Power distribution network state prediction and fault positioning method, device and terminal equipment

Abstract

The application provides a power distribution network state prediction and fault positioning method, device and terminal equipment, relates to the power distribution network fault detection technical field, and the method comprises the following steps: receiving state estimation data of at least one edge node; acquiring the topological structure of the power distribution network and the historical fault information of the power distribution network; taking the topological structure of the power distribution network, the historical fault information of the power distribution network and the state estimation data of each edge node as inputs, obtaining the fault section positioning information of the power distribution network through a global space-time graph neural network model constructed in advance based on the topological structure of the power distribution network; acquiring the switch state of the power distribution equipment of the fault section; taking the fault section positioning information and the switch state of the power distribution equipment as inputs, and determining the fault point positioning information of the power distribution network through a fault positioning identification model constructed based on the switch state of the power distribution equipment. The application effectively improves the accuracy of power distribution network state estimation and fault positioning, and improves the overall performance.

Description

Technical Field

[0001] This application relates to the field of power distribution network fault detection technology, specifically to a power distribution network status prediction and fault location method, a power distribution network status prediction and fault location device, a machine-readable storage medium, and a terminal device. Background Technology

[0002] As a crucial link directly connecting the power system to users, the safe and reliable operation of the distribution network is vital to the overall power supply quality. Traditional distribution network state estimation methods are mainly based on physical models and numerical calculations, such as least squares state estimation and weighted least squares. These methods rely on accurate mathematical models of the distribution network and a large amount of real-time measurement data to estimate the system state by solving a system of nonlinear equations. However, distribution networks typically have a radial structure, exhibit three-phase imbalance characteristics, and suffer from insufficient measurement equipment, limiting the effectiveness of traditional methods in distribution networks.

[0003] Currently, in fault location technology, traditional methods can be mainly divided into three categories: methods based on fault current analysis, methods based on traveling wave signals, and methods based on impedance calculation. Methods based on fault current analysis locate fault sections by monitoring changes in fault current parameters when a fault occurs in the distribution network, combined with the distribution network topology. These methods are simple in principle, but are easily affected by fault impedance and system operating modes, and are insensitive to high-impedance faults. Methods based on traveling wave signals utilize the propagation characteristics of transient traveling wave signals generated when a fault occurs in the distribution network for location. Although the location accuracy is high, it requires high sampling frequency and signal processing capabilities, and the traveling wave signal is prone to attenuation and distortion in the complex branch structure of the distribution network. Methods based on impedance calculation calculate the fault distance by measuring the voltage and current at the time of the fault, but their accuracy is greatly affected by the accuracy of line parameters and system operating modes.

[0004] With the development of artificial intelligence (AI) technology, its application in power distribution network state estimation and fault location is increasing. Deep learning-based methods learn fault characteristics from historical data through neural network models to achieve fault classification and location. For example, some studies use convolutional neural networks (CNNs) to extract features from image data of power distribution equipment to identify equipment damage and aging conditions; recurrent neural networks (RNNs) are used to process time series data to predict equipment fault trends and patterns. However, despite the progress made by AI technology in power distribution network fault location, existing methods still have many limitations, such as high dependence on data quality and integrity, poor model interpretability, and insufficient adaptability to complex fault scenarios. In addition, most existing AI methods focus on post-fault location analysis, with insufficient research on the collaborative optimization of state estimation and fault location. Summary of the Invention

[0005] The purpose of this application is to provide a method for predicting and locating the state of a power distribution network, a device for predicting and locating the state of a power distribution network, a machine-readable storage medium, and a terminal device to solve the above-mentioned problems.

[0006] To achieve the above objectives, the first aspect of this application provides a method for predicting the state of a distribution network and locating faults, comprising:

[0007] Receive state estimation data of at least one edge node, the state estimation data being calculated by each edge node from the multi-source operating data obtained by the corresponding edge node through an edge node spatiotemporal graph neural network model pre-constructed based on the topology of the corresponding edge node;

[0008] The topology and historical fault information of the distribution network are obtained. The fault location information of the distribution network is obtained by using the topology, historical fault information and state estimation data of each edge node as input and a global spatiotemporal graph neural network model pre-constructed based on the topology of the distribution network.

[0009] The switch status of the power distribution equipment in the fault section is obtained. The fault location information and the switch status of the power distribution equipment are used as inputs. The fault location information of the power distribution network is determined by the fault location and identification model constructed based on the switch status of the power distribution equipment.

[0010] Optionally, the training process of the spatiotemporal graph neural network model for edge nodes includes:

[0011] S1. Initialize the global model parameters of the global spatiotemporal graph neural network model;

[0012] S2. Send the global model parameters to each edge node so that each edge node uses the global model parameters as the initial model parameters of the spatiotemporal graph neural network model of the corresponding edge node;

[0013] S3. Receive the local model parameters of the corresponding edge node spatiotemporal graph neural network model sent by each edge node. The local model parameters are obtained by optimizing the corresponding edge node spatiotemporal graph neural network model with the local training data of each edge node.

[0014] S4. Obtain the number of samples of local training data for each edge node, determine the weight of each edge node based on the number of samples of local training data for each edge node, and perform a weighted average of the local model parameters of the corresponding edge node spatiotemporal graph neural network model using the weight of each edge node, and update the global model parameters with the weighted average result. If the preset maximum number of iterations is reached, execute step S5; otherwise, execute step S2.

[0015] S5. Use the updated global model parameters as the target global model parameters, use the target global model parameters as the final model parameters of the global spatiotemporal graph neural network model, and send the target global model parameters to each edge node so that each edge node uses the target global model parameters as the final model parameters of the corresponding edge node spatiotemporal graph neural network model.

[0016] Optionally, the weight of each edge node is determined based on the number of samples in the local training data of each edge node, including:

[0017] Determine the total number of samples in the local training data for all edge nodes;

[0018] For each edge node, the weight of the current edge node is the ratio of the number of samples in the local training data of the current edge node to the total number of samples.

[0019] Optionally, the multi-source operating data includes at least one of the voltage, current, power, frequency, ambient temperature, load forecast data, and power generation plan data of the corresponding edge node; the state estimation data includes the voltage amplitude and phase angle of the corresponding edge node; the training process of the spatiotemporal graph neural network model of each edge node includes:

[0020] Obtain the historical multi-source operating data of the current edge node and its corresponding actual voltage amplitude and actual phase angle value;

[0021] Based on the historical multi-source operating data of the current edge node and its corresponding actual voltage amplitude and actual phase angle, with the goal of minimizing the prediction error, the model parameters of the edge node spatiotemporal graph neural network model corresponding to the current edge node are optimized. The loss function of the edge node spatiotemporal graph neural network model is constructed by the measurement data fitting loss term and the physical constraint loss term.

