Distribution transformer heavy load rate prediction method and system based on multi-source data fusion

Abstract

The invention discloses a distribution transformer heavy load rate prediction method and system based on multi-source data fusion, and relates to the technical field of load prediction, and the method comprises the following steps: obtaining transformer operation data of a target region, and constructing a knowledge graph; entities and features having causal paths with the heavy load event of the target transformer are extracted from the knowledge graph, and a feature tensor is generated; performing graph embedding on the knowledge graph through a graph neural network, and inputting the graph embedding and feature tensor into a time sequence prediction model to construct a joint prediction network; performing simulation intervention on the input variables based on a causal reasoning method, and evaluating the influence intensity on a prediction result; according to the influence intensity of simulation intervention, outputting a heavy load rate prediction value of the target transformer in a future time period, and generating a main causal path causing the prediction value to change; according to the method, accurate and explainable prediction is realized through multi-source data fusion and the causal knowledge graph.

Description

Technical Field

[0001] The present invention relates to the technical field of load forecasting, and more particularly to a method and system for predicting the heavy load rate of a distribution transformer based on multi-source data fusion. Background Art

[0002] With the continuous development of power systems and the advancement of smart grid technology, the safe and stable operation of distribution networks, as an important component of power systems, is directly related to the reliability of power supply and the user experience. Distribution transformers, as the core equipment of distribution networks, undertake the critical task of converting electricity from high-voltage to low-voltage. The health of their operating status directly affects the stability of the entire distribution network. In recent years, with the widespread application of distributed power sources, big data technologies, and the Internet of Things (IoT) technologies, 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 the announcement number CN118396193A discloses a transformer reverse load / overload warning method and system based on a neural network, which relates to the technical field of power systems. It includes: a data acquisition step, a feature extraction step, a load rate prediction feature set and a data association graph construction step, a neural network construction and training step, and a prediction step. After constructing a data association graph based on the empirical knowledge and Euclidean distance of the distribution transformer data, the present invention aggregates and transmits data of similar scenarios through the GraphSAGE network, mines and further enhances the implicit relationship between variables, thereby improving the model's warning accuracy for high-dimensional small sample data; using LSTM as the downstream prediction task of GraphSAGE can mine the time series relationship of the distribution transformer data, further improving the model's fitting ability.

[0004] For example, patent publication CN113298318A discloses a novel distribution transformer overload prediction method, comprising the following steps: Step S1: Obtaining basic data from production and management systems related to distribution and dispatch, as well as from a meteorological data network; Step S2: Performing correlation analysis on various feature quantities to identify those with strong correlations with distribution transformer overload as inputs to a prediction model; and Step S3: Designing a prediction model, training, and testing the model using a dataset. This invention proposes a prediction model based on a multi-scale one-dimensional CNN-LSTM network using an attention mechanism. This model combines the speed and lightweight nature of a convolutional neural network with the information processing capabilities of the three gates of an LSTM network. The attention mechanism introduced at the network output assigns attention weights to the outputs of the LSTM hidden layer, highlighting the influence of key factors on distribution transformer overload. Ultimately, this method enables prediction of distribution transformer overload, providing guidance for practical production operations.

[0005] The above disclosed technical solutions have at least the following technical problems: 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, and lack the comprehensive utilization of multi-source information, resulting in limited prediction accuracy. Furthermore, traditional feature extraction methods struggle to capture the structural dependencies and temporal dynamics between paths and lack effective integration of contextual information.

[0006] In view of the above problems, the present invention proposes a solution. Summary of the Invention

[0007] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides 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 a joint prediction model based on multi-source data fusion, the problems of data heterogeneity, complex causal relationships and insufficient prediction accuracy in the prediction of the overload rate of distribution transformers are solved, and high-precision, explainable and dynamically adjustable overload rate prediction is achieved, thereby improving the intelligent operation and maintenance capabilities of the distribution system.

