The multidimensional state vector is input into the Double DQN network. The fault location is taken as the action and the minimization of reflected energy loss is taken as the immediate reward. The Q-value network is iteratively updated to obtain the location coordinates of the fault point.

[0154] Based on the location coordinates, determine the dual-mode features corresponding to the location coordinates; the dual-mode features include the transient impedance-frequency fingerprint and the reflected energy attenuation gradient of the fault point;

[0155] The transient impedance-frequency fingerprint and the reflected energy attenuation gradient are input into a multimodal fusion deep learning model for prediction, thereby obtaining the line loss cause type of the fault point and the confidence level corresponding to the line loss cause type.

[0156] Specifically, a multidimensional state vector is first constructed based on the reflection peak time difference, reflection coefficient, and attenuation gradient. This multidimensional state vector centrally reflects the electrical characteristics of the fault point, providing fundamental data for subsequent fault location and cause analysis. This multidimensional state vector is then input into a Double DQN network. In the Double DQN network, the fault point location is used as the action, and the goal is to minimize reflection energy loss, with an immediate reward set. By iteratively updating the Q-value network, the system can gradually optimize action selection. When the Q-value network update changes very little over several consecutive iterations (e.g., the change is less than a preset convergence threshold), the algorithm is considered to have converged, and iteration stops. Finally, the location coordinates of the fault point are determined. By utilizing the optimization capabilities of reinforcement learning algorithms, the action strategy is continuously adjusted to minimize reflection energy loss, thereby accurately locating the fault point. After determining the fault point location coordinates, the corresponding dual-mode features are further determined. Specifically, after determining the fault point location coordinates, the corresponding dual-mode features are further determined, including transient impedance-frequency fingerprint and reflection energy attenuation gradient. Transient impedance-frequency fingerprinting is generated by acquiring transient voltage and current signals at the fault point, performing frequency domain analysis, and calculating impedance values ​​at different frequencies. The dual-mode features include the transient impedance-frequency fingerprint of the fault point and the reflected energy attenuation gradient. The transient impedance-frequency fingerprint reflects the impedance characteristics of the fault point at different frequencies, while the reflected energy attenuation gradient provides detailed information on energy loss at the fault point. These two features describe the electrical state of the fault point from different perspectives, providing rich information for subsequent prediction of line loss causes.

[0157] Finally, the transient impedance-frequency fingerprint and reflected energy attenuation gradient are input into a multimodal fusion deep learning model for prediction. This model can comprehensively process different types of data. By learning patterns and regularities from a large amount of historical data, the model can accurately predict the type of line loss at a fault point and provide the corresponding confidence level. The confidence level reflects the reliability of the prediction results, helping maintenance personnel better understand and handle fault problems.

[0158] This embodiment achieves precise fault location by constructing a multi-dimensional state vector from the reflection peak time difference, reflection coefficient, and attenuation gradient, and iteratively optimizing the reflection path using the Double DQN reinforcement learning algorithm. The reinforcement learning algorithm aims to minimize reflection energy loss and efficiently determines the fault location coordinates by iteratively updating the Q-value network. Furthermore, by extracting the dual-mode features of the fault point and inputting them into a multi-modal fusion deep learning model for prediction, the line loss cause type and its corresponding confidence level can be accurately identified. This method not only improves the accuracy of fault location but also enhances the accuracy and reliability of line loss cause analysis, providing strong support for rapid fault diagnosis and maintenance in power systems.

[0159] Combination Figure 2 As shown, the present invention provides an artificial intelligence-based system for analyzing and locating the causes of line loss, comprising:

[0160] The excitation sequence sending module is used to send an excitation sequence to the main meter, circuit breaker and user meter of the target transformer area when the bus loss rate of the target transformer area exceeds the preset line loss threshold, so as to obtain the transient response data corresponding to the main meter, the circuit breaker and the user meter respectively.

[0161] The topology construction module is used to construct a dynamic topology map of the target transformer area based on the transient response data corresponding to the master meter, the circuit breaker, and the user meter, respectively.

[0162] The anomaly detection module is used to dynamically optimize the topology and detect anomalies based on the dynamic topology graph, combined with graph neural networks and attention mechanisms, and to determine abnormal connection points.

[0163] The associated user location module is used to determine the associated user's electricity meter downstream of the abnormal connection point based on the location information of the abnormal connection point.

[0164] The score evaluation module is used to determine the electricity consumption anomaly score of each associated user's electricity meter based on the load data of each associated user's electricity meter within a preset time period;

[0165] The traveling wave injection module is used to inject a steep-front traveling wave signal into the associated user's electricity meter with the highest electricity consumption anomaly score, so as to obtain the reflection peak time difference, reflection coefficient and attenuation gradient of the associated user's electricity meter;

[0166] The analysis and localization module is used to obtain the fault point of the target transformer area, the line loss cause type of the fault point, and the confidence level corresponding to the line loss cause type by iterating the reflection path through the Double DQN reinforcement learning algorithm based on the reflection peak time difference, the reflection coefficient, and the attenuation gradient.

