Power distribution network multi-source data fusion fault intelligent research and judgment system and method based on artificial intelligence

Abstract

This invention relates to the field of distribution network fault assessment technology, and is an intelligent fault assessment system and method based on artificial intelligence and multi-source data fusion for distribution networks. The system includes a data fusion and preprocessing unit, an AI intelligent analysis layer unit, and an application unit. This invention innovatively adopts a dynamic heterogeneous graph structure as the core data model. This model can not only integrate heterogeneous data such as SCADA, fault indicators, smart meters, protection device action signals, and even 95598 customer service work orders, but more importantly, through graph neural networks and long short-term memory network technology, it simultaneously captures the spatial propagation characteristics and temporal sequential evolution patterns of faults. This enables the system to understand complex fault chain reactions, effectively distinguish between real faults and instantaneous anomalies, thereby fundamentally reducing the common problems of misjudgment, missed judgment, and overjudgment caused by incomplete information in existing technologies, and achieving accurate location of fault sections.

Description

Technical Field

[0001] This invention relates to the field of distribution network fault assessment technology, and in particular to an intelligent fault assessment system and method for distribution networks based on artificial intelligence and multi-source data fusion. Background Technology

[0002] Traditional fault diagnosis techniques for distribution networks primarily rely on human experience, single data sources, and simple rule-based judgments, such as SCADA systems, fault indicators, or protective device action signals. Due to limited data sources and a lack of multi-dimensional data fusion analysis, fault location is often time-consuming and has low accuracy. Early fault diagnosis methods were typically based on threshold comparisons or fixed logic rules, such as overcurrent protection and zero-sequence current detection, which are ill-suited to the dynamic changes in complex distribution networks. Especially with a high proportion of distributed energy resources integrated, fault characteristics may be masked or interfered with, leading to misdiagnosis or missed diagnosis. Furthermore, traditional methods do not adequately consider the topological relationships of the distribution network, making it difficult to accurately assess the impact range of faults and resulting in low repair efficiency. With the development of smart grids, the scale of distribution networks is continuously expanding, and their structures are becoming increasingly complex. Traditional fault diagnosis methods based on single signals or simple rules can no longer meet the demands for high precision and rapid response. There is an urgent need for more intelligent and comprehensive fault diagnosis techniques to improve the efficiency and accuracy of fault handling.

[0003] Current advanced fault assessment schemes for distribution networks typically employ multi-source data fusion and machine learning techniques to improve the accuracy of fault detection and location. For example, they combine multi-dimensional information such as SCADA, AMI (Advanced Measuring Instrumentation), fault waveforms, and meteorological data, utilizing machine learning algorithms like random forests and support vector machines (SVM) for fault classification and probability prediction. Simultaneously, some studies introduce graph theory methods to analyze the distribution network topology and combine search algorithms such as Dijkstra's algorithm and A* to optimize fault segment location. In recent years, deep learning models (such as LSTM and CNN) have been applied to transient signal analysis to identify more complex fault modes, such as arc faults and intermittent grounding faults. Furthermore, some systems attempt to combine knowledge graph technology to construct a knowledge base linking distribution network equipment, operating status, and fault history to improve reasoning capabilities. However, existing schemes still face challenges such as strong dependence on data quality, insufficient model generalization ability, and limited real-time performance, especially in complex scenarios such as extreme weather or network reconfiguration, where the robustness of fault assessment still needs improvement. Future trends may move towards a hybrid intelligent assessment approach combining "data-driven" and "knowledge-based reasoning."

[0004] Current distribution network fault assessment technologies suffer from the following key shortcomings: traditional methods rely on static topology analysis, making it difficult to respond promptly to dynamic scenarios such as switch changes and distributed power source switching, resulting in low fault location accuracy; single data source criteria (such as relying solely on SCADA signals or fault indicators) are susceptible to interference, leading to a high false alarm rate; cross-system data (such as SCADA, electricity consumption information, and waveform data) lack effective integration, making collaborative assessment difficult; existing algorithms have insufficient ability to identify new fault modes (such as high-resistance grounding and intermittent faults); furthermore, the system's real-time performance, interpretability, and adaptability to complex power grid environments need improvement. These limitations severely restrict the efficiency and reliability of distribution network fault handling. Summary of the Invention

[0005] This invention provides an intelligent fault assessment system and method for distribution networks based on artificial intelligence and multi-source data fusion, which overcomes the shortcomings of the existing technologies and can effectively solve existing problems such as misjudgment, missed judgment, and multiple judgment.

[0006] To address the aforementioned problems, one of the technical solutions of this invention is implemented through the following means: an intelligent fault assessment system for distribution networks based on artificial intelligence multi-source data fusion, comprising:

[0007] The data fusion and preprocessing unit performs data fusion and preprocessing on multi-source data of the distribution network. The multi-source data includes integrated SCADA real-time data, smart meter information, fault indicator monitoring data, protection device action signals and 95598 customer service work orders.

[0008] The AI ​​intelligent analysis layer unit models the pre-processed multi-source data of the distribution network into a dynamic heterogeneous graph structure, performs time-series fault mode identification, and outputs fault assessment results.

[0009] The application unit automatically outputs an assessment report based on the fault assessment results and visualizes the fault by overlaying a single-line graph.

[0010] The aforementioned data fusion and preprocessing unit also includes a frequency synchronization module and a data detection module;

[0011] The frequency synchronization module uses a time series alignment algorithm to match different sampling rate data from multiple sources and synchronizes the sampling frequencies of the multiple sources to ensure that the data frequencies and times of the multiple sources are consistent.

[0012] The data detection module uses the isolated forest algorithm to detect abnormal data in multi-source data with synchronized sampling frequency. For abnormal data, it first classifies the abnormal data according to remote signaling data, telemetry data and topological relationship data, and then performs a comprehensive analysis of remote signaling data, telemetry data and topological relationship data.

[0013] The aforementioned AI intelligent analysis layer unit includes a dynamic graph construction module;

[0014] The dynamic graph construction module, based on graph neural networks and topology inference engine, models the preprocessed multi-source data of the distribution network into a dynamic heterogeneous graph structure. Nodes and edges are embedded with feature vectors, and graph attention network is used to calculate the fault propagation probability between nodes and locate fault sections.

