A Method and System for Predicting the Overload Rate of Distribution Transformers Based on Multi-Source Data Fusion

Abstract

This invention discloses a method and system for predicting the overload rate of distribution transformers based on multi-source data fusion, relating to the field of load forecasting technology. The method includes the following steps: acquiring transformer operation data for a target area and constructing a knowledge graph; extracting entities and features from the knowledge graph that have causal paths with the target transformer's overload event, generating a feature tensor; embedding the knowledge graph using a graph neural network, and inputting the graph embedding and feature tensor into a time series prediction model to construct a joint prediction network; simulating intervention on the input variables based on causal reasoning methods and evaluating the strength of the impact on the prediction results; outputting the predicted overload rate value of the target transformer for future time periods based on the strength of the simulated intervention, and generating the main causal paths leading to changes in the predicted value. This application achieves accurate and interpretable prediction through multi-source data fusion and a causal knowledge graph.

Description

Technical Field

[0001] This invention relates to the field of load forecasting technology, and more specifically, to a method and system for predicting the overload rate of distribution transformers based on multi-source data fusion. Background Technology

[0002] With the continuous development of power systems and the advancement of smart grid technology, the distribution network, as a crucial component of the power system, directly impacts the reliability of power supply and the user experience. Distribution transformers, as core equipment in the distribution network, undertake the critical task of converting electrical energy from the high-voltage side to the low-voltage side; their operational health directly affects the stability of the entire distribution network. In recent years, with the widespread application of distributed power sources, big data technology, and the Internet of Things (IoT), distribution transformer operating data has become multi-source, diverse, and dynamic. How to effectively utilize this multi-source data to accurately predict the load status of distribution transformers has become a key technical challenge in the field of smart distribution networks.

[0003] For example, the invention patent with announcement number CN118396193A discloses a method and system for early warning of transformer reverse overload / heavy load based on neural networks, relating to the field of power system technology. It includes: data acquisition steps, feature extraction steps, load rate prediction feature set and data association graph construction steps, neural network construction and training steps, and prediction steps. This invention, after constructing a data association graph based on empirical knowledge and Euclidean distance from distribution transformer data, aggregates and transmits data from similar scenarios through a GraphSAGE network, mining and further enhancing the implicit relationships between variables, thereby improving the model's early warning accuracy for high-dimensional, small-sample data. Using LSTM as the downstream prediction task of GraphSAGE can mine the time-series relationships of distribution transformer data, further improving the model's fitting ability.

[0004] For example, the invention patent with announcement number CN113298318A discloses a novel method for predicting heavy overload of distribution transformers, including the following steps: Step S1: Obtain basic data from the operation, distribution, and dispatching related production and management systems and meteorological data networks; Step S2: Perform correlation analysis on various feature quantities to identify feature quantities with strong correlation to heavy overload of distribution transformers as inputs to the prediction model; Step S3: Design the prediction model and train and test the model using the dataset. This invention proposes a prediction model based on a multi-scale one-dimensional CNN-LSTM network with an attention mechanism. This model combines a multi-scale one-dimensional CNN with an LSTM network, effectively integrating the speed and lightweight nature of convolutional neural networks with the information processing capabilities of the three gate pairs of LSTM networks. The attention mechanism introduced at the network output can allocate attention weights to the output of the LSTM hidden layers, highlighting the influence of key factors on heavy overload of distribution transformers, ultimately achieving the prediction of heavy overload of distribution transformers, and providing certain guidance for actual production work.

[0005] The above-disclosed technical solutions have at least the following technical problems:

[0006] Existing methods for predicting the overload rate of distribution transformers often rely on a single data source, such as historical load data or operating status data, lacking comprehensive utilization of multi-source information, which limits prediction accuracy. Furthermore, traditional feature extraction methods struggle to capture structural dependencies and temporal dynamic changes between paths and lack effective fusion of contextual information.

[0007] To address the above problems, this invention proposes a solution. Summary of the Invention

[0008] To overcome the aforementioned deficiencies in the prior art, embodiments of the present invention provide a method and system for predicting the overload rate of distribution transformers based on multi-source data fusion. By constructing a causal knowledge graph and joint prediction model based on multi-source data fusion, the method addresses the problems of data heterogeneity, complex causal relationships, and insufficient prediction accuracy in the prediction of the overload rate of distribution transformers. This achieves high-precision, interpretable, and dynamically adjustable overload rate prediction, thereby improving the intelligent operation and maintenance capabilities of the power distribution system.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] The method for predicting the overload rate of distribution transformers based on multi-source data fusion includes: acquiring transformer operation data of the target area and constructing a knowledge graph; extracting entities and features with causal paths to the overload events of the target transformer from the knowledge graph and generating feature tensors; embedding the knowledge graph using a graph neural network and inputting the graph embedding and feature tensors into a time series prediction model to construct a joint prediction network; simulating intervention on the input variables based on causal reasoning methods and evaluating the strength of the impact on the prediction results; and outputting the predicted overload rate of the target transformer for future time periods based on the strength of the simulated intervention, and generating the main causal paths that lead to changes in the predicted values.