[0008] To achieve the above object, the present invention provides the following technical solutions: A distribution transformer overload rate prediction method based on multi-source data fusion includes: obtaining transformer operation data in the target area and constructing a knowledge graph; extracting entities and features that have causal paths with the target transformer overload event from the knowledge graph to generate a feature tensor; embedding the knowledge graph through a graph neural network, and inputting the graph embedding and the feature tensor into a time series prediction model to construct a joint prediction network; simulating intervention on the input variables based on the causal reasoning method and evaluating the impact intensity on the prediction results; outputting the overload rate prediction value of the target transformer in the future time period based on the impact intensity of the simulated intervention, and generating the main causal path that leads to the change in the prediction value.

[0009] In a preferred embodiment, the transformer operation data of the target area is obtained and a knowledge graph is constructed, specifically as follows: a knowledge graph is constructed according to the business logic of the distribution system and the transformer operation data, and the core entity types and relationship types in the knowledge graph are determined; structured triples are extracted from the preprocessed transformer operation data according to the entity types and relationship types; for triples containing time series variables, a causal discovery algorithm is used to identify potential causal relationships, and causal edges with confidence levels higher than a threshold are output and screened to obtain corrected triples; a knowledge graph in the form of a directed graph is constructed based on the corrected triples; the prediction error between the output result of the overload rate prediction model and the actual observed data is obtained, and the key causal path is identified based on the prediction error; the confidence levels of the corresponding edges are dynamically adjusted based on the key causal paths, and the knowledge graph is updated.

[0010] In a preferred embodiment, the entities and features that have a causal path with the target transformer overload event are extracted from the knowledge graph to generate a feature tensor, specifically as follows: the target transformer number and prediction time interval in the prediction task are obtained, and the relevant nodes and historical time series labels are retrieved; starting from the target transformer entity node, according to a preset directed traversal strategy, a causal-oriented path is searched within a range of no more than L hops to form an initial candidate path set; the path score is calculated according to the causal edge confidence and path depth on the path, and paths with scores below a threshold are screened out, and high-value paths are retained to obtain the final candidate path set; the attribute features of the entities involved in the candidate path within a specified time period are extracted to form a path feature vector; the path feature vectors in each path are aligned in the time dimension, and spliced ​​to generate a path feature matrix; the matrix information is fused based on the path score and the causal confidence to generate the final feature tensor.

[0011] In a preferred embodiment, the matrix information based on path score and causal confidence is fused to generate the final feature tensor, specifically as follows: extract the node timing features of each path in the time window, and arrange them in topological order to form a feature sequence matrix; use the path score and causal confidence as attention weights, apply the graph attention mechanism to perform weighted aggregation on the feature sequence, and generate path aggregation features; extract the time context and regional context corresponding to the path, construct a context feature vector, and input the context gating function to generate an activation vector; multiply the activation vector and the aggregation feature vector element by element to obtain the context enhanced feature vector; in each time step, fuse the context enhanced features with the original time series 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 time sequence and path number, and finally construct a three-dimensional structured feature tensor.

[0012] In a preferred embodiment, the knowledge graph is embedded through a graph neural network, and the graph embedding and the feature tensor are input into the time series prediction model to construct a joint prediction network, specifically 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 vector to capture the topological dependencies and semantic correlations between nodes to generate the final embedding representation of each entity node; the embedding vectors are integrated through global average pooling to generate a global embedding vector, and the global embedding vector is linearly transformed to align it with the time dimension of the feature tensor; the aligned global embedding vector and the feature tensor are spliced ​​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.

[0013] In a preferred embodiment, a time series prediction model for transformer overload rate prediction is constructed based on the joint input tensor, specifically as follows: the Transformer model is used to compress the joint input tensor of each path in the time window, the semantic relationship and feature association of the key path are extracted, and a path feature vector is generated; the path feature vector is expanded according to the time series to form a path feature sequence at the time step level, and is input into the gated recurrent unit network for training to obtain a time series prediction model; the output result of the time series prediction model is input into a multi-layer perceptron, and the overload rate of the target transformer in the future t time steps is stepwise regressed to obtain an overload prediction value sequence; the mean square error between the overload rate prediction value sequence and the actual overload rate prediction value sequence is used as the prediction loss function; according to the output value of the prediction loss function, the parameters in the time series prediction model are iteratively updated through the back propagation algorithm with the goal of minimizing the prediction loss function.