[0167] The AI-based line loss cause analysis and location system of the present invention has the same advantages over the prior art as the AI-based line loss cause analysis and location method described above, and will not be repeated here.

[0168] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the artificial intelligence-based method for analyzing and locating line loss causes as described above.

[0169] The advantages of the computer-readable storage medium of the present invention compared to the prior art are the same as those of the above-mentioned artificial intelligence-based line loss cause analysis and location method compared to the prior art, and will not be repeated here.

[0170] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A method for analyzing and locating the causes of line loss based on artificial intelligence, characterized in that, include: When the bus loss rate of the target transformer area exceeds the preset line loss threshold, an excitation sequence is sent to the main meter, circuit breaker and user meter of the target transformer area to obtain the transient response data corresponding to the main meter, the circuit breaker and the user meter respectively. Based on the transient response data corresponding to the master meter, the circuit breaker, and the user meter, a dynamic topology map of the target distribution area is constructed. Based on the dynamic topology graph, combined with graph neural networks and attention mechanisms, dynamic optimization of the topology structure and anomaly detection are performed to determine abnormal connection points; Based on the location information of the abnormal connection point, determine the downstream associated user's electricity meter of the abnormal connection point; Based on the load data of each associated user's electricity meter within a preset time period, determine the electricity consumption anomaly score of each associated user's electricity meter; A steep-front traveling wave signal is injected into the electricity meter of the associated user with the highest electricity consumption anomaly score to obtain the reflection peak time difference, reflection coefficient, and attenuation gradient of the associated user's electricity meter. Based on the reflection peak time difference, the reflection coefficient, and the attenuation gradient, the reflection path is iterated using the Double DQN reinforcement learning algorithm to obtain the fault point of the target transformer area, the line loss cause type of the fault point, and the confidence level corresponding to the line loss cause type.

2. The method for analyzing and locating the causes of line loss based on artificial intelligence according to claim 1, characterized in that, When the bus loss rate of the target distribution area exceeds a preset line loss threshold, an excitation sequence is sent to the main meter, circuit breaker, and user meters of the target distribution area to obtain transient response data corresponding to the main meter, the circuit breaker, and the user meters, respectively, including: The bus loss rate of the target transformer area is determined based on the actual line loss value of each branch line in the target transformer area. When the bus loss rate exceeds the preset line loss threshold, the excitation sequence is generated and synchronously sent to the main meter of the target transformer area, the circuit breaker of each branch line, and the user meter through a parallel communication protocol. Simultaneously, the transient voltage and transient current of the main meter, the circuit breaker, and the user meter are acquired during a first preset acquisition time period; the first preset acquisition time period is set from a first preset time point before the excitation sequence is issued to a second preset time point after the excitation sequence is issued. The transient response data is obtained based on the transient voltage and transient current of the master meter, the circuit breaker, and the user meter within the preset acquisition time period.

3. The method for analyzing and locating the causes of line loss based on artificial intelligence according to claim 1, characterized in that, The step of constructing a dynamic topology map of the target distribution area based on the transient response data corresponding to the master meter, the circuit breaker, and the user meter includes: Feature extraction is performed on the transient response data to obtain the key feature vector of the transient response data; Construct a topology node system for the target transformer area, wherein the main meter is used as a first-level node, the circuit breaker of each branch line is used as a second-level node, and the user meter is used as a third-level node. Based on the key feature vectors corresponding to the master meter, the circuit breaker, and the user meter, the electrical correlation strength between each node in the topology node system is determined. The electrical association strength is input into a preset lightweight pulse-graph attention network. The key feature vector is used as the initial node embedding vector, and the electrical association strength is used as the initial edge weight. The node embedding vector is iteratively updated through a multi-head attention mechanism to obtain a dynamic adjacency matrix. The dynamic adjacency matrix is ​​subjected to threshold judgment and sparsification processing to generate the dynamic topology graph containing the dynamic connection relationships between the first-level nodes, the second-level nodes and the third-level nodes.

4. The method for analyzing and locating the causes of line loss based on artificial intelligence according to claim 3, characterized in that, The step of dynamically optimizing the topology and detecting anomalies based on the dynamic topology graph, combined with graph neural networks and attention mechanisms, and determining abnormal connection points includes: The dynamic topology map is compared with the initial preset topology of the target station area edge-to-edge to obtain the connection confidence and deviation of each edge in the dynamic topology map; Based on the connection confidence and the deviation, abnormal candidate edges in the dynamic topology graph are determined and a set of abnormal candidate edges is formed. For each of the abnormal candidate edges in the set of abnormal candidate edges, the abnormal probability of the abnormal candidate edge is determined based on the historical electrical parameter difference characteristics of the two end nodes of the abnormal candidate edge, the echo signal similarity, and the geographical coordinate offset. The candidate edges of anomalies whose anomalies are greater than or equal to a preset anomaly threshold are designated as anomaly connection edges. The abnormal connection point is determined based on the abnormal connection edge using the gated propagation mechanism of the graph neural network.