[0011] In a preferred embodiment, the acquisition of transformer operation data in the target area and the construction of a knowledge graph are specifically as follows: Based on the business logic of the power distribution system and the transformer operation data, a knowledge graph is constructed, and the core entity types and relationship types in the knowledge graph are determined; based on the entity types and relationship types, structured triples are extracted from the preprocessed transformer operation data; for triples containing time-series variables, a causal discovery algorithm is used to identify potential causal relationships, outputting and filtering causal edges with confidence levels higher than a threshold to obtain corrected triples; a directed graph-based knowledge graph is constructed based on the corrected triples; the prediction error between the output of the overload rate prediction model and the actual observed data is obtained, and key causal paths are identified based on the prediction error; based on the key causal paths, the confidence levels of the corresponding edges are dynamically adjusted, and the knowledge graph is updated.

[0012] In a preferred embodiment, the step of extracting entities and features with causal paths to the target transformer overload event from the knowledge graph and generating a feature tensor is as follows: Obtain the target transformer number and prediction time interval in the prediction task, and retrieve relevant nodes and historical time series labels; starting from the target transformer entity node, search for causally oriented paths within a range of no more than L hops according to a preset directed traversal strategy to form an initial candidate path set; calculate path scores based on the causal edge confidence and path depth, filter out paths with scores below a threshold, and retain high-value paths to obtain the final candidate path set; extract attribute features of entities involved in the candidate paths within a specified time period to form path feature vectors; align the path feature vectors in each path along the time dimension and concatenate them to generate a path feature matrix; generate the final feature tensor based on the fusion matrix information of path scores and causal confidence.

[0013] In a preferred embodiment, the step of generating the final feature tensor based on the fusion matrix information of path score and causal confidence is as follows: Extract the temporal features of nodes for each path within the time window, and arrange them in topological order to form a feature sequence matrix; use path score and causal confidence as attention weights, and apply a graph attention mechanism to weighted aggregate the feature sequences to generate path aggregate features; extract the temporal context and regional context corresponding to the path, construct a context feature vector, and input it into a context gating function to generate an activation vector; multiply the activation vector element-wise with the aggregate feature vector to obtain the context-enhanced feature vector; within each time step, fuse the context-enhanced features with the original temporal structure, and use average pooling to generate the final path features; organize the final path feature vectors of all time steps and paths according to temporal order and path number to finally construct a three-dimensional structured feature tensor.

[0014] In a preferred embodiment, the step of embedding the knowledge graph using a graph neural network and inputting the graph embedding and feature tensor into a time series prediction model to construct a joint prediction network is as follows: Iterative propagation and embedding learning are performed on the knowledge graph based on the graph neural network model to generate an embedding vector for each entity node; multiple rounds of iterative propagation and feature aggregation are performed on the embedding vectors to capture the topological dependencies and semantic relevance between nodes, generating the final embedding representation for each entity node; the embedding vectors are integrated using global average pooling to generate a global embedding vector, and a linear transformation is performed on the global embedding vector to align it with the time dimension of the feature tensor; the aligned global embedding vector and the feature tensor are concatenated along the feature dimension to form a joint input tensor; based on the joint input tensor, a time series prediction model for transformer overload rate prediction is constructed.

[0015] In a preferred embodiment, the step of constructing a time series prediction model for transformer overload rate prediction based on the joint input tensor is as follows: Using the Transformer model, the semantic relationships and feature associations of key paths are extracted from the joint input tensor of each path within the compressed time window to generate path feature vectors; the path feature vectors are expanded according to the time series to form a time-step-level path feature sequence, which is then input into a gated recurrent unit network for training to obtain the time series prediction model; the output of the time series prediction model is input into a multilayer perceptron to perform stepwise regression prediction of the overload rate of the target transformer for the next t time steps, obtaining a sequence of predicted overload rates; the mean square error between the predicted overload rate sequence and the actual predicted overload rate sequence is used as the prediction loss function; based on the output value of the prediction loss function, the parameters in the time series prediction model are iteratively updated using the backpropagation algorithm, with the goal of minimizing the prediction loss function.

[0016] In a preferred embodiment, the step of simulating intervention on input variables based on causal reasoning and evaluating the strength of the impact on the prediction results is as follows: Based on the knowledge graph, feature nodes with reachable causal paths to the target prediction variable are identified. A path dependency matrix is ​​constructed by combining path scores and historical covariate change data. Variables with causal path scores higher than a set threshold are selected as a set of candidate intervention variables. The frequency of each candidate variable in all causal paths is obtained, and variables are weighted according to path importance. Variables with scores higher than the threshold are selected to form the final set of intervention variables. Intervention operations are set for each intervention variable to generate an intervention sample input tensor. The feature tensors before and after intervention are input into the time series prediction model, and the transformer overload rate prediction results under the two inputs are recorded. The difference between the prediction results is calculated as the influence strength value of the prediction result. If the intervention variable appears in multiple paths, the influence strength on each path is calculated separately, and these paths are weighted and averaged according to path importance weights to form a normalized influence strength vector. The influence strength value of the prediction results under simulated intervention is corrected.