[0014] In a preferred embodiment, the causal reasoning method is used to simulate intervention on input variables and evaluate the impact on the prediction results, specifically as follows: according to the knowledge graph, feature nodes with reachable causal paths with the target prediction variables are identified, and the path dependency matrix is ​​constructed by combining the path score and the historical covariate change data, and variables with causal path scores higher than the set threshold are screened out as a set of candidate intervention variables; the frequency of occurrence of each candidate variable in all causal paths is obtained, and the variables are weighted according to the path importance score, and variables with scores higher than the threshold are screened to form a final set of intervention variables; an intervention operation is set for each intervention variable to generate an intervention sample input tensor; the feature tensors before and after the intervention are respectively input into the time series prediction model, the transformer overload rate prediction results under the two inputs are recorded, and the difference between the prediction results is calculated as the impact intensity value of the prediction result; if the intervention variable appears in multiple paths, the impact intensity on each path is calculated separately, and these paths are weighted averaged according to the path importance weight to form a normalized impact intensity vector; the impact intensity value of the prediction result under the simulated intervention is corrected.

[0015] In a preferred embodiment, the impact intensity value of the predicted result under the simulated intervention is corrected 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 impact intensity will be weakened proportionally to obtain a preliminary corrected impact value; multiple rounds of predictions are performed on samples before and after the intervention through a neural network based on MC Dropout. If the intervention operation of the variable causes large model fluctuations, the impact intensity is weighted and compressed according to the prediction uncertainty; for each candidate variable, multiple intervention disturbance amplitudes are set, and corresponding counterfactual input samples are constructed; multiple intervention disturbance amplitudes are set for each variable, counterfactual samples are generated, and an intervention response curve is constructed; if there is high nonlinear fluctuation in the response curve, the impact intensity under multiple disturbances is averaged and smoothed to obtain a local impact intensity value; the set of all causal paths involved in each variable in the graph is obtained, the stability score of each path is calculated, the local impact intensity value on each path is weightedly fused with the corresponding stability score, and finally the normalized impact intensity value of the variable is calculated.

[0016] In a preferred embodiment, the predicted value of the overload rate of the target transformer in the future time period is output according to the impact intensity of the simulated intervention, and the main causal path that causes the change of the predicted value is generated, specifically as follows: based on the modified variable impact intensity vector, the corresponding intervention variable in the joint input feature tensor is weighted and adjusted to generate a modified joint input tensor; the modified joint input tensor is input into the trained time series prediction model to output the sequence of predicted values ​​of the overload rate in the future time period; 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; according to the stability score of the node in the path The impact intensity of the simulated intervention is calculated, and the paths are weighted and scored. The contribution of each path to the prediction result is output, and the paths are sorted by contribution to screen several key causal paths. The intervention amplitude, predicted response value change and path score of each node in the key causal path are extracted to form a path-level causal impact explanatory vector. The prediction sequence is mapped and integrated with the key path causal explanatory vector to establish a causal path relationship between the intervention variable and the predicted change. According to the path mapping relationship, the predicted value sequence of the overload rate of the target transformer in the future time period is output, and together with its key causal path set and the corresponding causal impact explanatory vector, a prediction result set with causal explanation ability is formed.

[0017] A 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, and there are connections between the modules; the data acquisition module is used to obtain the transformer operation data of the target area and construct a knowledge graph; the extraction module is used to extract entities and features that have causal paths with the target transformer overload event from the knowledge graph and generate a feature tensor; the construction module is used to embed the knowledge graph through a graph neural network, and input the graph embedding and the feature tensor into a time series prediction model to construct a joint prediction network; the evaluation module is used to simulate intervention on the input variables based on the causal reasoning method and evaluate the impact intensity on the prediction result; the output module is used to output the overload rate prediction value of the target transformer in the future time period according to the impact intensity of the simulated intervention, and generate the main causal path that causes the change in the prediction value.

[0018] The technical effects and advantages of the present invention's method and system for predicting the overload rate of distribution transformers based on multi-source data fusion are as follows: 1. This invention integrates multi-source heterogeneous data to construct a structured knowledge graph, organically integrating the operating data, business logic, time series characteristics, and inherent causal relationships of distribution transformers to form a comprehensive and dynamic multidimensional data representation system. Using a causal discovery algorithm to filter causal edges and modify the knowledge graph for time series triples, this not only improves data accuracy and consistency but also ensures that model predictions are based on real, high-confidence causal relationships, making the prediction results more interpretable and trustworthy.