[0022] The measurement data fitting loss term is used to represent the error between the predicted voltage amplitude and predicted phase angle values ​​output by the spatiotemporal graph neural network model of the edge nodes and the corresponding actual voltage amplitude and actual phase angle values;

[0023] The physical constraint loss term is used to represent the error between the predicted value of the target operating parameter of the current edge node and the theoretical value of the target operating parameter obtained based on the specified circuit law, under the current predicted voltage amplitude and predicted phase angle. The value of the physical constraint loss term is directly proportional to the error between the predicted value of the target operating parameter and the theoretical value of the target operating parameter.

[0024] Optionally, the specified circuit law is Kirchhoff's current law; the target operating parameter is the outflow current of the current edge node; the calculation steps for the error between the predicted value of the target operating parameter of the current edge node and the theoretical value of the target operating parameter obtained based on the specified circuit law include:

[0025] Obtain the actual value of the inflow current of the current edge node, and determine the theoretical value of the outflow current of the current edge node based on the actual value of the inflow current and Kirchhoff's current law.

[0026] The predicted outflow current of the current edge node is determined based on the current predicted voltage magnitude and predicted phase angle of the current edge node.

[0027] The absolute difference between the predicted outflow current and the theoretical outflow current is taken as the error between the predicted outflow current and the theoretical outflow current.

[0028] Optionally, the loss function of the spatiotemporal graph neural network model of the edge nodes is constructed from a measurement data fitting loss term and a physical constraint loss term, including:

[0029] Determine the weights of the measurement data fitting loss term and the physical constraint loss term;

[0030] The loss function of the spatiotemporal graph neural network model of the edge nodes is constructed by weighting and summing the measurement data fitting loss term and the physical constraint loss term according to their respective weights.

[0031] Optionally, the fault location and identification model is constructed through the following steps:

[0032] The topology of each power distribution device in the fault section is determined based on the topology of the power distribution network.

[0033] Based on the topology of each power distribution device within the fault section, the fault section is divided into multiple fault sub-sections, and the weight of each fault sub-section is determined. Each fault sub-section includes one power distribution device.

[0034] A linear penalty term is constructed based on the fault state and weight of each fault sub-section, and an error penalty term is constructed based on the measured overcurrent information and theoretical overcurrent information of the upstream fault sub-section of each fault sub-section under the corresponding fault state and the corresponding power distribution equipment switching state.

[0035] Construct the objective function based on the sum of the linear penalty term and the error penalty term;

[0036] The fault location and identification model is constructed based on the objective function, with the goal of minimizing the objective function.

[0037] Optionally, a linear penalty term is constructed based on the fault state of each fault sub-segment and the weight of each fault sub-segment, including:

[0038] The linear penalty term is constructed by weighting and summing the fault states of each fault sub-segment based on their respective weights.

[0039] An error penalty term is constructed based on the measured and theoretical overcurrent information of the upstream fault sub-sections of each fault sub-section under the corresponding fault state and the corresponding power distribution equipment switching state, including:

[0040] Construct a switching function that represents the theoretical overcurrent information of the upstream fault sub-segment under the corresponding fault state and the switching state of its corresponding power distribution equipment;

[0041] The error penalty term is constructed by summing the absolute differences between the measured overcurrent information of all upstream fault sub-segments and the theoretical overcurrent information obtained from the switching function for any fault sub-segment under the corresponding fault state.

[0042] A second aspect of this application provides a power distribution network condition prediction and fault location device, comprising:

[0043] The data acquisition module is configured to receive state estimation data of at least one edge node. The state estimation data is obtained by each edge node through a spatiotemporal graph neural network model of the edge node pre-constructed based on the topology of the corresponding edge node, after calculating the multi-source operating data obtained by the corresponding edge node.

[0044] The first fault location module is configured to acquire the topology of the distribution network and the historical fault information of the distribution network. Taking the topology of the distribution network, the historical fault information of the distribution network and the state estimation data of each edge node as input, the fault location information of the distribution network is obtained through a global spatiotemporal graph neural network model pre-constructed based on the topology of the distribution network.

[0045] The second fault location module is configured to acquire the switch status of the power distribution equipment in the fault section, and use the fault section location information and the switch status of the power distribution equipment as input to determine the fault location information of the power distribution network through a fault location identification model constructed based on the switch status of the power distribution equipment.

[0046] In a third aspect, this application provides a machine-readable storage medium storing instructions that, when executed by a processor, configure the processor to perform the power distribution network state prediction and fault location method described above.

[0047] In a fourth aspect, this application provides a terminal device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the power distribution network status prediction and fault location method described above.

[0048] This application is based on constructing a spatiotemporal graph neural network for each node and the entire distribution network. By simultaneously capturing the spatial correlation and temporal dynamic changes between nodes, the accuracy of distribution network state estimation and fault location is effectively improved. At the same time, the spatiotemporal graph neural network model deployed on the edge nodes of the nodes performs state estimation for each node, and the global spatiotemporal graph neural network model deployed on the server performs preliminary location of fault sections of the distribution network. Based on the switching status of the distribution equipment in the fault section, the specific fault point in the fault section is accurately located. By co-optimizing state estimation and fault location, this application overcomes the limitation of treating the two problems in isolation in traditional methods and improves the overall performance.

[0049] Other features and advantages of the embodiments of this application will be described in detail in the following detailed description section. Attached Figure Description

[0050] The accompanying drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the following detailed description to explain the embodiments of this application, but do not constitute a limitation on the embodiments of this application. In the drawings:

[0051] Figure 1 A flowchart of a method for predicting the state of a power distribution network and locating faults, provided in a preferred embodiment of this application;

[0052] Figure 2 A schematic diagram of the system architecture provided for a preferred embodiment of this application;

[0053] Figure 3 A schematic diagram of a spatiotemporal graph neural network architecture provided in a preferred embodiment of this application;

[0054] Figure 4 A schematic block diagram of a power distribution network status prediction and fault location device provided in a preferred embodiment of this application;

[0055] Figure 5 A schematic diagram of a terminal device provided for a preferred embodiment of this application. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0057] It should be noted that the technical solutions of the various embodiments of this application can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by this application.

[0058] To solve the above problems, such as Figure 1 As shown, the first aspect of this application provides a method for predicting the state of a distribution network and locating faults, including:

[0059] S100. Receive state estimation data of at least one edge node. The state estimation data is obtained by each edge node through an edge node spatiotemporal graph neural network model pre-constructed based on the topology of the corresponding edge node, after calculating the multi-source operation data obtained by the corresponding edge node.

[0060] S200. Obtain the topology of the distribution network and its historical fault information. Using the topology of the distribution network, its historical fault information, and the state estimation data of each edge node as input, obtain the fault location information of the distribution network through a global spatiotemporal graph neural network model pre-built based on the topology of the distribution network.

[0061] S300: Obtain the switch status of the power distribution equipment in the fault section. Using the fault section location information and the switch status of the power distribution equipment as input, determine the fault location information of the power distribution network through the fault location identification model constructed based on the switch status of the power distribution equipment.