[0017] In a preferred embodiment, the correction of the influence intensity value of the predicted results under simulated intervention is specifically as follows: For each candidate intervention variable, a structural consistency function is constructed based on its causal path connectivity in the graph. If a variable is only connected by weak edges and path breaks, its influence intensity will be weakened proportionally to obtain a preliminarily corrected influence value. Multiple rounds of prediction are performed on the samples before and after intervention using a neural network based on MC Dropout. If the intervention operation of the variable causes large model fluctuations, the influence intensity is weighted and compressed according to the prediction uncertainty. For each candidate variable, multiple intervention perturbation amplitudes are set, and corresponding counterfactual input samples are constructed. Multiple intervention perturbation amplitudes are set for each variable to generate counterfactual samples and construct an intervention response curve. If the response curve has high nonlinear fluctuations, the influence intensity under multiple perturbations is averaged and smoothed to obtain a local influence intensity value. The set of all causal paths involved in each variable in the graph is obtained, and the stability score of each path is calculated. The local influence intensity value on each path is weighted and fused with the corresponding stability score to finally calculate the normalized influence intensity value of the variable.

[0018] In a preferred embodiment, the step of outputting the predicted load rate of the target transformer for future time periods based on the influence intensity of the simulated intervention, and generating the main causal path leading to the change in the predicted value, specifically includes: weighting the intervention variables corresponding to them in the joint input feature tensor based on the corrected variable influence intensity vector to generate a corrected joint input tensor; inputting the corrected joint input tensor into the trained time series prediction model to output the sequence of predicted load rates for future time periods; retrieving all reachable causal paths corresponding to each intervention variable based on a knowledge graph, and recording the stability score of each path; and calculating the stability score of each node in the path. The impact intensity of the simulated intervention is used to weight and score the paths, output the contribution of each path to the prediction results, and sort them by contribution to select several key causal paths. The intervention magnitude, predicted response value change, and path score of each node in the key causal paths are extracted to form a path-level causal impact explanation vector. The prediction sequence is mapped and integrated with the key path causal explanation vector to establish the causal path relationship between the intervention variable and the predicted change. Based on the path mapping relationship, the predicted value sequence of the target transformer's overload rate in the future time period is output, and together with its key causal path set and corresponding causal impact explanation vector, a prediction result set with causal explanation capability is formed.

[0019] The distribution transformer overload rate prediction system based on multi-source data fusion includes a data acquisition module, an extraction module, a construction module, an evaluation module, and an output module, with connections between the modules. The data acquisition module acquires transformer operation data for the target area and constructs a knowledge graph. The extraction module extracts entities and features from the knowledge graph that have causal paths with the target transformer overload event, generating a feature tensor. The construction module embeds the knowledge graph using a graph neural network and inputs the graph embedding and feature tensor into a time series prediction model to construct a joint prediction network. The evaluation module simulates intervention on the input variables based on causal reasoning methods and evaluates the strength of the impact on the prediction results. The output module outputs the predicted overload rate of the target transformer for future time periods based on the strength of the simulated intervention and generates the main causal paths leading to changes in the predicted values.

[0020] The technical effects and advantages of the present invention, which discloses a method and system for predicting the overload rate of distribution transformers based on multi-source data fusion, are as follows:

[0021] 1. This invention integrates multi-source heterogeneous data to construct a structured knowledge graph, organically fusing the operational data, business logic, temporal characteristics, and inherent causal relationships of distribution transformers to form a comprehensive and dynamic multi-dimensional data expression system. By utilizing a causal discovery algorithm to filter causal edges and correct the knowledge graph for time-series triples, it not only improves the accuracy and consistency of the data but also ensures that model predictions are based on real and high-confidence causal relationships, making the prediction results more interpretable and credible.

[0022] 2. This invention constructs a joint prediction network by deeply integrating graph neural networks with a time series prediction model. The graph neural network performs multi-round iterative propagation and feature aggregation on the knowledge graph, accurately mining the topological dependencies and semantic relevance between nodes to generate high-quality entity embedding vectors. Subsequently, through the synergistic effect of the Transformer and Gated Recurrent Unit (GRU) models, the dynamic features of time series between paths are deeply learned, achieving multi-step accurate prediction of transformer overload rate. This joint model significantly improves the prediction accuracy and robustness compared to traditional single models. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the process for predicting the overload rate of distribution transformers based on multi-source data fusion, as described in this invention.

