Fusion power distribution network fault prediction method and system based on multi-source data

Abstract

The invention relates to the technical field of power system automation, and discloses a fusion power distribution network fault prediction method and system based on multi-source data. The method comprises the following steps: acquiring real-time operation monitoring data of the power distribution network, and analyzing to obtain a real-time topological adjacency matrix and multi-source measurement data; performing feature fusion on the multi-source measurement data to obtain a fused heterogeneous feature vector; based on the real-time topological adjacency matrix, utilizing a graph neural network model to aggregate the fused heterogeneous feature vectors to obtain space node embedding; generating time sequence evolution characteristic representation by using a long short-term memory network model; determining potential risk nodes based on the feature space distance and matching a historical library to determine a fault type; and deducing a fault evolution trend and generating a disposal suggestion. According to the method, accurate fault early warning under a complex dynamic working condition can be realized.

Description

Technical Field

[0001] This invention relates to the field of power system automation and intelligent monitoring technology, and in particular to a method and system for predicting distribution network faults based on multi-source data fusion. Background Technology

[0002] Currently, with the integration of numerous distributed power sources, electric vehicle charging stations, and other interactive devices, the operating conditions of power distribution networks are becoming increasingly complex, significantly increasing the potential risks of faults. To shift from "passive emergency repair" to "proactive operation and maintenance," utilizing massive amounts of measurement data for real-time fault prediction and risk assessment has become an inevitable trend in the industry.

[0003] In existing technologies, fault prediction primarily relies on smart monitoring terminals deployed at various nodes of the power grid. These terminals typically carry MCUs (microcontrollers) and various sensor chips to collect electrical quantity data such as voltage and current. Traditional methods usually assume that the physical connections of the distribution network remain unchanged, using historical data to train machine learning models (such as support vector machines or shallow neural networks) to predict fault locations by identifying abnormal fluctuations in electrical quantity characteristics. However, the topology of actual distribution networks is not static but undergoes frequent dynamic reconfiguration due to scheduling operations, line maintenance, or fault isolation. Most existing prediction methods are trained under fixed network topology assumptions. Once line switching changes the connection relationships between nodes, the original model cannot adapt to the new spatial dependencies, leading to misjudgments of current and voltage propagation patterns. Simultaneously, the distribution network contains multiple heterogeneous data sources, such as smart meters (low-frequency) and micro PMUs (high-frequency). Existing technologies often struggle to address the dynamic alignment challenges of these multi-source heterogeneous data in time and space during topology changes, causing a sharp decline in the model's feature extraction capability under the new topology, resulting in a large number of false alarms or missed alarms.

[0004] Therefore, existing technologies suffer from low fault prediction accuracy and poor real-time performance. Summary of the Invention

[0005] This invention provides a method and system for predicting faults in distribution networks based on multi-source data, in order to solve the technical problems of low fault prediction accuracy and poor real-time performance in the prior art.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides a method for predicting faults in a distribution network based on multi-source data, comprising: Real-time operation monitoring data of the power distribution network is acquired, and topology analysis and data extraction are performed to obtain a real-time topology adjacency matrix and multi-source measurement data including smart meter data and micro phasor measurement data. Multi-scale time feature fusion is performed on the multi-source measurement data to obtain the fused heterogeneous feature vector of each network node in the distribution network; Based on the real-time topological adjacency matrix, a pre-set graph neural network model is used to aggregate the spatial dimensions of the fused heterogeneous feature vectors to obtain the spatial node embedding of the network nodes. A temporal dimension analysis of the spatial node embedding is performed using a pre-built long short-term memory network model to generate a temporal evolution feature representation of the network nodes; Obtain the normal operation reference vector corresponding to the current working condition, and calculate the feature space distance between the time-series evolution feature representation and the normal operation reference vector. Network nodes whose feature space distance exceeds a preset risk judgment threshold are identified as potential risk nodes. The temporal evolution feature representation of the potential risk node is matched with a pre-set historical fault pattern library to determine the potential fault type; Based on the location of the potential risk nodes in the topology and the potential fault types, the fault evolution trend is deduced and disposal suggestions are generated.

[0007] Secondly, the present invention provides a fusion distribution network fault prediction system based on multi-source data, comprising: The data acquisition and parsing module is used to acquire real-time operation monitoring data of the distribution network, and perform topology parsing and data extraction to obtain a real-time topology adjacency matrix and multi-source measurement data including smart meter data and micro phasor measurement data. The multi-scale feature fusion module is used to perform multi-scale time feature fusion on the multi-source measurement data to obtain the fused heterogeneous feature vector of each network node in the distribution network. The spatial dimension aggregation module is used to perform spatial dimension aggregation on the fused heterogeneous feature vector based on the real-time topological adjacency matrix and using a pre-set graph neural network model to obtain the spatial node embedding of the network node. The time dimension analysis module is used to perform time dimension analysis on the spatial node embedding using a pre-set long short-term memory network model, and generate a temporal evolution feature representation of the network node. The risk assessment module is used to obtain the normal operation reference vector corresponding to the current working condition, calculate the feature space distance between the time-series evolution feature representation and the normal operation reference vector, and identify network nodes whose feature space distance exceeds a preset risk assessment threshold as potential risk nodes. The fault characterization module is used to match the temporal evolution feature representation of the potential risk node with a pre-set historical fault pattern library to determine the potential fault type; The trend prediction and decision-making module is used to predict the evolution trend of the fault and generate disposal suggestions based on the location of the potential risk node in the topology and the potential fault type.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention constructs a real-time topological adjacency matrix by collecting real-time operational monitoring data and combines it with a graph neural network for spatial feature aggregation. Since the operation mode of the distribution network frequently changes due to switching operations, load transfer, or fault isolation, traditional models based on static topological assumptions often fail after network reconstruction. This invention utilizes the adaptive processing capability of graph neural networks for graph structure data, explicitly embedding the real-time parsed physical connection relationships into the node feature aggregation process, enabling the model to dynamically perceive changes in the node's neighborhood structure. This effectively solves the problem that existing technologies cannot adapt to dynamic changes in the distribution network topology, improving the generalization ability and positioning accuracy of the prediction model under complex and variable operating conditions.

[0009] (2) This invention constructs a dual-channel input containing long-period steady-state features and short-period dynamic features, and utilizes a multi-head attention mechanism to achieve adaptive weighted fusion of multi-source heterogeneous data. Addressing the challenge of significant differences in sampling frequencies between smart meters (low frequency) and micro phasor measurement units (high frequency), this invention abandons simple interpolation completion methods and instead uses an attention mechanism to allow the model to automatically learn the importance weights of features at different time scales at the current moment. Steady-state features provide background power flow constraints, while dynamic features provide details of abrupt changes; the two complement each other. This effectively solves the technical problem of aligning multi-source heterogeneous data in the same feature space, avoids noise interference introduced by interpolation, and improves the sensitivity to capturing weak early fault signs.