[0062] Thus, this application constructs a spatiotemporal graph neural network for each node and the entire distribution network. By simultaneously capturing the spatial correlation and temporal dynamic changes between nodes, the accuracy of distribution network state estimation and fault location is effectively improved. At the same time, the spatiotemporal graph neural network model deployed at the edge nodes of the nodes performs state estimation for each node, and the global spatiotemporal graph neural network model deployed on the server performs preliminary location of fault sections of the distribution network. Based on the switching status of the distribution equipment in the fault section, the specific fault point in the fault section is accurately located. By co-optimizing state estimation and fault location, this application overcomes the limitation of treating the two problems in isolation in traditional methods and improves the overall performance.

[0063] Understandable, such as Figure 2 As shown, the system architecture of this application comprises five core parts: a data acquisition and preprocessing module, a federated learning coordination module, a spatiotemporal graph neural network model, a state estimation and fault location module, and a decision-making and execution module. The system first collects multi-source data from the distribution network, such as current, voltage, power factor, and temperature, through intelligent sensors and monitoring devices at each node. Then, it uses a federated learning framework to train local models at each edge node and periodically uploads the model parameters to the aggregation server for global model updates. Finally, it analyzes the changes in the distribution network state through a spatiotemporal graph neural network to achieve accurate state estimation and fault location. It is understood that the method of this application is applied to the aggregation server.

[0064] In this application, the data acquisition and preprocessing module is responsible for real-time acquisition of multi-source heterogeneous data from various sensors in the power distribution network (such as PMUs, smart meters, and fault indicators) and external systems (such as meteorological systems). Its core tasks include data acquisition: automatically acquiring electrical measurement data such as voltage, current, power, and frequency, as well as non-electrical data such as equipment status and ambient temperature; and data preprocessing: cleaning, denoising, and formatting the raw data, and filling in missing values ​​to ensure the quality of the data input into subsequent models, laying a solid foundation for accurate analysis.

[0065] The Federated Learning Coordination Module coordinates edge nodes distributed across various areas of the power distribution network (such as substations and distribution rooms) to perform collaborative machine learning. This allows for global model optimization without data leaving the network domain, effectively protecting data privacy. Its main functions include: Training Scheduling: Dynamically allocating training tasks and setting reasonable training times based on the computing power and data volume of each edge node, preventing slow nodes from dragging down overall efficiency and improving training speed; Model Aggregation: Receiving local model parameters uploaded by each edge node, using aggregation algorithms (such as FedAvg) to generate a better global model, and then distributing it to each node; Contribution Evaluation: Evaluating the contribution of each participating node to the global model through a built-in incentive mechanism and evaluation model, incentivizing high-quality data nodes to actively participate, and ensuring the healthy development of the federated learning ecosystem.

[0066] The spatiotemporal graph neural network model is the core algorithm engine for achieving high-precision state estimation and fault location. It abstracts the distribution network as a graph structure, where nodes represent buses, loads, etc., and edges represent distribution lines. The main functions of the spatiotemporal graph neural network model include: spatial feature capture: using graph convolutional networks (GCNs) to analyze the connectivity and electrical coupling characteristics between nodes, understanding the propagation impact of faults on the topology; temporal dynamic modeling: using temporal convolutional networks (TCNs) or recurrent neural networks (RNNs) to analyze the time-series patterns of measurement data from each node, capturing the dynamic changes of the system; and spatiotemporal joint analysis: by alternately stacking spatial and temporal convolutional layers, it achieves joint modeling of the complex spatiotemporal correlation characteristics of the distribution network, thereby enabling more accurate deduction of the complete system state and capture of subtle fault symptoms.

[0067] The state estimation and fault location module is mainly used for state estimation: using a trained spatiotemporal graph neural network model, based on limited real-time measurement data and pseudo-measurement data, it calculates the complete voltage amplitude and phase angle information of all nodes in the distribution network, forming a comprehensive and accurate perception of the system's operating state; fault location: adopts a two-stage strategy combining section location and precise ranging. First, it quickly locates the line section where the fault occurred, and then performs refined calculations within that section to give the precise location of the fault point and evaluate the reliability of the location results.

[0068] The decision-making and execution module is responsible for translating the analysis results into actual operation and maintenance actions. Its main functions include decision generation: based on the fault location results, type and severity, combined with the expert knowledge base or rule engine, it automatically generates the optimal handling strategy, such as isolating the faulty section, adjusting the operation mode, and starting the backup power supply; control execution: through standard interfaces, it safely and reliably sends control commands to the intelligent switches, circuit breakers and other related control equipment on site to achieve rapid fault isolation and power restoration to non-faulty areas.

[0069] In this application, a corresponding spatiotemporal graph neural network model is pre-deployed at each edge node, and a global spatiotemporal graph neural network model is deployed on the central aggregation server. This application optimizes the edge node spatiotemporal graph neural network and the global spatiotemporal graph neural network model using a federated learning framework to address the issues of data silos and privacy protection. The federated learning framework consists of a central aggregation server and multiple edge nodes. Each edge node trains its model using local data, uploading only the model parameters rather than the raw data, thereby effectively protecting data privacy and reducing network transmission pressure.

[0070] Understandably, the method of this application can be applied to aggregation servers. Step S100, the training process of the spatiotemporal graph neural network model for edge nodes, includes:

[0071] S1. Initialize the global model parameters of the global spatiotemporal graph neural network model. The aggregation server first generates a set of initial global model parameters, which can be the weights of each layer in the neural network.

[0072] S2. The global model parameters are sent to each edge node so that each edge node uses the global model parameters as the initial model parameters for its corresponding spatiotemporal graph neural network model. After generating the initial global model parameters, the aggregation server sends the generated global model parameters to each node, and each node uses the received global model parameters as the model parameters for the spatiotemporal graph neural network model deployed locally on its edge node.

[0073] S3. Receive the local model parameters of the corresponding edge node spatiotemporal graph neural network model sent by each edge node. The local model parameters are obtained by optimizing the corresponding edge node spatiotemporal graph neural network model using the local training data of each edge node. For example, the spatiotemporal graph neural network model of the edge node is trained using at least one of the following multi-source operating data: voltage, current, power, frequency, ambient temperature, load forecast data, and power generation plan data of the corresponding edge node, and state estimation data: voltage amplitude and phase angle of the corresponding edge node, as output. The initial model parameters are then optimized, and the optimized model parameters are used as the local model parameters of the edge node spatiotemporal graph neural network model. Each node sends the obtained local model parameters to the aggregation server.

[0074] S4. Obtain the number of samples of local training data for each edge node, determine the weight of each edge node based on the number of samples of local training data for each edge node, and perform a weighted average of the local model parameters of the corresponding spatiotemporal graph neural network model of each edge node using the weight of each edge node. Update the global model parameters with the weighted average result. If the preset maximum number of iterations is reached or the global optimization target is minimized, proceed to step S5; otherwise, proceed to step S2.