[0019] 2. This invention deeply integrates graph neural networks with time series prediction models to construct a joint prediction network. The graph neural network uses multiple rounds of iterative propagation and feature aggregation on the knowledge graph to accurately mine topological dependencies and semantic relevance between nodes, generating high-quality entity embedding vectors. Subsequently, through the synergistic effect of the Transformer and Gated Recurrent Unit (GRU) models, the dynamic time series features between paths are deeply learned, achieving multi-step, accurate prediction of transformer overload rates. This joint model significantly improves prediction accuracy and robustness compared to traditional single models. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 The figure is a flow chart of the method for predicting the overload rate of distribution transformers based on multi-source data fusion according to the present invention.

[0021] Figure 2 The figure is a schematic diagram of the structure of the distribution transformer overload rate prediction system based on multi-source data fusion of the present invention. DETAILED DESCRIPTION

[0022] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0023] Example 1, Figure 1 The present invention provides a method for predicting the overload rate of distribution transformers based on multi-source data fusion, which includes the following steps: S1, obtains transformer operation data in the target area and builds a knowledge graph; In this embodiment, transformer operation data of the target area is obtained and a knowledge graph is constructed, as follows: Based on the business logic of the 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. The entities include but are not limited to distribution transformers, user groups, climate environment, geographical regions, and time tags. The relationship types include belonging to, depending on, affecting, and causing, which are used to represent the structural connection or causal dependency between entities. According to the entity type and relationship type, structured triples are extracted from the preprocessed transformer operation data. For structured data, a rule-based mapping method is adopted. For unstructured text data, a named entity recognition and relationship extraction model based on a pre-trained language model is preferably used to extract potential triples. For triplets containing time series variables, a causal discovery algorithm (such as the Granger causality test) is used to identify potential causal relationships and output a confidence score for each causal edge. Based on the confidence score, causal relationships with confidence levels below a set threshold are screened out, improving the accuracy of the knowledge graph and the reliability of reasoning, resulting in a revised triplet. A knowledge graph with a directed graph structure is constructed based on the modified 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 sources. Obtain the prediction error between the output of the overload rate prediction model and the actual observed data, and identify the causal chain that affects the prediction result, namely the critical causal path, based on the prediction error; According to 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 graph update mechanism.

[0024] The confidence scoring formula for each causal edge is as follows:

[0025] The formula for the lag order is as follows:

[0026] Where: represents the confidence score of the causal edge X→Y, is the natural base, is the adjustment parameter for causal confidence, is the F statistic, is the lag order obtained by the F test, which is used to determine whether the null hypothesis is rejected or not. is the significance level, which is usually set at 0.05. is the residual sum of squares without regression terms, is the residual sum of squares with regression terms, is the sample size, where and It is directly obtained through regression analysis of time series data.

[0027] S2, extracts entities and features that have causal paths with the target transformer overload event from the knowledge graph and generates feature tensors; In this embodiment, entities and features that have a causal path with the target transformer overload event are extracted from the knowledge graph to generate a feature tensor, as follows: Obtain the target transformer number required for prediction in the current prediction task, determine the corresponding prediction time interval, and retrieve the nodes and historical time series labels related to the target transformer from the knowledge graph based on the prediction time interval; Starting from the target transformer entity node, a preset directed traversal strategy (such as breadth-first, depth-first, or weighted random walk) is used to search for a set of paths that are causally oriented to the target transformer overload event in the current knowledge graph, and the causal paths whose path length does not exceed the preset number of hops L are retained to obtain the initial candidate path set; For each path, output the path score based on the confidence score of the causal edge on the path and the path depth; Sort the candidate paths by score from high to low, remove paths with scores below the set threshold, retain high-value causal paths for feature extraction, and obtain the final set of candidate paths; Extract attribute feature values ​​of the entity nodes involved in each path within a fixed time period according to the candidate path set to form a path feature vector. The attribute feature values ​​include temperature, current, load, operating status, etc. Align the entity feature vectors in each path in the time dimension and concatenate them to generate a path feature matrix; For each row in the path feature matrix, the score of the corresponding path and the causal confidence score of the edge in the path are fused to generate a feature tensor.