[0010] (3) This invention generates a dynamic normal operation reference vector corresponding to the current operating condition and performs risk assessment based on the feature space distance. Traditional fault monitoring usually uses fixed thresholds, which are prone to false alarms during load peaks or normal fluctuations. This invention adopts a "dynamic reference" strategy to reconstruct the "normal state" that should exist under the current operating condition in real time and calculate the degree of deviation between the actual state and the dynamic reference. This effectively solves the problem of high false alarm rate caused by normal load fluctuations in the distribution network, and achieves accurate "foresight" and proactive early warning of fault evolution trends while ensuring a low false alarm rate. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the process of the fusion distribution network fault prediction method based on multi-source data provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the integrated distribution network fault prediction system based on multi-source data provided in the second embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Reference Figure 1 The first embodiment of the present invention provides a method for predicting faults in a fusion distribution network based on multi-source data, comprising the following steps: S11: Obtain real-time operation monitoring data of the distribution network, perform topology analysis and data extraction to obtain real-time topology adjacency matrix and multi-source measurement data including smart meter data and micro phasor measurement data; S12, perform multi-scale time feature fusion on the multi-source measurement data to obtain the fused heterogeneous feature vector of each network node in the distribution network; S13, Based on the real-time topological adjacency matrix, the spatial dimension of the fused heterogeneous feature vector is aggregated using a pre-set graph neural network model to obtain the spatial node embedding of the network node. S14, using a pre-set long short-term memory network model to perform time dimension analysis on the spatial node embedding, and generate a temporal evolution feature representation of the network node; S15, obtain the normal operation reference vector corresponding to the current working condition, and calculate the feature space distance between the time-series evolution feature representation and the normal operation reference vector, and determine the network nodes whose feature space distance exceeds the preset risk judgment threshold as potential risk nodes; S16, Match the temporal evolution feature representation of the potential risk node with a pre-set historical fault mode library to determine the potential fault type; S17. Based on the location of the potential risk node in the topology and the potential fault type, deduce the fault evolution trend and generate handling suggestions.

[0014] In step S11, real-time operation monitoring data of the distribution network is acquired, and topology analysis and data extraction are performed to obtain a real-time topology adjacency matrix and multi-source measurement data containing smart meter data and micro phasor measurement data, including: The status of line switches, equipment operating parameters, and measurement device data of the power distribution network are collected as the real-time operation monitoring data; Separate and extract line switching status data and equipment operation status data from the real-time operation monitoring data, and identify the effective network nodes in the distribution network based on the line switching status data and the equipment operation status data; The on / off status of each connected line is determined based on the line switching status data, and the conductive connection path between the effective network nodes is determined. Construct an initial two-dimensional array, and assign connectivity values ​​at the row and column intersection positions corresponding to the effective network nodes and the connected connection paths in the initial two-dimensional array to obtain the real-time topology adjacency matrix; The smart meter data and the micro phasor measurement data are filtered and extracted from the real-time operation monitoring data to serve as the multi-source measurement data.

[0015] In one implementation, this embodiment utilizes a distribution automation master station system to establish communication links with edge gateways deployed at various physical sites. Based on the IEC 61850 communication protocol or Modbus-TCP protocol, it periodically polls the entire network data to collect the real-time operation monitoring data. This embodiment strictly sets the system sampling frequency to 100Hz, meaning data is refreshed every 10 milliseconds. This sampling frequency setting is based on a combination of the mechanical response time characteristics of distribution network circuit breaker actions and the Nyquist sampling theorem. Typically, the full opening time of a medium-voltage circuit breaker is between 30 and 100 milliseconds, and the cycle of power frequency AC is 20 milliseconds. Setting a sampling period of 10 milliseconds not only ensures that the system fully captures the dynamic change process of the switch action, preventing missed detections of topology state transitions due to undersampling, but also meets the minimum bandwidth requirement for half-wave reconstruction of the power frequency signal.

[0016] In one implementation, this embodiment employs a global static index mapping strategy to address the adaptation problem between dynamic changes in the distribution network topology and the fixed input dimension of the graph neural network, in order to identify the valid network nodes. The system pre-establishes a global asset index table containing all physically existing poles, transformers, switches, and buses in the distribution network, and assigns a fixed global index ID to each physical device, ranging from 0 to... ,in This represents the upper limit of the total number of physical nodes in the entire network. During the real-time operation phase, this embodiment traverses the real-time operation monitoring data and reads the device status bit and telemetry voltage value corresponding to each global index ID. Only when the status bit of a device is a logical "1" representing the operational state and its associated telemetry voltage value is greater than a preset energized threshold (e.g., 10% of the rated voltage), the system marks the global index ID as valid and includes it in the set of valid network nodes; otherwise, if the status bit is a logical "0" representing the maintenance or isolation state, it is marked as invalid.

[0017] It should be noted that the energization threshold is determined based on the induced voltage level of the distribution network lines, and is usually set to 10% of the rated voltage. The reason for choosing this value is that out-of-service lines often have false voltages due to electromagnetic induction from nearby energized lines. These induced voltages are usually lower than 5% of the rated value. Therefore, setting a threshold of 10% can effectively filter induced electrical interference and ensure the accuracy of topology node status identification.

[0018] In one implementation, for constructing the real-time topological adjacency matrix, this embodiment first constructs a matrix with a fixed dimension. An initial two-dimensional zero matrix is ​​constructed, whose rows and columns strictly correspond to the aforementioned global asset index table. Subsequently, this embodiment analyzes physical connection relationships based on the Common Information Model (CIM) of the distribution network and performs connectivity judgment in conjunction with the real-time collected remote signaling status of switches. For any two nodes with global index IDs u and v, if there is a physical line between them and the remote signaling values ​​of all series switches on that line are "1" (closed), then a conductive connection path is determined to exist. In this embodiment, the positions at coordinates (u,v) and (v,u) in the matrix are assigned preset connectivity values.

[0019] It should be noted that, to reflect the strength of electrical coupling between nodes, the connectivity value is set to the modulus of the admittance of the connecting line (i.e., the reciprocal of the impedance), rather than a simple binarized value. For nodes marked as invalid or non-conductive paths, their corresponding elements in the matrix remain zero. This zero-padding method ensures that the dimension of the output matrix remains unchanged regardless of how the power grid topology is reconstructed, thus meeting the computational requirements of the subsequent graph neural network for fixed tensor dimensions.