[0015] The expression for the dynamic heterogeneity graph is as follows:

[0016] G t =(V t E t ,X t V ,X t E ),

[0017] In the formula, G t V represents the overall topology and operating status of the distribution network at time t; t E represents the set of all nodes at time t; t X represents the set of all edges at time t; t V X represents the set of features or attributes of all nodes at time t; t E It represents the set of features or attributes of all edges at time t;

[0018] Edge e embedded electrical distance Z ij Topology dynamic updates are triggered by adjacency matrix reconstruction through switch state changes, as shown in the following expression:

[0019] ,

[0020] In the formula, This represents the number of nodes in the adjacency matrix at time step t. i With nodes j The connection weights between them; e represents the exponential decay function; Represents electrical distance Z ij The modulus; and Representing nodes respectively i and nodes j The on / off state;

[0021] ,

[0022] In the formula, express Time-based fault mode Below, distribution network nodes i For neighboring nodesj The fault-related attention coefficient; This indicates an exponential enhancement operator that amplifies the differential correlation between fault features through nonlinear mapping, thereby enhancing the distinguishability of abnormal patterns. This represents an asymmetric activation operator with leakage, which preserves fine-grained gradients of fault characteristics and prevents fault information from being annihilated during propagation. Indicates the fault mode The attention parameter vector is transposed to customize feature association weights for different fault types; This represents an element-wise modulation operator that achieves element-wise weighting of attention parameters and feature interaction vectors, accurately capturing the local correlations of fault features; express The spatiotemporal linear transformation matrix at time step, integrating the static topological characteristics of the distribution network nodes with... The dynamic operational characteristics at any given moment are used to complete the spatiotemporal mapping of fault characteristics; , and They represent Time Node i ,node j and nodes k The fault state feature vector; and They represent Time-based fault mode Next node ij and nodes ik The power imbalance eigenvector; express Time-based fault mode Next node i A dynamic set of neighbors; Represents nodes in fault scenarios i The normalized summation operator on the dynamic neighbor set constrains the distribution of attention coefficients and ensures the rationality of the weights for fault information propagation.

[0023] Overlaying spatiotemporal feature extraction layers:

[0024] ,

[0025] In the formula, Indicates at time step t The node feature matrix obtained after the spatiotemporal feature extraction layer; This represents a convolutional layer in a graph attention network, used to extract spatial dependency features between nodes from a graph structure. This represents a temporal convolutional layer, used to extract temporal evolutionary features from a sequence of node features; Indicates at time step t The original node feature matrix of the input model; Represents the edge index tensor;

[0026] The node failure probability is calculated as follows:

[0027] ,

[0028] In the formula, Represents a node i The probability of failure occurrence, subscript f It represents "fault"; This represents the Sigmoid activation function; This represents the learnable weight matrix of the fault classification layer; Represents a node i In the L The feature vector of the layer;

[0029] The method for calculating the location of the faulty section is as follows:

[0030] ,

[0031] In the formula, This indicates the finally located faulty section; This represents an optimization operation that iterates through all edges belonging to the edge set. edge Find the value of the objective function. The largest edge; Represents an edge, connecting nodes i and nodes j ; Represents the set of all edges; Representing an edge Two endpoint nodes i and j The absolute value of the difference between the failure probabilities; Represents a node j The probability of failure occurrence, subscript f It represents "fault"; Represents an edge The relevant indicator function or weighting coefficient.

[0032] The aforementioned AI intelligent analysis layer unit also includes a fault mode recognition module and a data decision module:

[0033] The fault mode identification module uses a long short-term memory network to analyze multi-source data, which integrates real-time SCADA data, smart meter information, fault indicator monitoring data, protection device action signals, and 95598 customer service work orders. It identifies short-circuit / ground fault characteristics, combines historical fault information of the distribution network, judges the fault situation of the distribution network, and outputs the corresponding 95598 customer service work order handling results according to the fault situation.

[0034] The data decision module employs deep reinforcement learning. By performing deep learning on a large amount of historical fault information of the power distribution network and the corresponding handling results of 95598 customer service work orders, it provides preliminary analysis results of the fault causes.

[0035] The aforementioned application units also include a data cleaning rules module and a feature processing rules module;

[0036] The data cleaning rules module is used to perform missing data repair, anomaly handling, and consistency verification.

[0037] The feature processing rule module is used to numerically encode and standardize the cleaned data.

[0038] The aforementioned AI intelligent analysis layer unit also includes an optimization module. The optimization module adopts a centralized learning framework to achieve centralized aggregation and collaborative analysis of cross-system data. It trains the DRL model through a historical fault case library to support the continuous optimization of fault judgment strategies.

[0039] The second technical solution of this invention is achieved through the following method: an intelligent fault assessment method for distribution networks based on artificial intelligence multi-source data fusion, which is implemented using an intelligent fault assessment system for distribution networks based on artificial intelligence multi-source data fusion, and includes the following steps:

[0040] Data fusion and preprocessing are performed on multi-source data of the distribution network. The multi-source data includes integrated SCADA real-time data, smart meter information, fault indicator monitoring data, protection device action signals and 95598 customer service work orders.

[0041] The preprocessed multi-source data of the distribution network is modeled into a dynamic heterogeneous graph structure, time-series fault mode identification is performed, and fault assessment results are output.

[0042] The system automatically generates an assessment report based on the fault assessment results and visualizes the fault by overlaying a single-line graph.

[0043] The above-mentioned data fusion and preprocessing of multi-source data from the distribution network includes:

[0044] Based on the time series alignment algorithm, data with different sampling rates from multiple sources are matched, and the sampling frequency of multiple sources is synchronized to make the data frequency and time of multiple sources consistent.

[0045] The isolated forest algorithm is used to detect abnormal data in multi-source data with synchronized sampling frequency. For abnormal data, anomalies are first classified according to remote signaling data, telemetry data and topological relationship data, and then the remote signaling data, telemetry data and topological relationship data are comprehensively analyzed.

[0046] The above-mentioned method models the preprocessed multi-source data of the distribution network into a dynamic heterogeneous graph structure, performs time-series fault mode identification, and outputs fault assessment results, including:

[0047] Based on graph neural networks and topology inference engines, the preprocessed multi-source data of the distribution network is modeled as a dynamic heterogeneous graph structure, with feature vectors embedded in nodes and edges. Graph attention networks are used to calculate the probability of fault propagation between nodes and to locate fault sections.