[0028] In this embodiment, each row in the path feature matrix is ​​fused according to the score of the corresponding path and the causal confidence score of the edge in the path to generate a feature tensor, as follows: For each causal path, extract its time series features within the time window and arrange the nodes according to the topological order of the path to 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. The path score and the causal confidence of each edge in the path are used as the weight input of the path attention, and the path feature sequence matrix is ​​weighted aggregated through the graph attention mechanism to generate the aggregated feature vector of the path; Extract the time context (such as hourly period, holiday, season) and regional context (such as urban / rural, voltage level, user density) corresponding to the path, construct a context feature vector, and input it into the context gating function to generate an activation vector. Multiply the activation vector and the aggregated feature vector element by element to obtain the enhanced feature vector after context enhancement; For all paths in each time step, their context-enhanced feature vectors are collected. Combined with the temporal structure in the path feature sequence matrix, the context-enhanced features are concatenated with the global structural features through the average pooling method to form the final path feature vector. The final path feature vectors of all time steps and paths are organized in time order and path number, and finally a three-dimensional structured feature tensor is constructed. ,in, For the time dimension, is the number of paths, is the feature dimension.

[0029] The activation gate vector is specifically as follows:

[0030] Where: represents the context-enhanced feature vector of the p-th path, represents the activation gating vector of the p-th path, represents the aggregated feature vector obtained by the p-th path in the attention aggregation stage, is the activation function, represents the weight matrix in the gating network, which is used to project the context features to the same dimension as the path aggregation feature vector. is the context feature vector, represents the bias vector in the gating network, Indicates a dimensional real vector space.

[0031] S3, embeds the knowledge graph through a graph neural network, and inputs the graph embedding and feature tensor into the time series prediction model to build a joint prediction network; 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 build a joint prediction network, as follows: Initialize each entity node in the graph into a basic feature vector, which may include the historical operating status, statistical characteristics, node type and context attributes corresponding to the node; Perform iterative propagation and embedding learning on the knowledge graph based on the graph neural network model to generate an embedding vector for 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, thereby generating the final embedding representation of each entity node. The vector is used as auxiliary modeling input in the subsequent prediction model; The embedding vectors are integrated through global average pooling to generate a global embedding vector, and the global embedding vector is linearly transformed to align it with the time dimension or path dimension of the feature tensor; Concatenate the aligned global embedding vector and the feature tensor 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.

[0032] In this embodiment, a time series prediction model for transformer overload rate prediction is constructed based on the joint input tensor, as follows: The Transformer model is used to compress the joint input tensor of each path within the time window, extract the semantic relationship and feature association between key paths, and generate a path feature vector; The path feature vector is expanded in time series to form a time-step path feature sequence, which is then input into a gated recurrent unit network for training. The overload rate evolution process is dynamically modeled to obtain a time series prediction model. The time series features output by the time series prediction model are input into a multi-layer perceptron (MLP) to perform stepwise regression prediction on the overload rate of the target transformer in the next t time steps to obtain a sequence of overload prediction values. The mean square error between the predicted value sequence of the overload rate and the actual predicted value sequence of the overload rate is used as the prediction loss function; According to the output value of the prediction loss function, the parameters in the time series prediction model are adjusted through the back propagation algorithm with the goal of minimizing the prediction loss function. The model is iteratively updated to finally obtain a robust and high-precision overload rate prediction model.