[0020] In one implementation, for extracting the multi-source measurement data, this embodiment performs data cleaning and splitting operations based on message signature codes. The real-time operation monitoring data stream contains mixed messages of multiple protocols. This embodiment sets up two parallel filtering channels. The first channel matches messages with signature codes DL / T 645 or AMI_Header, extracts and parses their payload data into low-frequency smart meter data, including active energy and effective voltage values. The second channel matches messages with signature codes C37.118 or PMU_Frame, extracts and parses their payload data into high-frequency micro-phasor measurement data, including synchronization phasors, frequency, and rate of change of frequency. After these two sets of data are extracted, they are stored in corresponding circular buffers, awaiting the next step of feature fusion.

[0021] In step S12, multi-scale time feature fusion is performed on the multi-source measurement data to obtain the fused heterogeneous feature vector of each network node in the distribution network, including: The high-frequency sampling time of the micro phasor measurement data is used as the reference time axis; Using a sample-and-hold algorithm or a trend extrapolation algorithm, the low-frequency sampled smart meter data is aligned to the reference time axis to construct a long-period steady-state feature sequence aligned with the timestamp of the micro phasor measurement data; The micro phasor measurement data is used as a short-period dynamic feature sequence and vector-concatenated with the long-period steady-state feature sequence to construct a heterogeneous data matrix. The heterogeneous data matrix is ​​input into a pre-set multi-head attention computing model, and the weight distribution of features at different time scales is calculated using the attention mechanism. Adaptive weighted fusion is then performed to obtain the fused heterogeneous feature vector.

[0022] In one implementation, for constructing the long-period steady-state feature sequence, this embodiment takes into account the order-of-magnitude difference in sampling frequency between smart meters (AMI) and micro-phasor measurement units (micro-PMU) (e.g., minute-level versus millisecond-level), and adopts a high-frequency reference alignment and zero-order hold strategy.

[0023] Specifically, this embodiment uses each high-frequency sampling time t of the micro-PMU as the primary key index to retrieve the most recently updated AMI data frame before that time, and directly copies and fills the active and reactive power values ​​of the AMI data into the data slot at time t. The physical significance of this processing method is to construct a time-varying bias constraint, that is, to use the power level of AMI data, which, although it presents a stepped pattern on the micro time scale, is macroscopically stable, to provide a background constraint for the high-frequency changing PMU voltage phasor data that can reflect the current power flow benchmark of the distribution network, thereby preventing the subsequent model from misjudging normal load fluctuations as fault symptoms.

[0024] In one implementation, before vector concatenation, this embodiment must perform strict data normalization to eliminate numerical imbalances caused by differences in the physical dimensions of different source data. Since smart meter data is measured in kilowatts (kW) or kilovars (kVar), its values ​​are typically in the hundreds to thousands range, while micro-phasor measurement data uses kilovolts (kV) or per-unit values ​​for voltage and radians for phase angle, resulting in a significant difference in magnitude. This embodiment employs the Z-Score normalization method. For each type of data, it first calculates the statistical mean and standard deviation in a pre-stored historical database. Then, it subtracts the statistical mean from the current measured value and divides the difference by the standard deviation, thereby mapping the data of different dimensions to a dimensionless distribution with a mean of zero and a variance of one. Subsequently, this embodiment concatenates the normalized micro-PMU features (including voltage amplitude, phase angle, and frequency) and the normalized AMI features (including active and reactive power) along the feature dimension to form the original heterogeneous feature vector at the current time t. .

[0025] In one implementation, adaptive weighted fusion is performed using a pre-built multi-head attention calculation model. This embodiment aims to address the issue of varying sensitivity of different fault types to different data sources. For example, high-impedance grounding faults are more reflected in the steady-state deviation of active power (AMI sensitivity), while short-circuit faults are more reflected in the abrupt change in voltage phase angle (PMU sensitivity). This embodiment employs a feature-dimensional self-attention mechanism.

[0026] Specifically, the model will input vectors Mapped to query vectors respectively Key vector Sum value vector The calculation formula is: in This is a learnable weight matrix. Then, the attention score is calculated. : This score characterizes the importance weight of different feature channels in the current fault identification task. Finally, this score is used to adjust the value vector. A weighted summation is performed, and the resulting fused heterogeneous feature vector is output after feature reconstruction. This vector automatically suppresses irrelevant noise features and enhances feature components that are strongly correlated with the current potential fault mode.

[0027] In step S13, based on the real-time topological adjacency matrix, a pre-set graph neural network model is used to aggregate the spatial dimensions of the fused heterogeneous feature vectors to obtain the spatial node embedding of the network nodes, including: The fused heterogeneous feature vector is mapped to the initial attribute features of the network node; The topological neighborhood range of the network node is determined based on the real-time topological adjacency matrix, and the attribute characteristics of adjacent nodes within the topological neighborhood range are obtained by traversal. The pre-set graph neural network model is used to perform weighted aggregation calculations on the attribute features of the adjacent nodes to generate aggregated features containing local spatial dependency information. The aggregation features are subjected to a nonlinear activation transformation to update the state representation of the network nodes, thereby obtaining the spatial node embedding.

[0028] In one implementation, regarding the mapping of the fused heterogeneous feature vectors to initial attribute features, this embodiment considers that the dimension of the feature vector output by S12 may not be consistent with the input dimension required by the hidden layer of the graph neural network. Therefore, a linear projection layer is set at the input of the model. This layer utilizes a learnable weight matrix. A linear transformation is performed on the input feature vector to map it to a unified hidden layer feature space, ensuring the consistency of data semantics in subsequent processing.

[0029] In one implementation, this embodiment employs an inductive learning model based on a spectral graph convolutional network (GCN) architecture as the pre-set graph neural network model. To address the computational anomalies caused by isolated nodes in the distribution network and to prevent the loss of node-specific feature information during feature aggregation, this embodiment adopts a self-loop renormalization strategy.

[0030] Specifically, this embodiment does not directly use the original real-time topological adjacency matrix. Instead, it constructs a modified adjacency matrix. ,in It is an identity matrix. This operation logically mandates that each network node has a degree of at least one, and its aggregation range explicitly includes the node itself.

[0031] In one implementation, for the weighted aggregation calculation performed through the pre-set graph neural network model, this embodiment executes a hierarchical propagation rule based on the symmetric normalized Laplacian matrix. The specific calculation formula is as follows: in, Indicates the first The node feature matrix of the layer (for the first layer, That is, the initial attribute features). The modified adjacency matrix with self-loops was added to the above. for The degree matrix, whose diagonal elements ; For the first A layer-learnable shared weight matrix; It is a non-linear activation function (e.g., ReLU); For the output of the first Layer feature matrix.