[0075] The weight of each edge node is determined based on the number of samples in the local training data of each edge node, including: determining the total number of samples in the local training data of all edge nodes; and for each edge node, the weight of the current edge node is the ratio of the number of samples in the local training data of the current edge node to the total number of samples.

[0076] Specifically, in this application, the global optimization objective of federated learning can be expressed as:

[0077]

[0078] Where F(ω) is the global loss function, which is the overall objective that the entire federated learning system aims to optimize. By minimizing F(ω), a global model that performs optimally across all participating data can be obtained; ω represents the global model parameters, i.e., the shared model parameters to be obtained; K is the total number of edge nodes. In a power distribution network scenario, each edge node can represent a smart substation, a distribution automation area, or an industrial park equipped with smart measurement devices; n k Let n be the number of local training data samples possessed by the k-th edge node, for example, the number of records of historical operating data such as voltage, current, and power stored in a substation over the past year; n is the total number of training data samples of all edge nodes, i.e., n = n1 + n2 + ... + n K ;(n k / n) is the weight coefficient, that is, the weight of the k-th edge node. In this application, the weight coefficient gives edge nodes with large data volume a greater weight in the model aggregation process. ω is the local loss function for the k-th edge node, which measures the performance of the current global model ω on the local data of that edge node (e.g., the magnitude of the prediction error). The smaller the value, the better the model performs locally at that edge node. The model can be represented as follows:

[0079]

[0080] Where L is the loss function, which is the innermost function used to calculate the difference or loss between the predicted value and the true value of a single sample. Common loss functions include mean squared error for regression tasks and cross-entropy loss for classification tasks. The model uses parameters ω to represent the input data. The predictions made, for example, This includes the voltage, current, power, and frequency of the edge nodes; This indicates the true label or value of this sample, such as the measured voltage amplitude and phase angle of an edge node, i.e. The calculation is the model's prediction loss for the i-th sample.

[0081] Understandable. The specific form can be determined based on the model task. For example, if the model task is state estimation (predicting node voltage). It could be mean square error, which measures the difference between the predicted voltage and the actual measured voltage; if the task is fault classification, it could be cross-entropy loss.

[0082] This application will use all n from the local training dataset of edge node k. k The losses for each training sample are summed. Then, to eliminate the influence of data size on the total loss, the sum is divided by the total number of training samples, n. k This allows us to obtain the average loss, thus making the losses between edge nodes with different amounts of data comparable. That is, The essence is the average prediction loss of the model parameters ω across the entire local training dataset of the edge node k.

[0083] S5. Use the updated global model parameters as the target global model parameters, use the target global model parameters as the final model parameters of the global spatiotemporal graph neural network model, and send the target global model parameters to each edge node so that each edge node uses the target global model parameters as the final model parameters of the spatiotemporal graph neural network model of the corresponding edge node.

[0084] After each training round, the aggregation server performs a weighted average of the local model parameters of each edge node based on the weights of each edge node, updates the global model parameters with the weighted average result, and then distributes the updated global model parameters to each edge node. Each edge node replaces its local model parameters with the received updated global model parameters and performs the next round of training until the convergence condition is met, i.e., the global loss function is minimized, or the maximum number of iterations is reached. The most recently updated global model parameters are used as the model parameters of each edge spatiotemporal graph neural network model and the global spatiotemporal graph neural network model.

[0085] This application, by treating federated learning as a large-scale distributed problem of minimizing empirical risk, has the following beneficial effects:

[0086] Minimizing global empirical risk: The objective function of this application guides the optimization direction of the model, aiming to achieve the optimal overall average performance across all edge node datasets, rather than just the optimal local performance of a single edge node; Balancing fairness and efficiency: Through (n kThe weight design of / n) allows the objective function to balance the interests of different data contributors, assigning higher weights to contributors with larger data volumes. This guides the model towards more robust convergence, as more data typically implies richer patterns. Adaptation to federated learning frameworks: The objective function structure of this application effectively aligns with the collaborative process of federated learning, allowing each party to optimize its own locally. The server then determines the weight (n) k Aggregate / n) to indirectly optimize the global objective F(ω).

[0087] In the federated learning scenario of power distribution networks, the objective function of this application can jointly optimize a global model without exchanging raw data, thereby overcoming the problem that data from various substations and power supply stations cannot be directly centralized due to privacy and security regulations. Since the load characteristics, distributed energy penetration rate, and fault types of power distribution networks in different regions may vary greatly (i.e., non-independent and identically distributed data), the weighted average objective function of this application helps to mitigate the negative impact of data distribution differences, thus obtaining a compromise model that can take into account the characteristics of different regions. Through local computation and periodic aggregation, the federated learning framework of this application reduces the need for transmitting a large amount of raw data over the network and reduces communication overhead.

[0088] To effectively capture the spatiotemporal characteristics of the distribution network, this application constructs a Spatial-Temporal Graph Neural Network (STGNN) model for edge node state estimation and distribution network fault location. The complete architecture of the STGNN in this application is as follows: Figure 3 As shown, the spatiotemporal graph neural network model simultaneously considers the spatial relationships and temporal dynamics between nodes, achieving joint modeling of the spatiotemporal characteristics of the power distribution network through alternating stacks of spatial and temporal convolutions. Specifically, the spatiotemporal graph neural network model includes:

[0089] (1) Graph structure construction: The distribution network is abstracted as a graph structure G=(V,E,A), where V is the set of nodes (buses, load points, etc.), E is the set of edges (distribution lines), and A is the adjacency matrix, representing the connection relationship between nodes. Edge features include electrical and physical parameters such as line resistance, reactance, and length. Node feature vectors include: electrical measurement features: voltage amplitude, phase angle, current, power, etc.; equipment status features: switch status, transformer tap position, etc.; environmental features: temperature, humidity, load rate, etc.; historical fault features: number of faults, fault type, processing records, etc.

[0090] (2) Spatial Convolution Module: The spatiotemporal graph neural network model uses a graph convolutional network to capture the spatial dependencies between nodes. The graph convolution operation is represented as:

[0091]

[0092] in, For the normalized adjacency matrix, Let the nodes of the l-th layer be represented as follows: For trainable weight matrix, This is the activation function.

[0093] Specifically, how does the graph convolution operation formula extract node features from the l-th layer of the graph convolution? Calculate the features of the next layer l+1. The calculation process includes:

[0094] Feature transformation: ,in, It is the feature matrix of all nodes in the l-th layer, and its shape is [number of nodes, current layer feature dimension]; It is a trainable parameter weight matrix of the l-th layer, responsible for mapping the features of the current layer to a new feature space. A linear transformation is performed on the feature vector of each node (equivalent to a fully connected layer), learning how to combine and transform the current features, which is similar to weight learning in traditional neural networks.