[0048] The expression for the dynamic heterogeneity graph is as follows:

[0049] G t =(V t E t ,X t V ,X t E ),

[0050] In the formula, G t V represents the overall topology and operating status of the distribution network at time t; t E represents the set of all nodes at time t; t X represents the set of all edges at time t; t V X represents the set of features or attributes of all nodes at time t; t E It represents the set of features or attributes of all edges at time t;

[0051] Edge e embedded electrical distance Z ij Topology dynamic updates are triggered by adjacency matrix reconstruction through switch state changes, as shown in the following expression:

[0052] ,

[0053] In the formula, This represents the number of nodes in the adjacency matrix at time step t. i With nodes j The connection weights between them; e represents the exponential decay function; Represents electrical distance Z ij The modulus; and Representing nodes respectively i and nodes j The on / off state;

[0054] ,

[0055] In the formula, express Time-based fault mode Below, distribution network nodesi For neighboring nodes j The fault-related attention coefficient; This indicates an exponential enhancement operator that amplifies the differential correlation between fault features through nonlinear mapping, thereby enhancing the distinguishability of abnormal patterns. This represents an asymmetric activation operator with leakage, which preserves fine-grained gradients of fault characteristics and prevents fault information from being annihilated during propagation. Indicates the fault mode The attention parameter vector is transposed to customize feature association weights for different fault types; This represents an element-wise modulation operator that achieves element-wise weighting of attention parameters and feature interaction vectors, accurately capturing the local correlations of fault features; express The spatiotemporal linear transformation matrix at time step, integrating the static topological characteristics of the distribution network nodes with... The dynamic operational characteristics at any given moment are used to complete the spatiotemporal mapping of fault characteristics; , and They represent Time Node i ,node j and nodes k The fault state feature vector; and They represent Time-based fault mode Next node ij and nodes ik The power imbalance eigenvector; express Time-based fault mode Next node i A dynamic set of neighbors; Represents nodes in fault scenarios i The normalized summation operator on the dynamic neighbor set constrains the distribution of attention coefficients and ensures the rationality of the weights for fault information propagation.

[0056] Overlaying spatiotemporal feature extraction layers:

[0057] ,

[0058] In the formula, Indicates at time step t The node feature matrix obtained after the spatiotemporal feature extraction layer; This represents a convolutional layer in a graph attention network, used to extract spatial dependency features between nodes from a graph structure. This represents a temporal convolutional layer, used to extract temporal evolutionary features from a sequence of node features; Indicates at time stept The original node feature matrix of the input model; Represents the edge index tensor;

[0059] The node failure probability is calculated as follows:

[0060] ,

[0061] In the formula, Represents a node i The probability of failure occurrence, subscript f Represents "fault"; it is the output of the model, a scalar value between 0 and 1, used to assess the likelihood that the node is a source of fault or affected by a fault. This represents the Sigmoid activation function, which maps any real-valued input to the interval (0, 1), thus outputting a value that can be interpreted as a probability. This represents the learnable weight matrix of the fault classification layer; Represents a node i In the L The feature vector of the layer;

[0062] The method for calculating the location of the faulty section is as follows:

[0063] ,

[0064] In the formula, This indicates the finally located faulty section; This represents an optimization operation that iterates through all edges belonging to the edge set. edge Find the value of the objective function. The largest edge; Represents an edge, connecting nodes i and nodes j ; Represents the set of all edges; Representing an edge Two endpoint nodes i and j The absolute value of the difference between the failure probabilities; Represents a node j The probability of failure occurrence, subscript f It represents "fault"; Represents an edge The relevant indicator function or weighting coefficient.

[0065] The above-mentioned process of modeling the preprocessed multi-source data of the distribution network into a dynamic heterogeneous graph structure, performing time-series fault mode identification, and outputting fault assessment results also includes:

[0066] Long Short-Term Memory Network (LSTM) is used to analyze multi-source data, which integrates real-time SCADA data, smart meter data, fault indicator monitoring data, protection device action signals, and 95598 customer service work orders. This identifies short-circuit / ground fault characteristics, combines historical fault information of the distribution network to assess the fault situation, and outputs the corresponding 95598 customer service work order handling results based on the fault situation.

[0067] By employing deep reinforcement learning, and through deep learning on a large amount of historical fault information of the power distribution network and the corresponding handling results of 95598 customer service work orders, preliminary results of fault cause analysis are given.

[0068] Compared with the prior art, the present invention has the following advantages:

[0069] 1. Unlike traditional methods that rely solely on data from a single system such as SCADA, this invention innovatively employs a dynamic heterogeneous graph structure (G). t This model serves as the core data model. It not only integrates heterogeneous data from SCADA, fault indicators, smart meters, protection device action signals, and even 95598 customer service work orders, but more importantly, it uses Graph Neural Networks (GNN) and Long Short-Term Memory (LSTM) technologies to simultaneously capture the spatial (along the power grid topology) propagation characteristics and temporal sequential evolution patterns of faults. This deep fusion analysis across both spatial and temporal dimensions enables the system to understand complex fault chain reactions, effectively distinguish between real faults and transient anomalies, thereby fundamentally reducing the common problems of misjudgment, missed judgment, and overjudgment caused by incomplete information in existing technologies, and achieving precise fault location.

[0070] 2. This invention constructs a closed-loop intelligent system integrating "rapid perception, intelligent analysis, self-optimization, and intuitive presentation." On the one hand, by learning from historical handling cases through a deep reinforcement learning (DRL) model, the system can continuously optimize its judgment strategy, forming an "expert brain" with self-learning capabilities, thus improving the accuracy of judgment. On the other hand, the judgment results not only automatically generate reports but also visualize them, accurately overlaying abstract fault information onto a single-line graph, greatly improving dispatchers' intuitive understanding of the fault situation and on-site handling efficiency. This comprehensive intelligence from backend algorithms to frontend applications significantly enhances the system's practical value and adaptability. Attached Figure Description

[0071] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0072] Figure 1 This is a system structure block diagram of Embodiment 1 of the present invention.

[0073] Figure 2This is a schematic diagram of data access and cleaning in Embodiment 1 of the present invention.

[0074] Figure 3 This is a diagram showing the relationship between dynamic heterogeneous graphs in Embodiment 1 of the present invention.

[0075] Figure 4 This is a flowchart of the method in Embodiment 2 of the present invention. Detailed Implementation

[0076] The present invention is not limited to the following embodiments, and the specific implementation can be determined according to the technical solution of the present invention and the actual situation.