[0032] It should be noted that the terms in the formula Topological normalization of the aggregation weights is achieved. Its physical significance lies in eliminating the impact of differences in node connection density on characteristic amplitudes in the distribution network. For example, the degree of a substation bus node is much greater than that of an end-user node; without normalization, the characteristic values ​​of high-density nodes will inflate dramatically during aggregation. By multiplying by the inverse square root of the degree matrix, the model can smooth this imbalance in numerical scale, allowing it to focus on learning the spatial physical coupling of voltage and current rather than the order-of-magnitude differences in topological connections.

[0033] In one implementation, to capture the long-distance fault propagation characteristics (e.g., voltage sag range) across multiple physical nodes in a distribution network, this embodiment uses K graph convolutional layers stacked in series (e.g., K=2 or K=3). The number of layers K is chosen based on the electrical effective influence radius of the distribution network; setting K layers means that each node can aggregate information from its K-order neighbors (i.e., the receptive field is K hops). The feature vector output after propagation through K layers is determined as the spatial node embedding of the network node. Due to the weight matrix... It is obtained by sharing the entire graph and training with a large number of samples from different topologies, and this model is capable of handling dynamic changes. The ability to correctly generate embedding vectors that reflect the current local structure after topological reconstruction.

[0034] In step S14, a temporal dimension analysis is performed on the spatial node embedding using a pre-set long short-term memory network model to generate a temporal evolution feature representation of the network nodes, including: For each of the network nodes, the spatial node embeddings of the network node at multiple consecutive time points are extracted using the sliding window algorithm to construct a temporal feature input tensor; The temporal features are input into the tensor and then into the pre-set long short-term memory network model. The gating mechanism is used to selectively remember and forget the historical state sequence, and the fluctuation trend features in the time dimension are extracted. The temporal evolution feature representation of the network node is obtained by performing a fully connected mapping on the fluctuation trend feature.

[0035] In one implementation, for constructing the temporal feature input tensor, this embodiment maintains a first-in-first-out data queue in memory to cache the spatial node embeddings calculated at historical time points. This embodiment sets a fixed time window length T, and each time S13 outputs the spatial node embedding vector at the current time t... At that time, the system pushes the data into a queue and extracts a continuous sequence of T vectors from time t-T+1 to time t. In this embodiment, the sequence is organized into a three-dimensional tensor structure, whose dimensions are, in order, batch size or number of nodes, time step T, and feature dimension.

[0036] It should be noted that the specific value of the time window length T is strictly calculated based on the physical evolution cycle of the distribution network fault and the sampling frequency. This embodiment statistically analyzes the average duration from the appearance of early weak signs (e.g., arc discharge at the initial stage of high-resistance grounding) to the complete stabilization of fault characteristics (e.g., protection tripping) in the historical fault database; for example, this duration is between 200 and 500 milliseconds. Combining this with the system sampling frequency (i.e., 100 Hz) determined in S11, this embodiment calculates the time window length by multiplying the average duration by the system sampling frequency. For example, if the average duration is 500 milliseconds, the time window length is set to 50 sampling points. This setting ensures that the time series of the input model can completely cover the entire process of a fault from occurrence and development to transient termination, avoiding contextual information truncation due to an excessively short window.

[0037] In one implementation, this embodiment uses a Long Short-Term Memory (LSTM) network as the core of time series analysis to extract fluctuation trend features using a gating mechanism. This model addresses the gradient vanishing problem in long sequence training by introducing cell states and three gating units. For each time step's input vector and the previous hidden state vector, the model executes the following rigorous computational logic: Forget gate calculation is performed, using the Sigmoid activation function to calculate a forgetting coefficient between zero and one, used to determine which historical information from the previous cell state should be discarded; Input gate calculation is performed simultaneously, on the one hand, using the hyperbolic tangent function to generate the candidate state update value for the current time step, and on the other hand, using the Sigmoid function to calculate input weights, used to determine which new information will be written into the cell state; The previous cell state is weighted using the forgetting coefficient, and the candidate state update value weighted by the input weights is superimposed to obtain the updated cell state for the current time step; Output gate calculation is performed, calculating the output weight based on the current input and historical hidden states, and applying this output weight to the current cell state after hyperbolic tangent transformation to obtain the final hidden state vector for the current time step. After recursive calculations for a preset time window length of T time steps, the hidden state vector output by the model in the last time step is determined to be the fluctuation trend feature that aggregates all historical information of the entire window.

[0038] In one implementation, a fully connected mapping is performed on the fluctuation trend features. In this embodiment, a fully connected layer is connected after the Long Short-Term Memory (LSTM) network layer. This layer contains a learnable weight matrix and a bias vector. This embodiment performs a linear transformation operation, specifically including performing matrix multiplication on the fluctuation trend features and the weight matrix, and adding the result to the bias vector, thereby mapping the high-dimensional hidden state output by the LSM network to a feature space of a preset dimension (e.g., 64 dimensions). The resulting vector after mapping is the temporal evolution feature representation.

[0039] It should be noted that the pre-set long short-term memory network model is not an isolated module that works independently. Instead, it is constructed as an end-to-end deep learning architecture with the pre-set multi-head attention calculation model described in S12 and the pre-set graph neural network model described in S13 during the offline stage, and is jointly trained.

[0040] The source and construction process of the sample dataset used in the joint training are as follows. The dataset consists of two parts. The first part comes from a pre-built historical operation database of the distribution network. This database is constructed by continuously collecting historical operation monitoring data of the target distribution network over the past several years (e.g., the last three years) through access to the distribution network SCADA system, AMI measurement and acquisition system, and micro-synchronous phasor measurement device (micro-PMU), and after noise reduction, time alignment, and manual review and cleaning. Specifically, time periods containing real fault waveform data, verified by manual verification, are selected as positive samples, and time periods with stable system operation and no alarm records are selected as negative samples. The second part is enhanced simulation data, which is generated by constructing a simulation model consistent with the actual field topology based on electromagnetic transient simulation software (e.g., PSCAD or ATP-EMTP), by batch setting different fault types (single-phase grounding, phase-to-phase short circuit, etc.), different transition resistances, and different fault locations.

[0041] During training, the two sets of data are mixed as input, and the contrastive loss function or triplet loss function is used as the optimization objective. The weight parameters of the multi-head attention calculation model, graph neural network model and long short-term memory network model are updated synchronously through the backpropagation algorithm, thereby ensuring that the entire feature extraction link can adaptively maximize the discriminative power between feature vectors of different working conditions.