[0095] Neighbor aggregation: ,in, It is an adjacency matrix with self-loops added, i.e. , where A is the original adjacency matrix and I is the identity matrix. In the equation, if nodes i and j are connected, then =1; at the same time, diagonal elements Set to 1, indicating that each node is connected to itself. Matrix multiplication is involved. The purpose of ×(transformed feature matrix) is to sum the transformed feature vectors of all its neighboring nodes (including itself) for each node, thereby forming a new feature vector of the node that incorporates neighborhood information. This process is called message passing, where each node receives message features from its neighbors and aggregates them.

[0096] Nonlinear activation: ,in, A non-linear activation function, such as ReLU or Sigmoid, is used to introduce non-linearity into the model, enabling the network to learn and represent more complex feature relationships and avoid degenerating into a simple linear transformation after stacking multiple layers.

[0097] (3) Temporal Convolutional Module: A Temporal Convolutional Network (TCN) is used to capture temporal dynamics. TCN utilizes dilated convolution to expand the receptive field, effectively capturing long-term temporal dependencies. Its basic operation can be expressed as:

[0098]

[0099] This formula describes how to calculate the output feature F(t) at time point t. F(t) is the output feature value at time point t, which is an abstract representation of the current moment after the model has analyzed historical data. This means that we will accumulate all the results from k=1 to k=K, where K is the size of the convolution kernel, i.e. how many time points the current window can cover.

[0100] f(k) represents the weight parameters of the convolution kernel at position k. These weights are learned by the model through training, with the aim of identifying meaningful local patterns in the time series (such as upward trends, downward trends, specific waveforms, etc.).

[0101] Indicates the input sequence at time point The value of determines which points in history the data are sampled from. In the distribution network state estimation and fault location scenario of this application, this value is usually multi-dimensional, containing various measurement data and state information collected from the power grid at a specific time. Specifically, it can be represented as a feature vector: X(time point) = [voltage, current, active power, reactive power, frequency, switch state, ...]. For example, X(t-3) may contain the set of readings of physical quantities such as voltage, current, and power at all key monitoring points in the distribution network at time t-3.

[0102] d represents the dilation factor, which determines the stride or sampling interval of the convolutional kernel when scanning historical data. The dilation factor d determines the step size of the convolutional kernel when backtracking on the time axis. Specifically, when d=1, the convolutional kernel examines the data at each historical moment (t-1, t-2, t-3, ...), which is suitable for capturing very fine, short-term fluctuation patterns; when d=2, the convolutional kernel samples at intervals of one moment (t-2, t-4, t-6, ...), which allows it to see a longer history at the same cost (kernel size K); when d=4, the sampling interval is larger (t-4, t-8, t-12, ...), and the receptive field (i.e., the range of history that the model can see) becomes larger, which is suitable for capturing long-term trends and periodic patterns.

[0103] TCN achieves an exponential expansion of the receptive field by stacking multiple dilated convolutional layers and making the dilation factor d of each layer grow exponentially (e.g., 1, 2, 4, 8, ...). This allows the lower layers of the network to process recent details and the higher layers to integrate long-term patterns, thus capturing very long temporal dependencies with only a few layers.

[0104] In the context of this application, this mechanism offers the following advantages:

[0105] Balancing short-term transients and long-term trends: The instantaneous electrical changes during a fault can be keenly captured by a smaller d (lower-level network), while certain slowly developing abnormal signs before a fault (such as gradual voltage drift) can be effectively identified by a larger d (higher-level network); Efficiently handling long-term dependencies: Compared to RNN series models, TCN establishes a more direct and efficient correlation between the current state and distant historical events through an expansion mechanism, which is beneficial for analyzing the root cause of faults; Stable training: The training process of TCN is more stable, without the gradient vanishing or exploding problems common in RNNs, ensuring the reliability of model training.

[0106] In step S200, the training process of the spatiotemporal graph neural network model for each edge node includes:

[0107] S210. Obtain the historical multi-source operating data of the current edge node and its corresponding actual voltage amplitude and actual phase angle. For example, obtain the multi-source operating data of the current edge node within one year, such as current, voltage, power factor, and temperature, and obtain the actual voltage amplitude and actual phase angle under the corresponding multi-source operating data as training sample data for the edge node spatiotemporal graph neural network model deployed on the current edge node.

[0108] S220. Based on the historical multi-source operating data of the current edge node and its corresponding actual voltage amplitude and actual phase angle, the model parameters of the spatiotemporal graph neural network model corresponding to the current edge node are optimized with the goal of minimizing the prediction error. The loss function of the spatiotemporal graph neural network model is constructed from a measurement data fitting loss term and a physical constraint loss term. The measurement data fitting loss term represents the error between the predicted voltage amplitude and predicted phase angle output by the spatiotemporal graph neural network model and the corresponding actual voltage amplitude and actual phase angle. The physical constraint loss term represents the error between the predicted target operating parameters of the current edge node and the theoretical target operating parameters obtained based on a specified circuit law, given the current predicted voltage amplitude and predicted phase angle. The value of the physical constraint loss term is directly proportional to the error between the predicted and theoretical target operating parameters.

[0109] In this application, the loss function of the spatiotemporal graph neural network model of the edge node is constructed from the measurement data fitting loss term and the physical constraint loss term, including: determining the weights of the measurement data fitting loss term and the physical constraint loss term; and constructing the loss function of the spatiotemporal graph neural network model of the edge node by weighted summation of the measurement data fitting loss term and the physical constraint loss term based on their respective weights.

[0110] Specifically, the spatiotemporal graph neural network model for edge nodes can be deployed as a state estimation module on edge nodes. Based on deep learning-based state estimation methods, this model directly learns the system state mapping from measurement data, avoiding reliance on precise physical models. The inputs to the edge node spatiotemporal graph neural network model are real-time measurement data (voltage, current, power, etc.) and pseudo-measurement data (load forecasting, generation plans, etc.), while the outputs are the estimated voltage amplitude and phase angle of the edge nodes.

[0111] The loss function of the spatiotemporal graph neural network model for edge nodes includes two parts: measurement fitting loss and physical constraint loss, which can be specifically expressed as:

[0112]

[0113] in, L SE Represents the loss function. The loss term for fitting the measurement data is represented by the mean square error function. The physical constraint loss term is used to ensure that the estimation results conform to the circuit laws, and α and β are weighting coefficients.

[0114] In this application, the measurement data fitting loss term is used to make the model's output as close as possible to the data actually measured by the sensor, i.e. It calculates the error between the predicted value and the actual measured value. Commonly used methods include mean squared error (MSE) or mean absolute error (MAE).

[0115] Physical constraint loss term The evaluation measures whether the model's output conforms to the fundamental physical laws of the power system, such as Kirchhoff's Current Law (KCL) or Voltage Law (KVL). For instance, according to KCL, for a grid node, the sum of the inflow currents should equal the sum of the outflow currents. If the model's estimation results significantly violate this law, This will result in a large penalty value. Minimizing this term is equivalent to adding a physical hard constraint to the model, ensuring that its output is physically reasonable.