[0077] Example 1: As Figure 1 As shown in the figure, this invention discloses an intelligent fault assessment system for distribution networks based on artificial intelligence multi-source data fusion, comprising:

[0078] The data fusion and preprocessing unit performs data fusion and preprocessing on multi-source data of the distribution network. The multi-source data includes integrated SCADA real-time data, smart meter information, fault indicator monitoring data, protection device action signals and 95598 customer service work orders.

[0079] The AI ​​intelligent analysis layer unit models the pre-processed multi-source data of the distribution network into a dynamic heterogeneous graph structure, performs time-series fault mode identification, and outputs fault assessment results.

[0080] The application unit automatically outputs an assessment report based on the fault assessment results and visualizes the fault by overlaying a single-line graph.

[0081] The analysis report includes information such as SCADA real-time data, smart meter (AMI) collected information, fault indicator monitoring data, protection device action signals, and data cleaning and feature processing rules for 95598 customer service work orders (TTU and FTU).

[0082] The integrated SCADA real-time data includes switch fault trip signals, switch remote signaling changes, switch protection action signals, switch measurements, and distribution transformer power outage and restoration signals.

[0083] like Figure 2 The data access and cleaning diagram shows that the data fusion and preprocessing unit includes a frequency synchronization module and a data detection module.

[0084] The frequency synchronization module, based on the time series alignment algorithm (DTW), matches multi-source data with different sampling rates and performs synchronization processing on the sampling frequencies of multi-source data to ensure that the data frequencies and times of multi-source data are consistent.

[0085] The data detection module uses the Isolation Forest algorithm to detect abnormal data in multi-source data with synchronized sampling frequency. For abnormal data, it first classifies the abnormal data according to remote signaling data, telemetry data and topological relationship data, and then performs comprehensive analysis on the remote signaling data, telemetry data and topological relationship data.

[0086] Abnormal data includes distorted voltage, distorted current, abnormal fault indicators, and abnormal tripping signals, among which abnormal fault indicators can be floating-point abnormal signals. Switch opening and closing signals and power outage / restoration signals from the distribution transformer acquisition terminal are all remote signaling signals. Current, voltage, active power, and reactive power of switches and distribution transformers are the signals to be measured. The ledger, from the power supply side to the load side, progressively shows the connection relationships and ownership relationships of the lines, forming the ledger topology data.

[0087] Among them, different sampling rates can be used for SCADA at the second level.

[0088] The aforementioned AI intelligent analysis layer unit includes a dynamic graph construction module; the dynamic graph construction module, based on graph neural networks and topology inference engine, models the preprocessed multi-source data of the distribution network into a dynamic heterogeneous graph structure, embeds feature vectors into nodes (switches, transformers, customers) and edges (electrical connections), and uses graph attention network to calculate the probability of fault propagation between nodes to locate fault sections;

[0089] The relationship diagram of the dynamic heterogeneity graph is as follows: Figure 3 As shown, the expression for the dynamic heterogeneous graph is as follows:

[0090] G t =(V t E t ,X t V ,X t E ),

[0091] In the formula, G t V represents the overall topology and operating status of the distribution network at time t; t E represents the set of all nodes in the graph at time t, signifying the collection of all entities in the distribution network that need to be monitored and analyzed at time t (such as large feeders, pole-mounted circuit breakers, pole-mounted transformers, etc.); t This represents the set of all edges in the graph at time t, defining the relationships between various entities in the distribution network at time t (such as the connections between stations, lines, transformers, and customers); X t V X represents the set of features or attributes of all nodes at time t, describing the real-time operating state and static attributes (such as transformer voltage, current, active power, reactive power, etc.) of each entity object at time t; tE It represents the set of features or attributes of all edges at time t, describing the state and attributes of each association itself at time t (such as the opening and closing state of a switch).

[0092] Edge e embedded electrical distance Z ij Topology dynamic updates are triggered by adjacency matrix reconstruction through switch state changes, as shown in the following expression:

[0093] ,

[0094] In the formula, This represents the number of nodes in the adjacency matrix at time step t. i With nodes j The connection weights between nodes; e represents the exponential decay function (exp), which uses the natural exponent to power the nodes. i and j The feature differences between them are transformed into a decay weight between 0 and 1; the larger the feature difference, the lower the weight. The smaller the value, the more likely the corresponding node is to be affected. i and j The lower the association weight. Represents electrical distance Z ij The modulus; and Representing nodes respectively i and nodes j The switch state (usually represented by 0 and 1); =1 indicates a node i and j Both belong to the same fault association cluster (i.e., both have a state label of 1, satisfying the topological or electrical association conditions for fault propagation). The logical AND operator (representing "AND") will only calculate the exponentially decaying association weights if this condition is met. ,otherwise( or (If the fault association conditions are not met), the association weight between the two. Set it directly to 0.

[0095] Design fault feature attention coefficient:

[0096] ,

[0097] In the formula, express Time-based fault mode Below, distribution network nodes i For neighboring nodes j The fault-related attention coefficient; This indicates an exponential enhancement operator that amplifies the differential correlation between fault features through nonlinear mapping, thereby enhancing the distinguishability of abnormal patterns. This represents an asymmetric activation operator with leakage (LeakyReLU instantiation), which preserves fine-grained gradients of fault characteristics (such as weak power deviations) to prevent fault information from being annihilated during propagation; Indicates the fault mode The attention parameter vector is transposed to customize feature association weights for different fault types (such as short circuit and grounding). This represents an element-wise modulation operator that achieves element-wise weighting of attention parameters and feature interaction vectors, accurately capturing the local correlations of fault features; express The spatiotemporal linear transformation matrix at time T integrates the static topological characteristics of the distribution network node with the dynamic operating characteristics (such as voltage and current) at time T to complete the spatiotemporal dimension mapping of fault characteristics. , and They represent Time Node i ,node j and nodes k The fault status feature vector includes multi-source heterogeneous features such as equipment ledger attributes, real-time measurement data, and historical fault tags. and They represent Time-based fault mode Next node ij and nodes ik The power imbalance feature vector covers multiple dimensions of fault propagation indicators, such as active / reactive power deviation and phase difference. express Time-based fault mode Next node i The dynamic neighbor set (faults may trigger temporary topology changes, such as switch actions); Represents nodes in fault scenarios i The normalized summation operator on the dynamic neighbor set constrains the distribution of attention coefficients and ensures the rationality of the weights for fault information propagation.