[0042] It is worth further explaining that, regarding the specific structural hyperparameters (including the number of hidden layer nodes, feature vector dimension, and number of encoder layers) and training hyperparameters (including initial learning rate and batch size) of the various neural network models involved in this embodiment (including the multi-head attention calculation model in S12, the graph neural network in S13, the long short-term memory network in S14, and the normal behavior reconstruction model in S15), this embodiment uses a grid search strategy based on the validation set to determine these parameters during the offline training phase. Specifically, the system first constructs a discrete search space for the hyperparameters, for example, setting the search set for the number of hidden layer nodes as {32, 64, 128, 256} and the search set for the initial learning rate as {0.01, 0.001, 0.0001}; then, using historical sample data, it performs multiple rounds of cross-validation training on each parameter combination within the search space, selecting the parameter combination with the smallest mean of the comprehensive loss function on the validation set as the final preset parameters.

[0043] Furthermore, during model training, this embodiment employs the Adam optimizer for gradient descent updates and introduces a dynamic learning rate decay mechanism. The slope of the loss function on the training set is monitored in real time. When a slowing trend is detected or after a preset decay period (e.g., 10 training epochs), the learning rate is multiplied by a decay coefficient (e.g., 0.9), balancing convergence speed in the early stages of training with optimization accuracy in the later stages. Simultaneously, to prevent overfitting, an early stopping mechanism is implemented. This mechanism sets a judgment window based on the oscillation tolerance of the validation set loss value. That is, when the loss value on the validation set no longer decreases within a certain number of consecutive training epochs (e.g., 20, the value being determined based on the average oscillation period of the validation set loss curve), the model is deemed to have reached its performance bottleneck, training is automatically terminated, and the current optimal weights are saved.

[0044] In step S15, a normal operation reference vector corresponding to the current operating condition is obtained, and the feature space distance between the temporal evolution feature representation and the normal operation reference vector is calculated. Network nodes whose feature space distance exceeds a preset risk judgment threshold are identified as potential risk nodes, including: Obtain the operating condition parameters at the current moment, and generate the normal operation reference vector that corresponds to the operating condition parameters and is in the same feature space based on the preset normal behavior reconstruction model; For each of the network nodes, calculate the multidimensional Euclidean distance or cosine similarity between the temporal evolution feature representation of the network node and the normal operation reference vector; The calculated distance value or similarity value is determined as the feature space distance, and it is determined whether the feature space distance exceeds the preset risk judgment threshold. If the feature space distance exceeds the preset risk judgment threshold, the corresponding network node is marked as the potential risk node, and the physical number of the network node is recorded as the potential fault location.

[0045] In one implementation, for generating the normal operation baseline vector based on a pre-set normal behavior reconstruction model, this embodiment employs a deep autoencoder architecture based on unsupervised learning. This model consists of an encoder and a decoder. The encoder contains a multi-layered (e.g., three to five layers) fully connected neural network or convolutional neural network with progressively lower dimensionality, while the decoder contains a progressively higher dimensionality network symmetrical to the number of encoder layers. Its physical meaning lies in learning the characteristic manifold distribution under normal operating conditions of the distribution network. During the online simulation phase, this embodiment uses the temporal evolution feature representation output in S14 as the model's input vector. The encoder first maps it to a low-dimensional latent space to extract core features, and the decoder then performs inverse reconstruction based on this latent space vector, outputting a reconstructed vector with the same dimension as the input vector. This reconstructed vector is defined as the normal operation baseline vector, representing the theoretical characteristic form that a node should exhibit if it is in a fully healthy state under the constraints of the input features at the current moment.

[0046] It is important to note that this model is trained offline using a clean training set of historical data from the distribution network historical operation database defined and constructed in S14. The data is rigorously selected from periods marked as operating normally and verified to have no abnormal events. The training objective is to minimize the mean squared error between the input vector and the reconstructed vector. Therefore, when the input contains abnormal features indicating fault symptoms, since the model has never learned the encoding rules for abnormal patterns, its reconstructed output will inevitably differ significantly from the actual input. This difference serves as the quantitative basis for fault identification.

[0047] In one implementation, this embodiment employs a data-driven method based on extreme value theory or quantile statistics to determine the preset risk judgment threshold, adapting to the long-tailed distribution characteristics typically exhibited by reconstruction errors. Specifically, this embodiment retains a portion of normal operating data that was not used in training as an independent validation set during the offline phase. The system inputs all samples from this validation set into a pre-trained normal behavior reconstruction model, calculates the reconstruction error for each sample, and thus obtains a baseline error distribution set. Subsequently, this embodiment determines the quantile position based on the system's maximum allowable false alarm rate (e.g., 0.1%), calculates the value of the corresponding percentile (e.g., 99.9% quantile) in the error distribution set, and strictly locks this value as the preset risk judgment threshold. This determination method ensures that the false alarm rate of the system under normal operating conditions is statistically controlled below the expected level, achieving an objective balance between sensitivity and robustness.

[0048] In one implementation, regarding risk assessment and node labeling, this embodiment calculates the Euclidean distance between the temporal evolution feature representation of the current node and the normal operation reference vector in real time, and defines it as the feature space distance. The system compares this distance with a previously determined preset risk assessment threshold. If the feature space distance is greater than the threshold, the system determines that the current operating state of the node has significantly deviated from the normal manifold and that there is a potential physical anomaly. This embodiment immediately labels the global index ID of the network node as a potential risk node, and generates a risk warning record by combining the current real-time timestamp and the specific value of the feature space distance, providing a location index for subsequent fault qualitative analysis.

[0049] In step S16, the temporal evolution feature representation of the potential risk node is matched with a preset historical fault pattern library to determine the potential fault type, including: The temporal evolution features of the potential risk nodes are represented as query vectors; Traverse the preset historical fault pattern library and calculate the similarity value between the query vector and each historical fault pattern vector in the preset historical fault pattern library; The historical fault pattern vector with the highest similarity value is selected as the target pattern, and the fault label associated with the target pattern is extracted. The fault label is then identified as the potential fault type.

[0050] In one implementation, to address the technical challenge of simple feature matching failing due to dynamic reconfiguration of the distribution network topology in constructing the pre-set historical fault mode library, this embodiment adopts a topology-aware storage strategy.

[0051] Specifically, during the offline database construction phase, the system extracts historical typical fault waveform data diagnosed manually from the distribution network's historical operation database, and combines this with fault sample data generated from simulation to manually label each set of data to determine its fault label. While extracting the temporal evolution feature representation corresponding to each type of fault, the system simultaneously extracts the local topology subgraph within a preset number of hops (e.g., 2 hops) for the node at the time of the fault occurrence. This embodiment uses a graph isomorphic hash algorithm (e.g., Weisfeiler-Lehman hash) to map this local topology subgraph into a unique topology fingerprint. Finally, the system uses the topology fingerprint as the primary key and the temporal evolution feature representation and fault label as association values, storing them in the database as key-value pairs. This construction method ensures that subsequent feature matching is performed only within a geometrically similar physical connection structure, eliminating the feature space misalignment problem caused by changes in the power grid's operation mode, i.e., topological heterogeneity.