[0024] Figure 2 This is a schematic diagram of the structure of the distribution transformer overload rate prediction system based on multi-source data fusion according to the present invention. Detailed Implementation

[0025] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0026] Example 1, Figure 1 The present invention provides a method for predicting the overload rate of distribution transformers based on multi-source data fusion, comprising the following steps:

[0027] S1. Obtain transformer operation data for the target area and construct a knowledge graph;

[0028] In this embodiment, transformer operation data for the target area is acquired, and a knowledge graph is constructed, as follows:

[0029] Based on the business logic of the power distribution system and the transformer operation data, a knowledge graph is constructed, and the core entity types and relationship types in the knowledge graph are determined. The entities include, but are not limited to, power distribution transformers, user groups, climate environment, geographical region and time tag, etc. The relationship types include belonging to, depending on, influencing, causing, etc., which are used to represent the structural relationship or causal dependency between entities.

[0030] Based on entity type and relation type, structured triples are extracted from preprocessed transformer operation data. For structured data, a rule-based mapping method is used. For unstructured text data, a named entity recognition and relation extraction model based on a pre-trained language model is preferred to extract potential triples.

[0031] For triples containing time-series variables, causal discovery algorithms (such as Granger causality test) are used to identify potential causal relationships and output the confidence score of each causal edge. Causal relationships with confidence scores below a set threshold are filtered out based on the confidence scores, thereby improving the accuracy and reliability of the knowledge graph and obtaining the corrected triples.

[0032] A directed graph structure is constructed based on the corrected triples, where nodes represent entities, edges represent structural dependencies or causal relationships between entities, and edge attributes are used to record causal confidence, time information, and data source.

[0033] Obtain the prediction error between the output of the heavy load rate prediction model and the actual observation data, and identify the causal chain that affects the prediction result, i.e. the key causal path, based on the prediction error.

[0034] Based on the key causal path, the confidence of the corresponding edge is dynamically adjusted and the knowledge graph is updated. The updated knowledge graph serves as the input data for the next round of overload rate prediction model, forming a dynamically evolving knowledge graph update mechanism.

[0035] The confidence score formula for each causal edge is as follows:

[0036]

[0037] The formula for the lag order is as follows:

[0038]

[0039] In the formula: This represents the confidence score of the causal edge X→Y. It is the natural base. It is the adjustment parameter for causal confidence. It is the F-statistic. It is the lag order obtained through the F-test, used to determine whether the null hypothesis is rejected. This is the significance level, typically set to 0.05. It is the sum of squared residuals without a regressor term. It is the sum of squared residuals with regressive terms. It is the sample size, where and It was obtained directly through regression analysis of time series data.

[0040] S2, extract entities and features that have a causal path with the target transformer overload event from the knowledge graph, and generate feature tensors;

[0041] In this embodiment, entities and features that have a causal path with the target transformer overload event are extracted from the knowledge graph, and a feature tensor is generated, as follows:

[0042] Obtain the target transformer number to be predicted in the current prediction task, determine the corresponding prediction time interval, and retrieve the nodes and historical time series tags related to the target transformer from the knowledge graph based on the prediction time interval;

[0043] Starting from the target transformer entity node, a set of paths with causal orientation to the target transformer overload event is searched in the current knowledge graph using a preset directed traversal strategy (such as breadth-first search, depth-first search, or weighted random walk). Causal paths with a path length not exceeding a preset number of hops L are retained to obtain an initial candidate path set.

[0044] For each path, output a path score based on the confidence score of the causal edges on the path and the path depth;

[0045] The candidate paths are sorted from high to low scores, and paths with scores below a set threshold are removed. High-value causal paths are retained for feature extraction, resulting in the final set of candidate paths.

[0046] Based on the candidate path set, the attribute feature values ​​of the entity nodes involved in each path within a fixed time period are extracted to form a path feature vector. The attribute feature values ​​include temperature, current, load, operating status, etc.

[0047] Align the entity feature vectors within each path along the time dimension and concatenate them to generate a path feature matrix;

[0048] For each row in the path feature matrix, the feature tensor is generated by fusing the corresponding path score and the causal confidence score of the edges in the path.

[0049] In this embodiment, for each row of the path feature matrix, the feature tensor is generated by fusing the corresponding path score and the causal confidence score of the edges in the path, as follows:

[0050] For each causal path, extract its time series features within the time window, arrange them according to the topological order of the nodes in the path, and obtain the feature sequence matrix of the path. The time series features include, but are not limited to, frequency deviation, circuit breaker opening and closing status, entropy value, etc.

[0051] The path score and the causal confidence of each edge in the path are used as the weight input for path attention, and the path feature sequence matrix is ​​weighted and aggregated through a graph attention mechanism to generate the aggregated feature vector of the path.

[0052] Extract the time context (e.g., hourly segment, whether it is a holiday, season) and regional context (e.g., city / rural, voltage level, user density) corresponding to the path, construct the context feature vector, and input it into the context gating function to generate the activation vector;

[0053] The activation vector is multiplied element-wise with the aggregated feature vector to obtain the context-enhanced feature vector;

[0054] For all paths within each time step, collect their context-enhanced feature vectors. Combined with the temporal structure in the path feature sequence matrix, the context-enhanced features and global structural features are concatenated and fused using the average pooling method to form the final path feature vector.