5. The method for analyzing and locating the causes of line loss based on artificial intelligence according to claim 3, characterized in that, The step of determining the downstream associated user meters of the abnormal connection point based on the location information of the abnormal connection point includes: Starting from the abnormal connection point, perform a breadth-first search along the outgoing edges of the dynamic topology graph until the third-level node is reached, thus obtaining a set of third-level nodes. The user meters corresponding to all the third-level nodes in the third-level node set are regarded as the associated user meters.

6. The method for analyzing and locating the causes of line loss based on artificial intelligence according to claim 3, characterized in that, The step of determining the electricity consumption anomaly score for each associated user's electricity meter based on the load data of each associated user's electricity meter within a preset time period includes: Based on the daily frozen electricity consumption, power curve, and transient response data of the associated user's electricity meter within the preset time period, a multi-dimensional load vector of the associated user's electricity meter is generated within the preset time period. By using a temporal convolutional network, feature extraction is performed on the multidimensional load vector to obtain the load dynamic features of the multidimensional load vector; Based on the load dynamic characteristics, the electricity consumption deviation of the associated user's electricity meter is determined; The abnormal electricity consumption score is obtained based on the electricity consumption deviation and the preset line loss influence coefficient.

7. The method for analyzing and locating the causes of line loss based on artificial intelligence according to claim 1, characterized in that, The step of injecting a steep-front traveling wave signal into the electricity meter of the associated user with the highest electricity consumption anomaly score to obtain the reflection peak time difference, reflection coefficient, and attenuation gradient of the associated user's electricity meter includes: A steep-front traveling wave signal is injected at the position in front of the electricity meter of the associated user with the highest electricity consumption anomaly score; Acquire the traveling wave reflection waveform of the associated user's electricity meter during the second preset acquisition time period; the second preset acquisition time period is set from the third preset time point before the injection of the steep leading edge traveling wave signal to the fourth preset time point after the injection of the steep leading edge traveling wave signal. Feature extraction is performed on the traveling wave reflection waveform to obtain the reflection peak time difference, the reflection coefficient, and the attenuation gradient.

8. The method for analyzing and locating the causes of line loss based on artificial intelligence according to claim 7, characterized in that, The step of iterating the reflection path using the Double DQN reinforcement learning algorithm based on the reflection peak time difference, the reflection coefficient, and the attenuation gradient to obtain the fault point of the target transformer area, the line loss cause type of the fault point, and the confidence level corresponding to the line loss cause type includes: A multidimensional state vector is formed based on the reflection peak time difference, the reflection coefficient, and the attenuation gradient; The multidimensional state vector is input into the Double DQN network. The fault location is taken as the action and the minimization of reflected energy loss is taken as the immediate reward. The Q-value network is iteratively updated to obtain the location coordinates of the fault point. Based on the location coordinates, determine the dual-mode features corresponding to the location coordinates; the dual-mode features include the transient impedance-frequency fingerprint and the reflected energy attenuation gradient of the fault point; The transient impedance-frequency fingerprint and the reflected energy attenuation gradient are input into a multimodal fusion deep learning model for prediction, thereby obtaining the line loss cause type of the fault point and the confidence level corresponding to the line loss cause type.

9. A system for analyzing and locating the causes of line loss based on artificial intelligence, characterized in that, include: The excitation sequence sending module is used to send an excitation sequence to the main meter, circuit breaker and user meter of the target transformer area when the bus loss rate of the target transformer area exceeds the preset line loss threshold, so as to obtain the transient response data corresponding to the main meter, the circuit breaker and the user meter respectively. The topology construction module is used to construct a dynamic topology map of the target transformer area based on the transient response data corresponding to the master meter, the circuit breaker, and the user meter, respectively. The anomaly detection module is used to dynamically optimize the topology and detect anomalies based on the dynamic topology graph, combined with graph neural networks and attention mechanisms, and to determine abnormal connection points. The associated user location module is used to determine the associated user's electricity meter downstream of the abnormal connection point based on the location information of the abnormal connection point. The score evaluation module is used to determine the electricity consumption anomaly score of each associated user's electricity meter based on the load data of each associated user's electricity meter within a preset time period; The traveling wave injection module is used to inject a steep-front traveling wave signal into the associated user's electricity meter with the highest electricity consumption anomaly score, so as to obtain the reflection peak time difference, reflection coefficient and attenuation gradient of the associated user's electricity meter; The analysis and localization module is used to obtain the fault point of the target transformer area, the line loss cause type of the fault point, and the confidence level corresponding to the line loss cause type by iterating the reflection path through the Double DQN reinforcement learning algorithm based on the reflection peak time difference, the reflection coefficient, and the attenuation gradient.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the artificial intelligence-based method for analyzing and locating line loss causes as described in any one of claims 1-8.

Images