[0098] Overlaying spatiotemporal feature extraction layers:

[0099] ,

[0100] In the formula, Indicates at time step t The node feature matrix obtained after passing through the spatiotemporal feature extraction layer; This represents a convolutional layer in a graph attention network, used to extract spatial dependency features between nodes from a graph structure. This represents a temporal convolutional layer, used to extract temporal evolutionary features from a sequence of node features; Indicates at time step t The input model's original node feature matrix typically contains the features of each node within a historical time window, with dimensions of [number of nodes, number of time steps, feature dimension]. The edge index tensor represents the graph, which defines the connection relationships (topology) between all nodes in the graph and is one of the inputs to the graph convolution operation;

[0101] The node failure probability is calculated as follows:

[0102] ,

[0103] In the formula, Represents a node i The probability of failure occurrence, subscript f Representing "fault", it is the output of the model, a scalar value between 0 and 1, used to assess the likelihood that the node is a source of fault or affected by a fault; This represents the Sigmoid activation function, which maps any real-valued input to the interval (0, 1), thus outputting a value that can be interpreted as a probability. This represents the learnable weight matrix of the fault classification layer; Represents a node i In the L The feature vector of the layer;

[0104] The method for calculating the location of the faulty section is as follows:

[0105] ,

[0106] In the formula, This indicates the finally located faulty section; This represents an optimization operation that iterates through all edges belonging to the edge set. edge Find the value of the objective function. The largest edge; Represents an edge, connecting nodes i and nodes j ; Represents the set of all edges; Representing an edge Two endpoint nodes i and j The absolute value of the difference between the failure probabilities is the edge. The larger the value, the more significant the "gradient" of the failure characteristics propagating along this edge is, and the more likely this edge is to be the boundary where the failure occurs. Represents a node j The probability of failure occurrence, subscript fIt represents "fault"; Represents an edge The relevant indicator function or weighting coefficient is used to modify the criteria by combining the physical properties of the edges, which can improve the accuracy of positioning.

[0107] The aforementioned AI intelligent analysis layer unit also includes a fault mode recognition module and a data decision module:

[0108] The fault mode identification module uses a long short-term memory network to analyze multi-source data, integrating real-time SCADA data, smart meter information, fault indicator monitoring data, protection device action signals, and 95598 customer service work orders. It identifies short-circuit / ground fault characteristics, combines historical fault information of the distribution network, judges the fault situation of the distribution network (such as cascading faults caused by overload), and outputs the corresponding 95598 customer service work order handling results according to the fault situation.

[0109] The data decision module employs deep reinforcement learning. By performing deep learning on a large amount of historical fault information of the power distribution network and the corresponding handling results of 95598 customer service work orders, it provides preliminary analysis results of the fault causes.

[0110] The aforementioned data cleaning rules and feature processing rules include:

[0111] Data cleaning rules are used to repair missing data, handle anomalies, and verify consistency.

[0112] Feature processing rules are used to numerically encode and standardize the cleaned data.

[0113] Specifically, data missing repair involves: for missing data, for discrete quantities such as key switch states and protection action signals, forward padding or setting them to a default invalid state is used; for continuous current and voltage measurements, data from adjacent or upstream nodes is used for interpolation based on topology, and if interpolation is not possible, an anomaly is marked.

[0114] The data is anomaly handled by: setting thresholds based on electrical laws (e.g., current cannot be negative and voltage should fluctuate within the rated range), identifying and eliminating obviously erroneous outliers; and detecting non-logical changes (e.g., frequent switching on and off within 1 second) as communication interference rather than a real state.

[0115] Perform consistency checks on the data, specifically: for example, check whether the switch status and the corresponding line current value logically match (if the switch is open but the current is not zero, an alarm should be triggered), etc.

[0116] Therefore, by processing the data using the aforementioned data cleaning and feature processing rules, a high-quality data foundation can be provided for subsequent analysis.

[0117] The feature processing rules transform the cleaned raw data into features more suitable for AI model understanding. Specifically, state signals such as "separation / combination" and "action / inaction" are binarized using 0 / 1 encoding; while category features such as device type and phase are encoded using one-hot encoding to avoid erroneous size correlations in the model. Feature processing reduces the learning difficulty of the model, improving the accuracy and generalization ability of fault diagnosis.

[0118] The aforementioned AI intelligent analysis layer unit also includes the use of a centralized learning framework to achieve centralized aggregation and collaborative analysis of cross-system data, and the training of DRL models through a historical fault case library to support continuous optimization of fault assessment strategies.

[0119] Example 2: Figure 4 As shown in the figure, this invention discloses an intelligent fault assessment method for distribution networks based on artificial intelligence multi-source data fusion. This method is implemented using an artificial intelligence-based intelligent fault assessment system for distribution networks based on multi-source data fusion, and includes the following steps:

[0120] S101 performs data fusion and preprocessing on multi-source data of the distribution network. The multi-source data includes integrated SCADA real-time data, smart meter information, fault indicator monitoring data, protection device action signals and 95598 customer service work orders.

[0121] S102, the preprocessed multi-source data of the distribution network is modeled into a dynamic heterogeneous graph structure, time-series fault mode identification is performed, and fault judgment results are output;

[0122] S103 automatically outputs an assessment report based on the fault assessment results and visualizes the fault by overlaying a single-line graph.

[0123] Step S101 involves data fusion and preprocessing of multi-source data from the distribution network, including:

[0124] Based on the time series alignment algorithm, data with different sampling rates from multiple sources are matched, and the sampling frequency of multiple sources is synchronized to make the data frequency and time of multiple sources consistent.

[0125] The isolated forest algorithm is used to detect abnormal data in multi-source data with synchronized sampling frequency. For abnormal data, anomalies are first classified according to remote signaling data, telemetry data and topological relationship data, and then the remote signaling data, telemetry data and topological relationship data are comprehensively analyzed.

[0126] Step S102 involves modeling the preprocessed multi-source data of the distribution network into a dynamic heterogeneous graph structure, performing time-series fault mode identification, and outputting fault assessment results, including:

[0127] Based on graph neural networks and topology inference engines, the preprocessed multi-source data of the distribution network is modeled as a dynamic heterogeneous graph structure, with feature vectors embedded in nodes and edges. Graph attention networks are used to calculate the probability of fault propagation between nodes and to locate fault sections.