[0055] The final path feature vectors of all time steps and paths are organized according to time order and path number to finally construct a three-dimensional structured feature tensor. ,in, For the time dimension, For the number of paths, For feature dimensions.

[0056] The activation gate vector is as follows:

[0057]

[0058] In the formula: This represents the context-enhanced feature vector of the p-th path. This represents the activation gate vector for the p-th path. This represents the aggregated feature vector obtained by the p-th path during the attention aggregation stage. For activation function, This represents the weight matrix in a gated network, used to project context features to the same dimension as the path-aggregated feature vector. It is a context feature vector. This represents the bias vector in the gated network. Represent a A real vector space.

[0059] S3 uses a graph neural network to embed the knowledge graph, and then inputs the graph embedding and feature tensor into the time series prediction model to construct a joint prediction network;

[0060] In this embodiment, a graph neural network is used to embed the knowledge graph, and the graph embedding and feature tensor are input into the time series prediction model to construct a joint prediction network, as follows:

[0061] Each entity node in the graph is initialized as a basic feature vector, which may include the node's historical running state, statistical features, node type, and context attributes.

[0062] Based on the graph neural network model, iterative propagation and embedding learning are performed on the knowledge graph to generate the embedding vector of each entity node;

[0063] Multiple rounds of iterative propagation and feature aggregation are performed on the embedding vectors to capture the topological dependencies and semantic relevance between nodes, thereby generating the final embedding representation of each entity node. The vectors are used as auxiliary modeling inputs in subsequent prediction models.

[0064] The embedding vectors are integrated by global average pooling to generate a global embedding vector, and then a linear transformation is performed on the global embedding vector to align it with the time dimension or path dimension of the feature tensor.

[0065] The aligned global embedding vector and the feature tensor are concatenated along the feature dimension to form a joint input tensor;

[0066] Based on the joint input tensor, a time series prediction model for transformer overload rate prediction is constructed.

[0067] In this embodiment, a time series prediction model for transformer overload rate prediction is constructed based on the joint input tensor, as follows:

[0068] The Transformer model is used to compress the joint input tensor of each path within the time window, extract the semantic relationships and feature associations between key paths, and generate path feature vectors.

[0069] The path feature vector is expanded according to the time series to form a path feature sequence at the time step level, and then input into the gated recurrent unit network for training. The evolution process of the load rate is dynamically modeled to obtain a time series prediction model.

[0070] The time series features of the output of the time series prediction model are input into a multilayer perceptron (MLP) to perform stepwise regression prediction on the reload rate of the target transformer in the next t time steps, and obtain the reload prediction value sequence.

[0071] The mean square error between the predicted heavy load rate sequence and the actual heavy load rate predicted sequence is used as the prediction loss function.

[0072] Based on the output value of the prediction loss function, the parameters in the time series prediction model are iteratively updated using the backpropagation algorithm with the goal of minimizing the prediction loss function, ultimately obtaining a robust and high-precision overload rate prediction model.

[0073] S4, based on causal reasoning methods, simulate intervention on input variables and assess the strength of the impact on the prediction results;

[0074] In this embodiment, the input variables are simulated and intervened based on a causal reasoning method, and the strength of the impact on the prediction results is evaluated, as follows:

[0075] Based on the knowledge graph, feature nodes with causal paths that can reach the target predictor variable (i.e., transformer overload rate) are identified. A path dependency matrix is ​​constructed by combining path scores and historical covariate changes. Variables with causal path scores higher than a set threshold are selected as a set of candidate intervention variables.

[0076] The frequency of each candidate variable in all causal paths is obtained, and the variables are weighted according to the importance of the path. Variables with scores higher than the threshold are selected to form the final set of intervention variables.

[0077] For each intervention variable, an intervention operation is set to generate an intervention sample input tensor. The intervention operation only modifies the value of the variable and its causal successor nodes, while keeping the values ​​of other unrelated nodes consistent, so as to ensure that the simulation experiment is locally consistent within the causal structure.

[0078] The feature tensors before and after the intervention are input into the time series prediction model, the prediction results of the transformer overload rate under the two inputs are recorded, and the difference between the prediction results is calculated as the influence strength value of the prediction result.

[0079] If the intervention variable appears in multiple paths, the influence intensity on each path is calculated separately, and these paths are weighted and averaged according to their importance to form a normalized influence intensity vector, which represents the influence intensity of each intervention variable on the prediction target.

[0080] The impact strength value of the prediction results under simulated intervention is corrected.