[0052] In one implementation, regarding the calculation of similarity scores and the determination of potential fault types, this embodiment first calculates the real-time topological fingerprint of the current potential risk node and filters out a set of candidate patterns that match the fingerprint from the database. Subsequently, this embodiment uses cosine similarity as a metric, taking the temporal evolution feature representation as a query vector and comparing it with each historical fault pattern vector in the candidate set. Specifically, the system calculates the dot product of the query vector and the historical fault pattern vector, and simultaneously calculates the Euclidean norm (i.e., magnitude) of these two vectors. Finally, the dot product is divided by the product of the magnitudes of the two vectors to obtain a similarity score between -1 and +1. The closer this score is to +1, the more consistent the feature evolution directions of the two are. This embodiment selects the pattern with the highest similarity score as the initial target and introduces a statistical confidence threshold determination mechanism.

[0053] It is important to note that the confidence threshold, or matching threshold, is determined based on the posterior probability distribution of the validation set. During the model training phase, this embodiment statistically analyzes the similarity values ​​between all correctly identified fault samples in the validation set and their corresponding cluster centers, constructing a similarity probability density function. This embodiment selects the left 5% quantile value of this distribution as the matching threshold. During online matching, if the current highest similarity value is still lower than this threshold, the system determines that the current fault feature is in the blind spot of historical experience (i.e., a novel or complex fault that has not been seen before). Instead of forcibly outputting a specific label, it marks it as "unknown anomaly" and triggers a manual review process, thereby effectively avoiding the risk of forced misclassification under open set conditions.

[0054] In step S17, based on the location of the potential risk node in the topology and the potential fault type, the fault evolution trend is deduced and a handling suggestion is generated, including: The location information is obtained by parsing the physical number of the potential risk node, and the corresponding real-time operating status data stream is extracted based on the location information; Based on the potential fault type, the location information, and the real-time operating status data stream, an initial state of fault evolution is constructed. The pre-set fault evolution probability model is invoked, and the propagation path and cascading impact range of the fault are iteratively deduced from the initial state of the fault evolution in the future multiple time steps. Based on the cascading impact range, a pre-set emergency response strategy library is matched to generate fault handling suggestions containing isolation strategies. These suggestions are then packaged with the potential fault types and the location information to generate a final prediction report.

[0055] In one implementation, to construct the initial state vector for fault evolution, this embodiment first maps the physical number of the potential risk node to the logical node index in the current topology graph by querying the real-time topology adjacency matrix generated in S11. Then, the system forcibly assigns values ​​to the electrical state variables of the logical node based on the potential fault type output in S16 (e.g., "A-phase metallic grounding" or "AB-phase short circuit"). For example, if the fault type is "A-phase grounding," the A-phase voltage state of the node is set to zero, and the zero-sequence voltage state is set to the system rated voltage. This forcibly assigned state vector, combined with the real-time power flow state of the entire network at the current moment (provided by the fusion features in S12), constitutes the zero-time boundary conditions used for subsequent deduction, i.e., the initial state vector for fault evolution.

[0056] In one implementation, iterative deduction is performed using a pre-defined fault evolution probability model. This embodiment employs a Dynamic Bayesian Network (DBN) as the core of the deduction and focuses on addressing the issues of the transparency of model parameter sources and computational efficiency. The network structure of the pre-defined fault evolution probability model is dynamically isomorphic to the real-time topology obtained from S11 analysis; that is, each physical node in the power grid corresponds to a random variable node in the Bayesian network, and the physical connections of the power grid correspond to dependency edges in the network.

[0057] It is important to note that the conditional probability table (CPT) for node state transitions in this network is strictly constructed based on the time-current characteristic curve (TCC) of the distribution network relay protection devices. The time-current characteristic curve is a digital curve automatically generated by the system based on the inverse-time action equations defined in the International Electrotechnical Commission (IEC) or Institute of Electrical and Electronics Engineers (IEEE) standards (e.g., the inverse-time formula in IEC 60255), combined with pre-entered setting values ​​(including parameters such as starting current and time multiples) of each level of protection device; or it can directly call the measured action characteristic data spectrum provided by the equipment manufacturer.

[0058] Specifically, the system pre-stores the protection setting values ​​for all levels of circuit breakers and fuses across the entire network. During the simulation calculation, the model calculates the fault current value flowing through a certain branch and looks up the protection operation probability based on the corresponding operating time on the TCC curve. For example, if the fault current exceeds the instantaneous overcurrent protection setting value, the probability of the circuit breaker in that branch "tripping" in the next time step is set to be close to 1 (e.g., 0.99); otherwise, the probability is close to 0. This modeling method based on physical protection logic ensures that the probabilistic simulation process conforms to the actual physical response law of the power system.

[0059] In one implementation, to ensure complex probabilistic deductions are completed within millisecond-level time windows, this embodiment employs a probability truncation and parallel computing strategy. When the system uses the Monte Carlo method to sample and infer the DBN, a probability truncation threshold (e.g., 0.1%) is set. During the search of the evolutionary tree, for any branch path with a cumulative probability lower than this threshold, the system treats it as an extremely low-probability event and prunes it directly, ceasing further deduction. Simultaneously, this embodiment utilizes the parallel computing units of the GPU to simultaneously launch thousands of simulation threads, computing evolutionary paths under different random seeds in parallel. Finally, the system counts the set of nodes affected (i.e., experiencing power outages or voltage overruns) in all unpruned evolutionary paths, takes their union as the cascading impact range, and selects the propagation sequence with the highest probability as the propagation path.

[0060] In one implementation, to generate disposal suggestions, this embodiment first constructs the pre-set emergency disposal strategy library using a data-driven method combining historical data association mining and topology simulation verification in the offline stage. Specifically, the system connects to the distribution network dispatch management system through a data interface, batch exports historical fault disposal logs and switch operation sequence data from the past several years across the entire network, and automatically filters out invalid samples marked as failures or secondary reclosing based on the disposal result status bits in the logs, retaining only positive sample operation sequences that successfully isolated the fault and restored power supply. Next, the system uses frequent pattern mining algorithms, such as the FP-Growth algorithm or the Apriori algorithm, to perform association analysis between the fault location and type and the switch operation sequence, calculates the support and confidence of each operation sequence, and automatically extracts high-frequency operation sequences with confidence levels higher than a preset statistical threshold, such as 95%, as candidate historical plans.