[0128] The expression for the dynamic heterogeneity graph is as follows:

[0129] G t =(V t E t ,X t V ,X t E ),

[0130] In the formula, G t V represents the overall topology and operating status of the distribution network at time t; t E represents the set of all nodes at time t; t X represents the set of all edges at time t; t V X represents the set of features or attributes of all nodes at time t; t E It represents the set of features or attributes of all edges at time t;

[0131] Edge e embedded electrical distance Z ij Topology dynamic updates are triggered by adjacency matrix reconstruction through switch state changes, as shown in the following expression:

[0132] ,

[0133] In the formula, This represents the number of nodes in the adjacency matrix at time step t. i With nodes j The connection weights between them; e represents the exponential decay function; Represents electrical distance Z ij The modulus; and Representing nodes respectively i and nodes j The on / off state;

[0134] ,

[0135] In the formula, express Time-based fault mode Below, distribution network nodes i For neighboring nodes j The fault-related attention coefficient; This indicates an exponential enhancement operator that amplifies the differential correlation between fault features through nonlinear mapping, thereby enhancing the distinguishability of abnormal patterns. This represents an asymmetric activation operator with leakage, which preserves fine-grained gradients of fault characteristics and prevents fault information from being annihilated during propagation. Indicates the fault mode The attention parameter vector is transposed to customize feature association weights for different fault types; This represents an element-wise modulation operator that achieves element-wise weighting of attention parameters and feature interaction vectors, accurately capturing the local correlations of fault features; express The spatiotemporal linear transformation matrix at time step, integrating the static topological characteristics of the distribution network nodes with... The dynamic operational characteristics at any given moment are used to complete the spatiotemporal mapping of fault characteristics; , and They represent Time Node i ,node j and nodes k The fault state feature vector; and They represent Time-based fault mode Next node ij and nodes ik The power imbalance feature vector; express Time-based fault mode Next node i A dynamic set of neighbors; Represents nodes in fault scenarios i The normalized summation operator on the dynamic neighbor set constrains the distribution of attention coefficients and ensures the rationality of the weights for fault information propagation.

[0136] Overlaying spatiotemporal feature extraction layers:

[0137] ,

[0138] In the formula, Indicates at time step t The node feature matrix obtained after passing through the spatiotemporal feature extraction layer; This represents a convolutional layer in a graph attention network, used to extract spatial dependency features between nodes from a graph structure. This represents a temporal convolutional layer, used to extract temporal evolutionary features from a sequence of node features; Indicates at time step t The original node feature matrix of the input model; Represents the edge index tensor;

[0139] The node failure probability is calculated as follows:

[0140] ,

[0141] In the formula, Represents a node i The probability of failure occurrence, subscript f It represents "fault"; This represents the Sigmoid activation function; This represents the learnable weight matrix of the fault classification layer; Represents a node i In the L The feature vector of the layer;

[0142] The method for calculating the location of the faulty section is as follows:

[0143] ,

[0144] In the formula, This indicates the finally located faulty section; This represents an optimization operation that iterates through all edges belonging to the edge set. edge Find the value of the objective function. The largest edge; Represents an edge, connecting nodes i and nodes j ; Represents the set of all edges; Representing an edge Two endpoint nodes i and j The absolute value of the difference between the failure probabilities; Represents a node j The probability of failure occurrence, subscript f It represents "fault"; Represents an edge The relevant indicator function or weighting coefficient.

[0145] Step S102, which involves modeling the preprocessed multi-source data of the distribution network into a dynamic heterogeneous graph structure, performing time-series fault mode identification, and outputting fault assessment results, also includes:

[0146] Long Short-Term Memory Network (LSTM) is used to analyze multi-source data, which integrates real-time SCADA data, smart meter data, fault indicator monitoring data, protection device action signals, and 95598 customer service work orders. This identifies short-circuit / ground fault characteristics, combines historical fault information of the distribution network to assess the fault situation, and outputs the corresponding 95598 customer service work order handling results based on the fault situation.

[0147] By employing deep reinforcement learning, and through deep learning on a large amount of historical fault information of the power distribution network and the corresponding handling results of 95598 customer service work orders, preliminary results of fault cause analysis are given.

[0148] Unlike traditional methods that rely solely on data from a single system such as SCADA, this invention innovatively employs a dynamic heterogeneous graph structure (G). t This model serves as the core data model. It not only integrates heterogeneous data from SCADA, fault indicators, smart meters, protection device action signals, and even 95598 customer service work orders, but more importantly, it uses Graph Neural Networks (GNN) and Long Short-Term Memory (LSTM) technologies to simultaneously capture the spatial (along the power grid topology) propagation characteristics and temporal sequential evolution patterns of faults. This deep fusion analysis across both spatial and temporal dimensions enables the system to understand complex fault chain reactions, effectively distinguish between real faults and transient anomalies, thereby fundamentally reducing the common problems of misjudgment, missed judgment, and overjudgment caused by incomplete information in existing technologies, and achieving precise fault location.

[0149] This invention constructs a closed-loop intelligent system integrating "rapid perception, intelligent analysis, self-optimization, and intuitive presentation." On one hand, by learning from historical handling cases through a deep reinforcement learning (DRL) model, the system can continuously optimize its judgment strategies, forming an "expert brain" with self-learning capabilities, thus improving the accuracy of judgments. On the other hand, the judgment results not only automatically generate reports but also visualize them, accurately overlaying abstract fault information onto a single-line graph, greatly improving dispatchers' intuitive understanding of the fault situation and on-site handling efficiency. This comprehensive intelligence, from backend algorithms to frontend applications, significantly enhances the system's practical value and adaptability.

[0150] Example 3: This embodiment of the invention uses data from a historical fault case database in a certain area to test and verify the dynamic GNN method of the present invention compared with the traditional static analysis method. The results are shown in Tables 1 and 2, which illustrate the calculation tables for the improvement in positioning accuracy and the reduction in misjudgment rate.

[0151] As shown in Table 1, the static topology used in traditional methods may no longer match the actual network structure when a fault occurs, leading to misjudgments. This invention effectively solves this core problem through dynamic topology analysis, thus improving the accuracy of fault location.