[0081] In this embodiment, the influence intensity value of the prediction result under simulated intervention is corrected as follows:

[0082] For each candidate intervention variable, a structural consistency function is constructed based on its causal path connectivity in the graph. This function measures whether there is an effective path for the variable to propagate to the target predictor variable. If a variable is only connected by weak edges or the path is broken, its influence will be reduced proportionally to obtain a preliminary adjusted influence value. ,in, Let represent the causal structure consistency coefficient of the i-th variable. This indicates the strength of the impact after structural modification. This indicates the influence strength of the predicted result calculated under the i-th variable;

[0083] For each variable, the generated feature tensors before and after intervention are used to perform multiple rounds of prediction using a neural network based on MC Dropout. The mean and standard deviation of the predicted output are obtained. If the intervention of a variable causes significant fluctuations in the model, it indicates that the variable does not have stable explanatory power in the current context, and its impact strength is weighted and compressed according to uncertainty. ,in, This represents the standard deviation of the prediction for the i-th variable after multiple rounds of prediction. This indicates the strength of the impact after the uncertainty perception correction;

[0084] For each candidate variable, multiple intervention perturbation magnitudes (e.g., +5%, -5%) are set, and corresponding counterfactual input samples are constructed;

[0085] Each group of samples is input into the same prediction model, the magnitude of change in the predicted output is calculated, and an intervention response curve is formed.

[0086] If the response curve of a variable exhibits high nonlinear fluctuations, it indicates that the local causal response is unstable. By averaging and smoothing the predicted influence intensity values ​​under multiple disturbances, a stable influence intensity, i.e., the local influence intensity value, can be obtained. ,in, Indicates the amplitude of the disturbance The strength of the intervention effect on the i-th variable. This indicates that the average of the results under multiple perturbations is taken. This represents the magnitude of the j-th set intervention disturbance, such as +5% or -5%.

[0087] Obtain the set of all causal paths involved in the graph for each variable, calculate the stability score for each path, and weight and fuse the local influence intensity value on each path with the corresponding stability score to finally calculate the normalized influence intensity value of the variable. The stability score is analyzed by the SHAP method to analyze the sensitivity of the predicted value to the input features, indirectly reflecting the local prediction residual, and the stability score is measured by the local prediction residual.

[0088] The specific normalized influence intensity values ​​are as follows:

[0089]

[0090] In the formula: Representing variables The final normalized influence strength is obtained by weighting multiple paths. Representing variables Local influence intensity value on path p It is the stability score of path p. Representing variables The set of all valid causal paths involved.

[0091] S5, based on the intensity of the impact of the simulated intervention, outputs the predicted value of the overload rate of the target transformer in the future time period, and generates the main causal path that leads to the change of the predicted value;

[0092] In this embodiment, based on the intensity of the simulated intervention, the predicted overload rate of the target transformer for future time periods is output, and the main causal path leading to the change in the predicted value is generated, as follows:

[0093] Based on the corrected variable influence intensity vector, the corresponding intervention variables in the joint input feature tensor are weighted and adjusted to generate a joint input tensor corrected by simulated intervention.

[0094] The modified joint input tensor is input into the trained time series prediction model, and the output is a sequence of predicted reload rates of the target transformer in the future time period.

[0095] Based on the knowledge graph, all reachable causal paths corresponding to each intervention variable are retrieved, and the stability score of each path is recorded.

[0096] Based on the stability scores of nodes in the path and the intensity of the impact of simulated intervention, the path is weighted and scored, and the contribution of each path to the prediction results is output.

[0097] We sorted the paths by contribution from highest to lowest and selected several key causal paths that had a significant impact on the change of the target predicted value.

[0098] For each key causal path, information such as the intervention magnitude, predicted response value changes, and path score of each node in the path is extracted to form a path-level causal impact explanation vector.

[0099] The predicted heavy load rate sequence of the target transformer in the future time period is mapped and integrated with the causal impact explanation vector of each critical path to establish a path mapping relationship from input intervention variables to changes in predicted values.

[0100] Based on the path mapping relationship, the predicted heavy load rate sequence of the target transformer in the future time period is output, and together with its key causal path set and corresponding causal influence explanation vector, a prediction result set with causal explanation capability is formed to assist in power regulation and decision optimization tasks.

[0101] The intervention variables in the joint input feature tensor are weighted and adjusted based on the modified variable influence intensity vector to generate a joint input tensor modified by simulated intervention, as follows:

[0102] By using the influence intensity vector as weights, the values ​​of the corresponding intervention variables in the joint input tensor are adjusted, thereby constructing a joint input tensor that reflects the multivariate causal intervention effect;

[0103] Subsequently, the adjusted joint input tensor is input into the trained time series prediction model to obtain the heavy load rate prediction sequence of the target transformer in multiple future time steps. This prediction result can comprehensively reflect the causal influence of each intervention variable and achieve accurate prediction of transformer load dynamics.