[0033] S4, simulates interventions on input variables based on causal inference methods and evaluates the intensity of the impact on the predicted results; In this embodiment, a causal inference method is used to simulate interventions on input variables and evaluate the impact on the predicted results, as follows: Based on the knowledge graph, we identify feature nodes that have a reachable causal path with the target predictor variable (i.e., transformer overload rate). We then build a path dependency matrix by combining the path score with historical covariate changes, and select variables with causal path scores higher than a set threshold as candidate intervention variables. Obtain the frequency of occurrence of each candidate variable in all causal paths, assign variable weights based on the path importance score, and screen variables with scores above the threshold to form the final set of intervention variables; 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 node, while the other non-related nodes remain consistent to ensure that the simulation experiment is locally consistent within the causal structure. The feature tensors before and after the intervention are input into the time series prediction model respectively, the transformer overload rate prediction results under the two inputs are recorded, and the difference between the prediction results is calculated as the impact intensity value of the prediction result; If the intervention variable appears in multiple paths, the impact intensity on each path is calculated separately, and these paths are weighted averaged according to the path importance weight to form a normalized impact intensity vector, which represents the impact intensity of each intervention variable on the predicted target; The impact intensity values ​​of the predicted results under simulated interventions are corrected.

[0034] In this embodiment, the impact intensity value of the predicted result under the simulated intervention is corrected as follows: For each candidate intervention variable, a structural consistency function is constructed based on its causal path connectivity in the graph to measure whether the variable has an effective transmission path to the target predictor variable. If a variable is only connected by a weak edge or the path is broken, its influence strength will be weakened proportionally to obtain a preliminary revised influence value. ,in, represents the causal structure consistency coefficient of the i-th variable, represents the impact strength after structural modification, Indicates the impact intensity of the prediction result calculated under the i-th variable; For each variable, the feature tensor generated before and after the intervention is used to perform multiple rounds of prediction using a neural network based on MC Dropout to obtain the mean and standard deviation of the prediction output. If the intervention operation of the variable causes significant model fluctuations, it means that the variable does not have stable explanatory power in the current context, and its influence strength is weighted and compressed according to the uncertainty. ,in, It represents the prediction standard deviation of the i-th variable after multiple rounds of prediction, represents the intensity of the effect after the correction of uncertainty perception; For each candidate variable, set multiple intervention disturbance amplitudes (such as +5%, –5%) and construct corresponding counterfactual input samples; Each group of samples is input into the same prediction model, and the magnitude of change in the predicted output is calculated to form an intervention response curve; If the response curve of the variable shows high nonlinear fluctuation, it means that the local causal response is unstable. By averaging and smoothing the predicted impact intensity values ​​under multiple disturbances, a stable impact intensity, that is, the local impact intensity value, is obtained. ,in, Indicates the disturbance amplitude The intervention impact intensity of the i-th variable, It means taking the average of the results under multiple groups of disturbances. represents the j-th set intervention disturbance amplitude, such as +5%, −5%; Obtain the set of all causal paths involved in each variable in the graph, calculate the stability score of each path, perform weighted fusion on the local impact strength value of each path and the corresponding stability score, and finally calculate the normalized impact strength value of the variable. The stability score analyzes the sensitivity of the predicted value to the input features through the SHAP method, indirectly reflects the local prediction residual, and measures the stability score with the local prediction residual.

[0035] The normalized impact intensity values ​​are as follows:

[0036] Where: Representing variables The final normalized impact intensity after integrating multiple paths weighted, Representing variables The local influence strength value on path p, is the stability score of path p, Representing variables The set of all valid causal paths involved.

[0037] S5, based on the impact intensity 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 leading to the change of the predicted value; In this embodiment, based on the impact 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, as follows: Based on the modified variable influence intensity vector, the corresponding intervention variable in the joint input feature tensor is weighted and adjusted to generate the joint input tensor modified by the simulated intervention; Input the modified joint input tensor into the trained time series prediction model, and output the predicted value sequence of the target transformer's overload rate in the future time period; Based on the knowledge graph, retrieve all reachable causal paths corresponding to each intervention variable and record the stability score of each path; According to the stability score of the nodes in the path and the impact intensity of the simulated intervention, the paths are weighted and scored, and the contribution of each path to the prediction results is output; Sort by contribution from high to low, and select several key causal paths that have a significant impact on the change of the target predicted value; For each key causal path, we extract information such as the intervention magnitude, predicted response value change, and path score of each node in the path to form a path-level causal impact explanation vector; The predicted value sequence of the overload rate of the target transformer in the future time period is mapped and integrated with the causal impact explanation vector of each key path to establish a path mapping relationship from the input intervention variable to the predicted value change; According to the path mapping relationship, a sequence of predicted values ​​of the target transformer's overload rate in the future time period is output, and together with its key causal path set and the corresponding causal impact explanation vector, a prediction result set with causal explanation capability is constructed to assist power regulation and decision optimization tasks.