[0061] Furthermore, to cover blank topology nodes not covered by historical data, the system further executes a graph-based path search algorithm. Centered on each node in the distribution network topology graph, the system uses a breadth-first search algorithm to traverse outwards to find the nearest operable switchgear to determine the theoretical isolation boundary. Simultaneously, based on the minimum spanning tree algorithm or the maximum flow minimum cut principle, it calculates the optimal transfer path from the non-faulty power outage area to the backup power source, thereby generating a computational completion plan. Subsequently, the system inputs the mined candidate historical plans and the generated computational completion plan one by one into the built-in distribution network power flow simulation engine, simulating the execution of the switching actions in the plans and performing power flow calculations to verify whether the line load rate and node voltage deviation meet the safe operation constraints.

[0062] It should be further explained that the built-in power flow simulation engine is a numerical calculation module developed based on power system analysis principles. This module integrates a steady-state calculation core conforming to industry standards such as IEEE or IEC standards in its software architecture. Its power flow calculation operation is implemented based on the Newton-Raphson iterative method. Specifically, this method calculates the voltage magnitude and phase angle of each network node, as well as the power transmission status of each line, by solving the nonlinear power balance equations of the distribution network nodes. This ensures the verification of line load rate and node voltage deviation after the implementation of the contingency plan, possessing high accuracy and physical reliability.

[0063] Finally, the system structures all the contingency plans that have passed the simulation verification into key-value pairs and stores them in the database to complete the database construction. The database stores standard handling contingency plans for different fault locations and impact ranges, such as "cut off F1 feeder switch" or "close L2 tie switch".

[0064] During the online decision-making phase, the system transforms the derived cascading impact range, i.e., the damaged topology subgraph, into a structured query statement and performs a matching search within the pre-set emergency response strategy library. If a perfectly matching plan exists, it is output directly; otherwise, the system uses a graph edit distance algorithm to find the plan corresponding to the most similar scenario. Finally, the system encapsulates the potential fault types, the derived propagation path graph, and the generated response suggestions into a standardized XML or JSON format report, which is then pushed to the operations and maintenance main station via the scheduling data network.

[0065] In summary, this invention deeply couples the dynamic analysis of the real-time topology of the distribution network with the spatial aggregation mechanism of graph neural networks, and utilizes a multi-dimensional attention mechanism to adaptively fuse multi-source heterogeneous data from smart meters and micro-synchronous phasor measurement units. Simultaneously, it introduces dynamic risk assessment based on normal behavior reconstruction and fault mode matching strategies based on topological fingerprints. This enables accurate capture of potential fault symptoms and physically interpretable deduction of evolution trends under complex operating conditions characterized by frequent reconfiguration of distribution network operation modes and significant differences in the spatiotemporal scales of multi-source data. It effectively solves the problems of false alarms and missed alarms caused by fixed topology assumptions and rigid threshold settings in existing technologies, significantly improving the real-time performance and reliability of proactive operation and maintenance of the distribution network.

[0066] Reference Figure 2 The second embodiment of the present invention provides a fusion distribution network fault prediction system based on multi-source data, comprising: The data acquisition and parsing module is used to acquire real-time operation monitoring data of the distribution network, and perform topology parsing and data extraction to obtain a real-time topology adjacency matrix and multi-source measurement data including smart meter data and micro phasor measurement data. The multi-scale feature fusion module is used to perform multi-scale time feature fusion on the multi-source measurement data to obtain the fused heterogeneous feature vector of each network node in the distribution network. The spatial dimension aggregation module is used to perform spatial dimension aggregation on the fused heterogeneous feature vector based on the real-time topological adjacency matrix and using a pre-set graph neural network model to obtain the spatial node embedding of the network node. The time dimension analysis module is used to perform time dimension analysis on the spatial node embedding using a pre-set long short-term memory network model, and generate a temporal evolution feature representation of the network node. The risk assessment module is used to obtain the normal operation reference vector corresponding to the current working condition, calculate the feature space distance between the time-series evolution feature representation and the normal operation reference vector, and identify network nodes whose feature space distance exceeds a preset risk assessment threshold as potential risk nodes. The fault characterization module is used to match the temporal evolution feature representation of the potential risk node with a pre-set historical fault pattern library to determine the potential fault type; The trend prediction and decision-making module is used to predict the evolution trend of the fault and generate disposal suggestions based on the location of the potential risk node in the topology and the potential fault type.

[0067] It should be noted that the multi-source data-based integrated distribution network fault prediction system provided in this embodiment of the invention is used to execute all the process steps of the multi-source data-based integrated distribution network fault prediction method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0068] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a multi-source data-based integrated distribution network fault prediction program. When the processor executes the computer program, it implements the steps described in the various embodiments of the multi-source data-based integrated distribution network fault prediction method, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above system embodiments, such as the data acquisition and parsing module.

[0069] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0070] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0071] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0072] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0073] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0074] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0075] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for predicting faults in a distribution network based on multi-source data fusion, characterized in that, include: Real-time operation monitoring data of the power distribution network is acquired, and topology analysis and data extraction are performed to obtain a real-time topology adjacency matrix and multi-source measurement data including smart meter data and micro phasor measurement data. Multi-scale time feature fusion is performed on the multi-source measurement data to obtain the fused heterogeneous feature vector of each network node in the distribution network; Based on the real-time topological adjacency matrix, a pre-set graph neural network model is used to aggregate the spatial dimensions of the fused heterogeneous feature vectors to obtain the spatial node embedding of the network nodes. A temporal dimension analysis of the spatial node embedding is performed using a pre-built long short-term memory network model to generate a temporal evolution feature representation of the network nodes; Obtain the normal operation reference vector corresponding to the current working condition, and calculate the feature space distance between the time-series evolution feature representation and the normal operation reference vector. Network nodes whose feature space distance exceeds a preset risk judgment threshold are identified as potential risk nodes. The temporal evolution feature representation of the potential risk node is matched with a pre-set historical fault pattern library to determine the potential fault type; Based on the location of the potential risk nodes in the topology and the potential fault types, the fault evolution trend is deduced and disposal suggestions are generated.