[0104] Example 2, Figure 2 The present invention provides a distribution transformer overload rate prediction system based on multi-source data fusion, including a data acquisition module, an extraction module, a construction module, an evaluation module, and an output module, with connections between the modules;

[0105] The data acquisition module is used to acquire transformer operation data in the target area and build a knowledge graph;

[0106] The extraction module is used to extract entities and features that have a causal path with the target transformer overload event from the knowledge graph and generate feature tensors;

[0107] The module is used to embed the knowledge graph into a graph neural network and input the graph embedding and feature tensor into a time series prediction model to build a joint prediction network.

[0108] The evaluation module is used to simulate interventions on input variables based on causal reasoning methods and assess the strength of the impact on the prediction results.

[0109] The output module is used to output the predicted value of the overload rate of the target transformer in the future time period based on the intensity of the impact of the simulated intervention, and to generate the main causal path that leads to the change in the predicted value.

[0110] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0111] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0112] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0113] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0114] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0115] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the overload rate of distribution transformers based on multi-source data fusion, characterized in that, include: Obtain transformer operation data for the target area and construct a knowledge graph; Extract entities and features that have a causal path with the target transformer overload event from the knowledge graph and generate a feature tensor; Knowledge graphs are embedded using graph neural networks, and the embedded graphs and feature tensors are input into a time series prediction model to construct a joint prediction network. The input variables are simulated and intervened based on causal reasoning methods, and the strength of their impact on the prediction results is assessed. Based on the intensity of the simulated intervention, the predicted value of the overload rate of the target transformer in the future time period is output, and the main causal path leading to the change of the predicted value is generated. The knowledge graph is embedded using a graph neural network, and the embedded graph and feature tensor are input into the time series prediction model to construct a joint prediction network, as detailed below: Based on the graph neural network model, iterative propagation and embedding learning are performed on the knowledge graph to generate the embedding vector of each entity node; Perform multiple rounds of iterative propagation and feature aggregation on the embedding vector to capture the topological dependencies and semantic relevance between nodes, and generate the final embedding representation of each entity node; The embedding vectors are integrated by global average pooling to generate a global embedding vector, and then a linear transformation is performed on the global embedding vector to align it with the time dimension of the feature tensor. The aligned global embedding vector and the feature tensor are concatenated along the feature dimension to form a joint input tensor; By using the Transformer model to extract the semantic relationships and feature associations of key paths from the joint input tensor of each path within the compressed time window, a path feature vector is generated. The path feature vectors are expanded according to the time series to form a path feature sequence at the time step level, and then input into a gated recurrent unit network for training to obtain a time series prediction model. The output of the time series prediction model is input into the multilayer perceptron to perform stepwise regression prediction on the reload rate of the target transformer in the next t time steps, and the reload prediction value sequence is obtained. The mean square error between the predicted heavy load rate sequence and the actual heavy load rate predicted sequence is used as the prediction loss function. Based on the output value of the prediction loss function, the parameters in the time series prediction model are iteratively updated using the backpropagation algorithm with the goal of minimizing the prediction loss function, thus obtaining the overload rate prediction model. Based on the intensity of the simulated intervention, the system outputs a predicted value for the overload rate of the target transformer over a future time period and generates the main causal path leading to the change in the predicted value, as follows: Based on the corrected variable influence intensity vector, the corresponding intervention variables in the joint input feature tensor are weighted and adjusted to generate the corrected joint input tensor; The modified joint input tensor is input into the heavy load rate prediction model, and the output is a sequence of predicted heavy load rate values ​​for future time periods. Based on the knowledge graph, all reachable causal paths corresponding to each intervention variable are retrieved, and the stability score of each path is recorded. Based on the stability scores of nodes in the path and the impact intensity of simulated interventions, the path is weighted and scored, the contribution of each path to the prediction results is output, and several key causal paths are selected by sorting by contribution. Extract the intervention magnitude, predicted response value changes, and path scores of each node in the key causal path to construct a path-level causal impact explanation vector; By integrating the predicted sequence with the critical path causal explanation vector mapping, a causal path relationship between intervention variables and predicted changes is established. Based on the path mapping relationship, the predicted heavy load rate sequence of the target transformer in the future time period is output, and together with its key causal path set and corresponding causal influence explanation vector, a prediction result set with causal explanation capability is formed.

2. The method for predicting the overload rate of distribution transformers based on multi-source data fusion according to claim 1, characterized in that, The process of acquiring transformer operation data for the target area and constructing a knowledge graph is as follows: Based on the business logic of the power distribution system and transformer operation data, a knowledge graph is constructed, and the core entity types and relationship types in the knowledge graph are determined. Based on entity type and relation type, extract structured triples from the preprocessed transformer operation data; For triples containing time-series variables, a causal discovery algorithm is used to identify potential causal relationships, output and filter causal edges with confidence scores higher than a threshold, and obtain modified triples. Construct a directed graph-like knowledge graph based on the modified triples; Obtain the prediction error between the output of the heavy load rate prediction model and the actual observation data, and identify key causal paths based on the prediction error; Based on key causal paths, the confidence level of corresponding edges is dynamically adjusted, and the knowledge graph is updated.