[0116] The weighting coefficients are hyperparameters used to control the relative importance of trusting the data and adhering to physics in the overall objective. If α is much larger than β, the model will focus more on fitting the measured data, but may produce some physically unreliable results (especially when the data is noisy or missing). If β is much larger than α, the model will strictly guarantee that the output conforms to physical laws, but may not be able to fully capture the complex patterns in the data, leading to insufficient estimation of the details of the real situation. Adjusting these two parameters aims to find the optimal balance between data-driven and physics-driven approaches.

[0117] Specifically, in this application, the specified circuit law is Kirchhoff's current law; the target operating parameter is the outflow current of the current edge node; the calculation steps for the error between the predicted value of the target operating parameter of the current edge node and the theoretical value of the target operating parameter obtained based on the specified circuit law include: obtaining the actual value of the inflow current of the current edge node, and determining the theoretical value of the outflow current of the current edge node based on the actual value of the inflow current and Kirchhoff's current law; determining the predicted value of the outflow current of the current edge node based on the current predicted voltage amplitude and predicted phase angle value of the current edge node; and using the absolute difference between the predicted value of the outflow current and the theoretical value of the outflow current as the error between the predicted value of the outflow current and the theoretical value of the outflow current.

[0118] Understandably, the actual inflow current of the current edge node can be acquired in real time by sensors, and the theoretical outflow current of the current edge node can be determined based on the topology of the current edge node, using the actual inflow current and Kirchhoff's Current Law. Kirchhoff's Current Law is existing technology, and its calculation process is not limited here. Then, the predicted outflow current of the current edge node is determined based on the current predicted voltage amplitude and predicted phase angle. For example, the constraint loss term based on KCL can be expressed as:

[0119]

[0120] Where N(i) is the set of all neighboring nodes connected to node i. Based on the model-predicted voltage magnitude and phase angle values ​​at nodes i and j, the line admittance is used. The calculated line current flowing from node i to node j, i.e., the predicted outflow current from node i, is given by the following: ,in, Line admittance is the line impedance. The reciprocal of the line impedance It can be obtained directly. V is the complex voltage at node i predicted by the model. i Let be the voltage amplitude at node i. Let i be the voltage phase angle at node i. V is the complex voltage at node j predicted by the model.j Let J be the voltage magnitude at node j. Let J be the voltage phase angle at node j. The actual injected current at node i can be calculated from the injected power measurement and node voltage measurement at node i, or derived from the current measurement of adjacent lines. It is the 2-norm (the square of the modulus) of a complex number.

[0121] This application, by embedding physical constraints into the loss function, effectively improves the model's generalization ability compared to traditional pure data-driven methods. Even in extreme operating conditions or fault scenarios not covered by training data, the physical constraints force the model to adhere to the most basic physical conservation laws, thus guiding the model to provide relatively reasonable estimates. Simultaneously, this application enhances the model's robustness. When measurement data contains noise or is partially missing, the physical constraints can correct biases, preventing the model from being misled by poor data and outputting results that clearly violate physical laws. Furthermore, this application reduces the model's dependence on the amount of data. Physical laws themselves provide strong prior knowledge; combined with physical constraints, the model can learn more fundamental mapping relationships without extremely large datasets, alleviating the data hunger problem to some extent. In other words, in the distribution network state estimation and fault location scenario of this application, the loss function ensures that the AI-based model can learn from limited, potentially noisy, distribution network measurement data, and its output system state (such as node voltage and phase angle) and fault location results can always remain consistent with the actual physical behavior of the power grid.

[0122] In this application, the training process of the global spatiotemporal graph neural network model is the same as that of the edge node spatiotemporal graph neural network model, and will not be described again here. For example, the aggregation server obtains historical data on voltage amplitude and phase angle of each node, as well as the topology and historical fault information of the distribution network. The historical fault information may include information such as the number of historical faults and historical fault sections of the distribution network. The global spatiotemporal graph neural network model is trained using the historical data on voltage amplitude and phase angle of each node, the topology and historical fault information of the distribution network as input, and the fault section location information of the distribution network as output. The model parameters of the trained global spatiotemporal graph neural network model can be distributed to each node as the initial global model parameters.

[0123] This application employs a multi-scale fault location strategy, comprising two stages: section location and precise fault point location. The section location stage uses a graph neural network to analyze changes in node characteristics to determine the faulty line section. Specifically, the distribution network topology, real-time operational data (such as current and voltage), and historical alarm information are first used as features of nodes and edges, input into the graph neural network model. The model ultimately outputs a probability distribution indicating the likelihood of a fault occurring in each section, thus identifying the most probable faulty section. For example, the section location stage outputs the location information of the faulty bus and obtains the opening and closing status information of distribution network equipment switches from distribution automation terminals (such as FTUs).

[0124] In the precise fault location stage, the fault location and identification model takes the fault area information output by the previous layer, namely the global spatiotemporal graph neural network model (segment location stage), and the opening and closing status of the distribution network equipment switches as input. Combined with the power grid physical equations, it performs precise fault location within the narrowed search space. Specifically, the fault location and identification model of this application is constructed through the following steps:

[0125] S310. Determine the topology of each power distribution device in the fault section based on the topology of the power distribution network. For example, after obtaining the fault section location information, for example, the global spatiotemporal graph neural network model identifies that there is a fault between node 1 and node 2, then obtain the topology of each power distribution device between node 1 and node 2.

[0126] S320. Based on the topology of each power distribution device within the fault section, divide the fault section into multiple fault sub-sections and determine the weight of each fault sub-section. Each fault sub-section includes one power distribution device. For example, according to the topology of the fault section, divide each power distribution device into different sub-sections.

[0127] S330. Construct a linear penalty term based on the fault state and weight of each fault sub-section, and construct an error penalty term based on the measured overcurrent information and theoretical overcurrent information of the upstream fault sub-section of each fault sub-section under the corresponding fault state and the corresponding power distribution equipment switching state.

[0128] The linear penalty term is constructed based on the fault state and weight of each fault sub-segment, including: constructing a linear penalty term by weighted summation of the fault states of each fault sub-segment based on the weight of each fault sub-segment.

[0129] An error penalty term is constructed based on the measured and theoretical overcurrent information of the upstream fault sub-segments of each fault sub-segment under the corresponding fault state and the corresponding power distribution equipment switching state. This includes: constructing a switching function that represents the theoretical overcurrent information of the upstream fault sub-segments of any fault sub-segment under the corresponding fault state and the corresponding power distribution equipment switching state; and constructing an error penalty term based on the sum of the absolute differences between the measured overcurrent information of all upstream fault sub-segments of any fault sub-segment under the corresponding fault state and the theoretical overcurrent information obtained from the switching function.

[0130] In this application, within the fault area, each possible line segment is defined as a suspected fault unit, and a binary variable x is used to represent it. i Indicate its state, for example, x i =1 indicates a fault in this section, x i =0 represents normal. The goal of precise fault location is to find a set of fault segments from these candidate sub-segments that best match the actual monitored fault information. This application transforms the fault location problem into an optimization problem by introducing physical constraints, and obtains the optimal fault location through an optimization algorithm.