[0152] As shown in Table 2, the present invention can effectively identify and filter most of the false signals generated by abnormal acquisition devices through multimodal fusion and cross-validation mechanism, thereby significantly reducing the overall system's misjudgment rate (by more than 30%) on a high baseline, and significantly improving the reliability of power distribution network fault assessment.

[0153] Table 1 Calculation Table for Improving Positioning Accuracy

[0154]

[0155] Table 2 Calculation Table for Reducing False Positive Rate

[0156]

Claims

1. An intelligent fault assessment system for distribution networks based on artificial intelligence and multi-source data fusion, characterized in that, include: The data fusion and preprocessing unit performs data fusion and preprocessing on multi-source data of the distribution network. The multi-source data includes integrated SCADA real-time data, smart meter information, fault indicator monitoring data, protection device action signals and 95598 customer service work orders. The AI ​​intelligent analysis layer unit models the pre-processed multi-source data of the distribution network into a dynamic heterogeneous graph structure, performs time-series fault mode identification, and outputs fault assessment results. The application unit automatically outputs an assessment report based on the fault assessment results and visualizes the fault by overlaying a single-line graph. The AI ​​intelligent analysis layer unit includes a dynamic graph construction module; The dynamic graph construction module, based on graph neural networks and topology inference engine, models the preprocessed multi-source data of the distribution network into a dynamic heterogeneous graph structure. Nodes and edges are embedded with feature vectors, and graph attention network is used to calculate the fault propagation probability between nodes and locate fault sections. The expression for the dynamic heterogeneity graph is as follows: G t =(V t ,E t ,X t V ,X t E ), In the formula, G t V represents the overall topology and operating status of the distribution network at time t; t E represents the set of all nodes at time t; t Let X represent the set of all edges at time t; t V X represents the set of features or attributes of all nodes at time t; t E It represents the set of features or attributes of all edges at time t; Edge e embedded electrical distance Z ij Topology dynamic updates are triggered by adjacency matrix reconstruction through switch state changes, as shown in the following expression: ， In the formula, This represents the number of nodes in the adjacency matrix at time step t. i With nodes j The connection weights between them; e represents the exponential decay function; Represents electrical distance Z ij The modulus; and Representing nodes respectively i and nodes j The on / off state; Design fault feature attention coefficient: ， In the formula, express Time-based fault mode Below, distribution network nodes i For neighboring nodes j The fault-related attention coefficient; This indicates an exponential enhancement operator that amplifies the differential correlation between fault features through nonlinear mapping, thereby enhancing the distinguishability of abnormal patterns. This represents an asymmetric activation operator with leakage, which preserves fine-grained gradients of fault characteristics and prevents fault information from being annihilated during propagation. Indicates the fault mode The attention parameter vector is transposed to customize feature association weights for different fault types; This represents an element-wise modulation operator that achieves element-wise weighting of attention parameters and feature interaction vectors, accurately capturing the local correlations of fault features; express The spatiotemporal linear transformation matrix at time step, integrating the static topological characteristics of the distribution network nodes with... The dynamic operational characteristics at any given moment are used to complete the spatiotemporal mapping of fault characteristics; , and They represent Time Node i ,node j and nodes k The fault state feature vector; and They represent Time-based fault mode Next node ij and nodes ik The power imbalance feature vector; express Time-based fault mode Next node i A dynamic set of neighbors; Represents nodes in fault scenarios i The normalized summation operator on the dynamic neighbor set constrains the distribution of attention coefficients and ensures the rationality of the weights for fault information propagation. Overlaying spatiotemporal feature extraction layers: ， In the formula, Indicates at time step t The node feature matrix obtained after the spatiotemporal feature extraction layer; This represents a convolutional layer in a graph attention network, used to extract spatial dependency features between nodes from a graph structure. This represents a temporal convolutional layer, used to extract temporal evolutionary features from a sequence of node features; Indicates at time step t The original node feature matrix of the input model; Represents the edge index tensor; The node failure probability is calculated as follows: ， In the formula, Represents a node i The probability of failure occurrence, subscript f It represents "fault"; This represents the Sigmoid activation function; This represents the learnable weight matrix of the fault classification layer; Represents a node i In the L The feature vector of the layer; The method for calculating the location of the faulty section is as follows: ， In the formula, This indicates the finally located faulty section; This represents an optimization operation that iterates through all edges belonging to the edge set. edge Find the value of the objective function. The largest edge; Represents an edge, connecting nodes i and nodes j ; Represents the set of all edges; Representing an edge Two endpoint nodes i and j The absolute value of the difference between the failure probabilities; Represents a node j The probability of failure occurrence, subscript f It represents "fault"; Represents an edge The relevant indicator function or weighting coefficient.

2. The intelligent fault assessment system for distribution networks based on artificial intelligence multi-source data fusion according to claim 1, characterized in that, The data fusion and preprocessing unit also includes a frequency synchronization module and a data detection module; The frequency synchronization module uses a time series alignment algorithm to match different sampling rate data from multiple sources and synchronizes the sampling frequencies of the multiple sources to ensure that the data frequencies and times of the multiple sources are consistent. The data detection module uses the isolated forest algorithm to detect abnormal data in multi-source data with synchronized sampling frequency. For abnormal data, it first classifies the abnormal data according to remote signaling data, telemetry data and topological relationship data, and then performs a comprehensive analysis of remote signaling data, telemetry data and topological relationship data.

3. The intelligent fault assessment system for multi-source data fusion in power distribution networks based on artificial intelligence as described in claim 1, characterized in that, The AI ​​intelligent analysis layer unit also includes a fault mode recognition module and a data decision module: The fault mode identification module uses a long short-term memory network to analyze multi-source data, which integrates real-time SCADA data, smart meter information, fault indicator monitoring data, protection device action signals, and 95598 customer service work orders. It identifies short-circuit / ground fault characteristics, combines historical fault information of the distribution network, judges the fault situation of the distribution network, and outputs the corresponding 95598 customer service work order handling results according to the fault situation. The data decision module employs deep reinforcement learning. By performing deep learning on a large amount of historical fault information of the power distribution network and the corresponding handling results of 95598 customer service work orders, it provides preliminary analysis results of the fault causes.

4. The intelligent fault assessment system for distribution networks based on artificial intelligence multi-source data fusion according to claim 1, characterized in that, The application unit also includes a data cleaning rule module and a feature processing rule module; The data cleaning rules module is used to perform missing data repair, anomaly handling, and consistency verification. The feature processing rule module is used to numerically encode and standardize the cleaned data.