2. The method for predicting distribution network faults based on multi-source data according to claim 1, characterized in that, The process of acquiring real-time operation monitoring data of the distribution network, performing topology analysis and data extraction to obtain a real-time topology adjacency matrix and multi-source measurement data including smart meter data and micro phasor measurement data includes: The status of line switches, equipment operating parameters, and measurement device data of the power distribution network are collected as the real-time operation monitoring data; Separate and extract line switching status data and equipment operation status data from the real-time operation monitoring data, and identify the effective network nodes in the distribution network based on the line switching status data and the equipment operation status data; The on / off status of each connected line is determined based on the line switching status data, and the conductive connection path between the effective network nodes is determined. Construct an initial two-dimensional array, and assign connectivity values ​​at the row and column intersection positions corresponding to the effective network nodes and the connected connection paths in the initial two-dimensional array to obtain the real-time topology adjacency matrix; The smart meter data and the micro phasor measurement data are filtered and extracted from the real-time operation monitoring data to serve as the multi-source measurement data.

3. The method for predicting distribution network faults based on multi-source data according to claim 1, characterized in that, The process of fusing multi-scale time features of the multi-source measurement data to obtain a fused heterogeneous feature vector for each network node in the distribution network includes: The high-frequency sampling time of the micro phasor measurement data is used as the reference time axis; Using a sample-and-hold algorithm or a trend extrapolation algorithm, the low-frequency sampled smart meter data is aligned to the reference time axis to construct a long-period steady-state feature sequence aligned with the timestamp of the micro phasor measurement data; The micro phasor measurement data is used as a short-period dynamic feature sequence and vector-concatenated with the long-period steady-state feature sequence to construct a heterogeneous data matrix. The heterogeneous data matrix is ​​input into a pre-set multi-head attention computing model, and the weight distribution of features at different time scales is calculated using the attention mechanism. Adaptive weighted fusion is then performed to obtain the fused heterogeneous feature vector.

4. The method for predicting distribution network faults based on multi-source data according to claim 1, characterized in that, The step of aggregating the spatial dimensions of the fused heterogeneous feature vectors based on the real-time topological adjacency matrix using a pre-set graph neural network model to obtain the spatial node embedding of the network nodes includes: The fused heterogeneous feature vector is mapped to the initial attribute features of the network node; The topological neighborhood range of the network node is determined based on the real-time topological adjacency matrix, and the attribute characteristics of adjacent nodes within the topological neighborhood range are obtained by traversal. The pre-set graph neural network model is used to perform weighted aggregation calculations on the attribute features of the adjacent nodes to generate aggregated features containing local spatial dependency information. The aggregation features are subjected to a nonlinear activation transformation to update the state representation of the network nodes, thereby obtaining the spatial node embedding.

5. The method for predicting distribution network faults based on multi-source data according to claim 1, characterized in that, The step of performing a time-dimensional analysis on the spatial node embedding using a pre-built long short-term memory network model to generate a temporal evolution feature representation of the network nodes includes: For each of the network nodes, the spatial node embeddings of the network node at multiple consecutive time points are extracted using the sliding window algorithm to construct a temporal feature input tensor; The temporal features are input into the tensor and then into the pre-set long short-term memory network model. The gating mechanism is used to selectively remember and forget the historical state sequence, and the fluctuation trend features in the time dimension are extracted. The temporal evolution feature representation of the network node is obtained by performing a fully connected mapping on the fluctuation trend feature.

6. The method for predicting distribution network faults based on multi-source data according to claim 1, characterized in that, The step of obtaining the normal operation reference vector corresponding to the current operating condition, calculating the feature space distance between the temporal evolution feature representation and the normal operation reference vector, and identifying network nodes whose feature space distance exceeds a preset risk judgment threshold as potential risk nodes includes: Obtain the operating condition parameters at the current moment, and generate the normal operation reference vector that corresponds to the operating condition parameters and is in the same feature space based on the preset normal behavior reconstruction model; For each of the network nodes, calculate the multidimensional Euclidean distance or cosine similarity between the temporal evolution feature representation of the network node and the normal operation reference vector; The calculated distance value or similarity value is determined as the feature space distance, and it is determined whether the feature space distance exceeds the preset risk judgment threshold. If the feature space distance exceeds the preset risk judgment threshold, the corresponding network node is marked as the potential risk node, and the physical number of the network node is recorded as the potential fault location.

7. The method for predicting distribution network faults based on multi-source data according to claim 1, characterized in that, The step of matching the temporal evolution feature representation of the potential risk node with a pre-set historical fault pattern library to determine the potential fault type includes: The temporal evolution features of the potential risk nodes are represented as query vectors; Traverse the preset historical fault pattern library and calculate the similarity value between the query vector and each historical fault pattern vector in the preset historical fault pattern library; The historical fault pattern vector with the highest similarity value is selected as the target pattern, and the fault label associated with the target pattern is extracted. The fault label is then identified as the potential fault type.

8. The method for predicting distribution network faults based on multi-source data according to claim 1, characterized in that, The step of inferring the failure evolution trend and generating handling suggestions based on the location of the potential risk nodes in the topology and the potential failure types includes: The location information is obtained by parsing the physical number of the potential risk node, and the corresponding real-time operating status data stream is extracted based on the location information; Based on the potential fault type, the location information, and the real-time operating status data stream, an initial state of fault evolution is constructed. The pre-set fault evolution probability model is invoked, and the propagation path and cascading impact range of the fault are iteratively deduced from the initial state of the fault evolution in the future multiple time steps. Based on the cascading impact range, a pre-set emergency response strategy library is matched to generate fault handling suggestions containing isolation strategies. These suggestions are then packaged with the potential fault types and the location information to generate a final prediction report.

9. A fusion-based distribution network fault prediction system based on multi-source data, characterized in that, include: The data acquisition and parsing module is used to acquire real-time operation monitoring data of the distribution network, and perform topology parsing and data extraction to obtain a real-time topology adjacency matrix and multi-source measurement data including smart meter data and micro phasor measurement data. The multi-scale feature fusion module is used to perform multi-scale time feature fusion on the multi-source measurement data to obtain the fused heterogeneous feature vector of each network node in the distribution network. The spatial dimension aggregation module is used to perform spatial dimension aggregation on the fused heterogeneous feature vector based on the real-time topological adjacency matrix and using a pre-set graph neural network model to obtain the spatial node embedding of the network node. The time dimension analysis module is used to perform time dimension analysis on the spatial node embedding using a pre-set long short-term memory network model, and generate a temporal evolution feature representation of the network node. The risk assessment module is used to obtain the normal operation reference vector corresponding to the current working condition, calculate the feature space distance between the time-series evolution feature representation and the normal operation reference vector, and identify network nodes whose feature space distance exceeds a preset risk assessment threshold as potential risk nodes. The fault characterization module is used to match the temporal evolution feature representation of the potential risk node with a pre-set historical fault pattern library to determine the potential fault type; The trend prediction and decision-making module is used to predict the evolution trend of the fault and generate disposal suggestions based on the location of the potential risk node in the topology and the potential fault type.

Images