[0131] S340. Construct the objective function based on the sum of the linear penalty term and the error penalty term. Specifically, the objective function of this application, i.e., the fitness function, is as follows:

[0132]

[0133] Where, x i This indicates the state of faulty sub-segment i (0 indicates non-faulty, 1 indicates faulty). The weight of faulty sub-segment i, This indicates the measured overcurrent information at switch j. This represents the calculated value of the overcurrent information at switch j under fault state x. m represents the number of upstream fault sub-segments of fault sub-segment i, and n represents the number of upstream specific power distribution equipment or specific switches of power distribution equipment or switch j. The upstream fault sub-segments of fault sub-segment i and the upstream specific switches of switch j can be determined by the power distribution equipment topology of the fault sub-segment.

[0134] in, For linear penalty terms, This is an error penalty term. This involves modeling the switching function, i.e., the physical constraint. The switching function uses a mathematical logic formula to characterize the physical law that if a section fails, a specific upstream switch can detect an overcurrent signal. This application uses an objective function to represent the fault state x of the section. i Overcurrent information detected by the switch Connect them. For example, a simplified switching function can be expressed as: =Logical function(x) i ,...),in It is the overcurrent state that switch j should theoretically report under a given fault state combination x.

[0135] S350. With minimizing the objective function as the objective, a fault location and identification model is constructed based on the objective function.

[0136] After constructing the objective function, the model transforms the precise location problem into an optimization problem, namely, finding a set of fault sub-segment states x such that the theoretical switching signal calculated based on this set of states... With the actual transmitted switch signal The difference between them is minimal.

[0137] Among them, the error term Measuring the difference between theoretical and actual values ​​is the core matching indicator; regularization term This is a penalty for the number of faulty segments, which tends to choose simpler solutions with fewer faulty segments (i.e., consistent with the common sense of single faults or multiple but finite faults in reality), thus preventing overfitting and improving fault tolerance. When minimizing the overall objective function f(x), not only should the error term be minimized, but the regularization term should also be minimized as much as possible. In this way, the algorithm will find solutions that can basically explain most of the observed phenomena with the fewest critical fault points, which is consistent with reality, where a fault event is usually caused by a single or very few root causes. Understandably, the optimization process of the objective function can be implemented using existing optimization algorithms, such as particle swarm optimization and genetic algorithms, and this is not limited here.

[0138] In summary, this application combines a federated learning framework with edge computing to construct a distributed intelligent fault handling system, achieving a balance between localized data processing and global knowledge sharing, effectively solving the problems of data silos and privacy protection. Simultaneously, it constructs a spatiotemporal graph neural network specifically for distribution networks, capturing both spatial relationships and dynamic temporal changes between nodes, improving the accuracy of state estimation and fault location. This model organically combines graph convolutional networks and temporal convolutional networks, achieving efficient modeling of the spatiotemporal characteristics of distribution networks. Through a bidirectional information feedback mechanism, it achieves collaborative optimization of state estimation and fault location, overcoming the limitations of traditional methods that treat the two problems in isolation, thus improving overall performance. It adopts a two-stage fault location strategy combining segment location and precise ranging, first determining the line segment where the fault occurred, and then performing precise ranging within that segment, balancing location efficiency and accuracy. Furthermore, it dynamically adjusts the aggregation weights in federated learning based on the quality, quantity, and age of node data, optimizing global model performance and effectively improving model convergence speed.

[0139] like Figure 4 As shown, in a second aspect, this application provides a power distribution network condition prediction and fault location device, comprising:

[0140] The data acquisition module is configured to receive state estimation data from at least one edge node. The state estimation data is obtained by each edge node through a spatiotemporal graph neural network model of the edge node pre-constructed based on the topology of the corresponding edge node, after calculating the multi-source operating data obtained by the corresponding edge node.

[0141] The first fault location module is configured to acquire the topology of the distribution network and the historical fault information of the distribution network. Taking the topology of the distribution network, the historical fault information of the distribution network and the state estimation data of each edge node as input, the fault location information of the distribution network is obtained through a global spatiotemporal graph neural network model pre-constructed based on the topology of the distribution network.

[0142] The second fault location module is configured to acquire the switch status of the power distribution equipment in the fault section. Taking the fault section location information and the switch status of the power distribution equipment as input, the fault location information of the power distribution network is determined by the fault location identification model constructed based on the switch status of the power distribution equipment.

[0143] It is understood that those skilled in the art will clearly recognize that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0144] In a third aspect, this application provides a machine-readable storage medium storing instructions that, when executed by a processor, configure the processor to perform the power distribution network state prediction and fault location method described above.

[0145] In a fourth aspect, this application provides a terminal device including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the power distribution network state prediction and fault location method described above.

[0146] like Figure 5 The diagram shown is a schematic representation of a terminal device provided in an embodiment of this application. Figure 5 As shown, the terminal device 10 of this embodiment includes a processor 100, a memory 101, and a computer program 102 stored in the memory 101 and executable on the processor 100. When the processor 100 executes the computer program 102, it implements the steps in the above method embodiments. Alternatively, when the processor 100 executes the computer program 102, it implements the functions of each module / unit in the above device embodiments.

[0147] For example, computer program 102 may be divided into one or more modules / units, one or more of which are stored in memory 101 and executed by processor 100 to complete this application. 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 computer program 102 in terminal device 10.

[0148] Terminal device 10 may be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. Terminal device 10 may include, but is not limited to, a processor 100 and a memory 101. Those skilled in the art will understand that... Figure 5 This is merely an example of terminal device 10 and does not constitute a limitation on terminal device 10. It may include more or fewer components than shown, or combine certain components, or different components. For example, terminal device may also include input / output devices, network access devices, buses, etc.

[0149] The processor 100 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. The general-purpose processor can be a microprocessor or any conventional processor.

[0150] The memory 101 can be an internal storage unit of the terminal device 10, such as a hard disk or RAM of the terminal device 10. The memory 101 can also be an external storage device of the terminal device 10, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or FlashCard equipped on the terminal device 10. Furthermore, the memory 101 can include both internal and external storage units of the terminal device 10. The memory 101 is used to store computer programs and other programs and data required by the terminal device 10. The memory 101 can also be used to temporarily store data that has been output or will be output.