5. The intelligent fault assessment system for distribution networks based on artificial intelligence multi-source data fusion according to claim 1, characterized in that, The AI ​​intelligent analysis layer unit also includes an optimization module. The optimization module adopts a centralized learning framework to realize the centralized aggregation and collaborative analysis of cross-system data. It trains the DRL model through a historical fault case library to support the continuous optimization of fault judgment strategies.

6. A method for intelligent fault assessment of distribution networks based on artificial intelligence multi-source data fusion, characterized in that, This method is implemented using the AI-based multi-source data fusion fault intelligent judgment system for distribution networks as described in any one of claims 1-5, and includes the following steps: Data fusion and preprocessing are performed on multi-source data of the distribution network. The multi-source data includes integrated SCADA real-time data, smart meter information, fault indicator monitoring data, protection device action signals and 95598 customer service work orders. The preprocessed multi-source data of the distribution network is modeled into a dynamic heterogeneous graph structure, time-series fault mode identification is performed, and fault assessment results are output. The system automatically generates an assessment report based on the fault assessment results and visualizes the fault by overlaying single-line graphs. The process of modeling the preprocessed multi-source data of the distribution network into a dynamic heterogeneous graph structure, performing time-series fault mode identification, and outputting fault assessment results includes: Based on graph neural networks and topology inference engines, the preprocessed multi-source data of the distribution network is modeled as a dynamic heterogeneous graph structure, with feature vectors embedded in nodes and edges. Graph attention networks are used to calculate the probability of fault propagation between nodes and to locate fault sections. The expression for the dynamic heterogeneity graph is as follows: G t =(V t ,E t ,X t V ,X t E ), In the formula, G t V represents the overall topology and operating status of the distribution network at time t; t E represents the set of all nodes at time t; t Let X represent the set of all edges at time t; t V X represents the set of features or attributes of all nodes at time t; t E It represents the set of features or attributes of all edges at time t; Edge e embedded electrical distance Z ij Topology dynamic updates are triggered by adjacency matrix reconstruction through switch state changes, as shown in the following expression: ， In the formula, This represents the number of nodes in the adjacency matrix at time step t. i With nodes j The connection weights between them; e represents the exponential decay function; Represents electrical distance Z ij The modulus; and Representing nodes respectively i and nodes j The on / off state; Design fault feature attention coefficient: ， In the formula, express Time-based fault mode Below, distribution network nodes i For neighboring nodes j The fault-related attention coefficient; This indicates an exponential enhancement operator that amplifies the differential correlation between fault features through nonlinear mapping, thereby enhancing the distinguishability of abnormal patterns. This represents an asymmetric activation operator with leakage, which preserves fine-grained gradients of fault characteristics and prevents fault information from being annihilated during propagation. Indicates the fault mode The attention parameter vector is transposed to customize feature association weights for different fault types; This represents an element-wise modulation operator that achieves element-wise weighting of attention parameters and feature interaction vectors, accurately capturing the local correlations of fault features; express The spatiotemporal linear transformation matrix at time step, integrating the static topological characteristics of the distribution network nodes with... The dynamic operational characteristics at any given moment are used to complete the spatiotemporal mapping of fault characteristics; , and They represent Time Node i ,node j and nodes k The fault state feature vector; and They represent Time-based fault mode Next node ij and nodes ik The power imbalance feature vector; express Time-based fault mode Next node i A dynamic set of neighbors; Represents nodes in fault scenarios i The normalized summation operator on the dynamic neighbor set constrains the distribution of attention coefficients and ensures the rationality of the weights for fault information propagation. Overlaying spatiotemporal feature extraction layers: ， In the formula, Indicates at time step t The node feature matrix obtained after the spatiotemporal feature extraction layer; This represents a convolutional layer in a graph attention network, used to extract spatial dependency features between nodes from a graph structure. This represents a temporal convolutional layer, used to extract temporal evolutionary features from a sequence of node features; Indicates at time step t The original node feature matrix of the input model; Represents the edge index tensor; The node failure probability is calculated as follows: ， In the formula, Represents a node i The probability of failure occurrence, subscript f It represents "fault"; This represents the Sigmoid activation function; This represents the learnable weight matrix of the fault classification layer; Represents a node i In the L The feature vector of the layer; The method for calculating the location of the faulty section is as follows: ， In the formula, This indicates the finally located faulty section; This represents an optimization operation that iterates through all edges belonging to the edge set. edge Find the value of the objective function. The largest edge; Represents an edge, connecting nodes i and nodes j ; Represents the set of all edges; Representing an edge Two endpoint nodes i and j The absolute value of the difference between the failure probabilities; Represents a node j The probability of failure occurrence, subscript f It represents "fault"; Represents an edge The relevant indicator function or weighting coefficient.

7. The method for intelligent fault assessment of distribution networks based on artificial intelligence multi-source data fusion according to claim 6, characterized in that, The data fusion and preprocessing of multi-source data from the distribution network includes: Based on the time series alignment algorithm, data with different sampling rates from multiple sources are matched, and the sampling frequency of multiple sources is synchronized to make the data frequency and time of multiple sources consistent. The isolated forest algorithm is used to detect abnormal data in multi-source data with synchronized sampling frequency. For abnormal data, anomalies are first classified according to remote signaling data, telemetry data and topological relationship data, and then the remote signaling data, telemetry data and topological relationship data are comprehensively analyzed.

8. The method for intelligent fault assessment of distribution networks based on artificial intelligence multi-source data fusion according to claim 6, characterized in that, The step of modeling the preprocessed multi-source data of the distribution network into a dynamic heterogeneous graph structure, performing time-series fault mode identification, and outputting fault assessment results also includes: Long Short-Term Memory Network (LSTM) is used to analyze multi-source data, which integrates real-time SCADA data, smart meter data, fault indicator monitoring data, protection device action signals, and 95598 customer service work orders. This identifies short-circuit / ground fault characteristics, combines historical fault information of the distribution network to assess the fault situation, and outputs the corresponding 95598 customer service work order handling results based on the fault situation. By employing deep reinforcement learning, and through deep learning on a large amount of historical fault information of the power distribution network and the corresponding handling results of 95598 customer service work orders, preliminary results of fault cause analysis are given.

Images