3. The method for predicting the overload rate of distribution transformers based on multi-source data fusion according to claim 2, characterized in that, The process of extracting entities and features from the knowledge graph that have a causal path to the target transformer overload event and generating a feature tensor is as follows: Obtain the target transformer number and prediction time interval in the prediction task, and retrieve relevant nodes and historical time series tags; Starting from the target transformer entity node, and following a pre-defined directed traversal strategy, search for causally guided paths within a range of no more than L hops to form an initial candidate path set. The path score is calculated based on the confidence of the causal edges on the path and the path depth. Paths with scores below the threshold are filtered out, and high-value paths are retained to obtain the final set of candidate paths. Extract the attribute features of entities involved in the candidate path within a specified time period to form a path feature vector; Align the path feature vectors within each path along the time dimension and concatenate them to generate a path feature matrix; The final feature tensor is generated based on the fusion matrix information of path score and causal confidence.

4. The method for predicting the overload rate of distribution transformers based on multi-source data fusion according to claim 3, characterized in that, The final feature tensor is generated based on the fusion matrix information of path scoring and causal confidence, as follows: Extract the temporal features of each node within the time window for each path, and arrange them in topological order to form a feature sequence matrix; By using path score and causal confidence as attention weights, a graph attention mechanism is applied to weighted aggregate of feature sequences to generate path aggregate features. Extract the temporal and regional contexts corresponding to the path, construct a context feature vector, and input it into a context gating function to generate an activation vector; The activation vector is multiplied element-wise with the aggregated feature vector to obtain the context-enhanced feature vector; Within each time step, context-enhanced features are fused with the original temporal structure, and average pooling is used to generate the final path features. The final path feature vectors of all time steps and paths are organized according to time order and path number to finally construct a three-dimensional structured feature tensor.

5. The method for predicting the overload rate of distribution transformers based on multi-source data fusion according to claim 4, characterized in that, The method of causal inference is used to simulate intervention on input variables and assess the strength of their impact on the prediction results, as detailed below: Based on the knowledge graph, feature nodes that have a causal path to the target predictor variable are identified. By combining the path score and historical covariate change data, a path dependency matrix is ​​constructed, and variables with causal path scores higher than a set threshold are selected as a set of candidate intervention variables. The frequency of each candidate variable in all causal paths is obtained, and the variables are weighted according to the importance of the path. Variables with scores higher than the threshold are selected to form the final set of intervention variables. For each intervention variable, define the intervention operation to generate the intervention sample input tensor; The feature tensors before and after the intervention are input into the time series prediction model, the prediction results of the transformer overload rate under the two inputs are recorded, and the difference between the prediction results is calculated as the influence strength value of the prediction result. If the intervention variable appears in multiple paths, the influence intensity on each path is calculated separately, and these paths are weighted and averaged according to their importance to form a normalized influence intensity vector. The impact strength value of the prediction results under simulated intervention is corrected.

6. The method for predicting the overload rate of distribution transformers based on multi-source data fusion according to claim 5, characterized in that, The correction of the influence intensity value of the prediction results under simulated intervention is as follows: For each candidate intervention variable, a structural consistency function is constructed based on its causal path connectivity in the graph. If a variable is only connected by weak edges and broken paths, its influence will be weakened proportionally to obtain the preliminary corrected influence value. Multiple rounds of predictions were performed on samples before and after the intervention using a neural network based on MC Dropout. If the intervention of variables caused large fluctuations in the model, the impact intensity was weighted and compressed according to the prediction uncertainty. For each candidate variable, multiple intervention perturbation magnitudes are set, and corresponding counterfactual input samples are constructed; Multiple intervention perturbation magnitudes are set for each variable to generate counterfactual samples and construct intervention response curves; If the response curve exhibits high nonlinear fluctuations, the influence intensity under multiple disturbances is averaged and smoothed to obtain the local influence intensity value. Obtain the set of all causal paths involved in each variable in the graph, calculate the stability score for each path, and weight and fuse the local influence intensity value on each path with the corresponding stability score to finally calculate the normalized influence intensity value of the variable.

7. A system using the multi-source data fusion-based method for predicting the overload rate of distribution transformers as described in any one of claims 1-6, characterized in that, It includes a data acquisition module, an extraction module, a construction module, an evaluation module, and an output module, and the modules are interconnected. The data acquisition module is used to acquire transformer operation data in the target area and build a knowledge graph; The extraction module is used to extract entities and features that have a causal path with the target transformer overload event from the knowledge graph and generate feature tensors; The module is used to embed the knowledge graph into a graph neural network and input the graph embedding and feature tensor into a time series prediction model to build a joint prediction network. The evaluation module is used to simulate interventions on input variables based on causal reasoning methods and assess the strength of the impact on the prediction results. The output module is used to output the predicted value of the overload rate of the target transformer in the future time period based on the intensity of the impact of the simulated intervention, and to generate the main causal path that leads to the change in the predicted value.

Images