[0151] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0152] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0153] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for predicting the condition of a power distribution network and locating faults, characterized in that, include: Receive state estimation data of at least one edge node, the state estimation data being calculated by each edge node from the multi-source operating data obtained by the corresponding edge node through an edge node spatiotemporal graph neural network model pre-constructed based on the topology of the corresponding edge node; The topology and historical fault information of the distribution network are obtained. The fault location information of the distribution network is obtained by using the topology, historical fault information and state estimation data of each edge node as input and a global spatiotemporal graph neural network model pre-constructed based on the topology of the distribution network. The switch status of the power distribution equipment in the fault section is obtained. The fault location information of the fault section and the switch status of the power distribution equipment are used as inputs. The fault location information of the power distribution network is determined by the fault location identification model constructed based on the switch status of the power distribution equipment. The fault location and identification model is constructed through the following steps: The topology of each power distribution device in the fault section is determined based on the topology of the power distribution network. Based on the topology of each power distribution device within the fault section, the fault section is divided into multiple fault sub-sections, and the weight of each fault sub-section is determined. Each fault sub-section includes one power distribution device. A linear penalty term is constructed by weighting and summing the fault states of each fault sub-segment based on the weight of each fault sub-segment. Construct a switching function that represents the theoretical overcurrent information of the upstream fault sub-segment under the corresponding fault state and the switching state of its corresponding power distribution equipment; An error penalty term is constructed by summing the absolute differences between the measured overcurrent information of all upstream fault sub-segments and the theoretical overcurrent information obtained based on the switching function for any fault sub-segment under the corresponding fault state. Construct the objective function based on the sum of the linear penalty term and the error penalty term; The fault location and identification model is constructed based on the objective function, which aims to minimize the objective function. The objective function is: Where, x i This represents the fault state of faulty sub-segment i, where 0 indicates no fault and 1 indicates a fault. The weight of faulty sub-segment i, This indicates the measured overcurrent information at switch j. This represents the theoretical overcurrent information at switch j under fault state x, m represents the number of upstream fault sub-segments of fault sub-segment i, and n represents the number of specific power distribution equipment or specific switches upstream of power distribution equipment or switch j.

2. The method for predicting the state of a power distribution network and locating faults according to claim 1, characterized in that, The training process of the spatiotemporal graph neural network model for edge nodes includes: S1. Initialize the global model parameters of the global spatiotemporal graph neural network model; S2. Send the global model parameters to each edge node so that each edge node uses the global model parameters as the initial model parameters of the spatiotemporal graph neural network model of the corresponding edge node; S3. Receive the local model parameters of the corresponding edge node spatiotemporal graph neural network model sent by each edge node. The local model parameters are obtained by optimizing the corresponding edge node spatiotemporal graph neural network model with the local training data of each edge node. S4. Obtain the number of samples of local training data for each edge node, determine the weight of each edge node based on the number of samples of local training data for each edge node, and perform a weighted average of the local model parameters of the corresponding edge node spatiotemporal graph neural network model using the weight of each edge node, and update the global model parameters with the weighted average result. If the preset maximum number of iterations is reached, execute step S5; otherwise, execute step S2. S5. Use the updated global model parameters as the target global model parameters, use the target global model parameters as the final model parameters of the global spatiotemporal graph neural network model, and send the target global model parameters to each edge node so that each edge node uses the target global model parameters as the final model parameters of the corresponding edge node spatiotemporal graph neural network model.

3. The method for predicting the state of a power distribution network and locating faults according to claim 2, characterized in that, The weights of each edge node are determined based on the number of samples in the local training data of each edge node, including: Determine the total number of samples in the local training data for all edge nodes; For each edge node, the weight of the current edge node is the ratio of the number of samples in the local training data of the current edge node to the total number of samples.

4. The method for predicting the state of a power distribution network and locating faults according to claim 1, characterized in that, The multi-source operating data includes at least one of the voltage, current, power, frequency, ambient temperature, load forecast data and power generation plan data of the corresponding edge node, and the state estimation data includes the voltage amplitude and phase angle value of the corresponding edge node. The training process of the spatiotemporal graph neural network model for each edge node includes: Obtain the historical multi-source operating data of the current edge node and its corresponding actual voltage amplitude and actual phase angle value; Based on the historical multi-source operating data of the current edge node and its corresponding actual voltage amplitude and actual phase angle, with the goal of minimizing the prediction error, the model parameters of the edge node spatiotemporal graph neural network model corresponding to the current edge node are optimized. The loss function of the edge node spatiotemporal graph neural network model is constructed by the measurement data fitting loss term and the physical constraint loss term. The measurement data fitting loss term is used to represent the error between the predicted voltage amplitude and predicted phase angle values ​​output by the spatiotemporal graph neural network model of the edge nodes and the corresponding actual voltage amplitude and actual phase angle values; The physical constraint loss term is used to represent the error between the predicted value of the target operating parameter of the current edge node and the theoretical value of the target operating parameter obtained based on the specified circuit law, under the current predicted voltage amplitude and predicted phase angle. The value of the physical constraint loss term is directly proportional to the error between the predicted value of the target operating parameter and the theoretical value of the target operating parameter.

5. The method for predicting the state of a distribution network and locating faults according to claim 4, characterized in that, The specified circuit law is Kirchhoff's current law; the target operating parameter is the outflow current of the current edge node; The steps for calculating the error between the predicted target operating parameters of the current edge node and the theoretical target operating parameters obtained based on the specified circuit law include: Obtain the actual value of the inflow current of the current edge node, and determine the theoretical value of the outflow current of the current edge node based on the actual value of the inflow current and Kirchhoff's current law. The predicted outflow current of the current edge node is determined based on the current predicted voltage magnitude and predicted phase angle of the current edge node. The absolute difference between the predicted outflow current and the theoretical outflow current is taken as the error between the predicted outflow current and the theoretical outflow current.

6. The method for predicting the state of a power distribution network and locating faults according to claim 4, characterized in that, The loss function of the spatiotemporal graph neural network model for edge nodes is constructed from a measurement data fitting loss term and a physical constraint loss term, including: Determine the weights of the measurement data fitting loss term and the physical constraint loss term; The loss function of the spatiotemporal graph neural network model of the edge nodes is constructed by weighting and summing the measurement data fitting loss term and the physical constraint loss term according to their respective weights.

7. A distribution network condition prediction and fault location device, employing the distribution network condition prediction and fault location method as described in any one of claims 1-6, characterized in that, include: The data acquisition module is configured to receive state estimation data of at least one edge node. The state estimation data is obtained by each edge node through a spatiotemporal graph neural network model of the edge node pre-constructed based on the topology of the corresponding edge node, after calculating the multi-source operating data obtained by the corresponding edge node. The first fault location module is configured to acquire the topology of the distribution network and the historical fault information of the distribution network. Taking the topology of the distribution network, the historical fault information of the distribution network and the state estimation data of each edge node as input, the fault location information of the distribution network is obtained through a global spatiotemporal graph neural network model pre-constructed based on the topology of the distribution network. The second fault location module is configured to acquire the switch status of the power distribution equipment in the fault section, and use the fault section location information and the switch status of the power distribution equipment as input to determine the fault location information of the power distribution network through a fault location identification model constructed based on the switch status of the power distribution equipment.

8. A machine-readable storage medium storing instructions thereon, characterized in that, When executed by the processor, the instruction causes the processor to be configured to perform the power distribution network condition prediction and fault location method as described in any one of claims 1-6.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the power distribution network status prediction and fault location method as described in any one of claims 1-6.

Images