[0038] Based on the modified variable influence intensity vector, the corresponding intervention variable in the joint input feature tensor is weighted and adjusted to generate the joint input tensor modified by the simulated intervention, as follows: Using the impact intensity vector as a weight, the values ​​of the corresponding intervention variables in the joint input tensor are adjusted to construct a joint input tensor that reflects the causal intervention effects of multiple variables. Subsequently, the adjusted joint input tensor is input into the trained time series prediction model to obtain the target transformer's overload rate prediction sequence in multiple future time steps. The prediction result can comprehensively reflect the causal influence of each intervention variable and achieve accurate prediction of the transformer load dynamics.

[0039] Example 2, Figure 2 The invention provides a distribution transformer overload rate prediction system based on multi-source data fusion, which includes a data acquisition module, an extraction module, a construction module, an evaluation module and an output module, and there are connections between the modules; The data acquisition module is used to obtain 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; A construction module is used to embed the knowledge graph through a graph neural network, and input the graph embedding and feature tensor into the time series prediction model to build a joint prediction network; Evaluation module, used to simulate interventions on input variables based on causal reasoning methods and evaluate the intensity of the impact on the predicted results; The output module is used to output the predicted value of the heavy load rate of the target transformer in the future time period according to the impact intensity of the simulated intervention, and generate the main causal path leading to the change of the predicted value.

[0040] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0041] The above embodiments may be implemented in whole or in part through software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product.

[0042] Those skilled in the art will appreciate that the modules and algorithm steps of each example 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 performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel 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.

[0043] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0044] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0045] Finally: 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 in the scope of protection of the present invention.

Claims

1. A distribution transformer overload rate prediction method based on multi-source data fusion, characterized in that: include: Obtain transformer operation data in the target area and build a knowledge graph; Extract entities and features that have causal paths with the target transformer overload event from the knowledge graph and generate feature tensors; The knowledge graph is embedded through a graph neural network, and the graph embedding and feature tensor are input into the time series prediction model to build a joint prediction network; Simulate interventions on input variables based on causal inference methods and assess the magnitude of the impact on predicted outcomes; According to the impact 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.

2. The method for predicting the overload rate of distribution transformers based on multi-source data fusion according to claim 1 is characterized in that: The transformer operation data of the target area is obtained and the knowledge graph is constructed 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; Extract structured triples from preprocessed transformer operation data according to entity types and relationship types; For triples containing time series variables, a causal discovery algorithm is used to identify potential causal relationships, and causal edges with confidence levels higher than a threshold are output and filtered to obtain modified triples. Construct a knowledge graph in the form of a directed graph based on the modified triples; Obtain the prediction error between the output of the overload rate prediction model and the actual observed data, and identify the key causal path based on the prediction error; According to the key causal path, the confidence of the corresponding edge 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 is characterized in that: The entities and features that have a causal path with the target transformer overload event are extracted from the knowledge graph to generate a feature tensor, as follows: Obtain the target transformer number and prediction time interval in the prediction task, and retrieve related nodes and historical time series labels; Starting from the target transformer entity node, according to the preset directed traversal strategy, a causal-oriented path is searched within the range of hop number not exceeding L to form an initial candidate path set; Calculate the path score based on the confidence of the causal edge and the path depth on the path, filter out paths with scores below the threshold, retain high-value paths, and obtain the final set of candidate paths; Extract the attribute features of the entities involved in the candidate path within the specified time period to form a path feature vector; Align the path feature vectors in each path in the time dimension and concatenate them to generate a path feature matrix; The final feature tensor is generated based on the matrix information of path score and causal confidence fusion.

4. The method for predicting the overload rate of distribution transformers based on multi-source data fusion according to claim 3 is characterized in that: The path score and causal confidence fusion matrix information are used to generate the final feature tensor, as follows: Extract the node timing features of each path within the time window and arrange them in topological order to form a feature sequence matrix; Taking the path score and causal confidence as attention weights, the graph attention mechanism is applied to perform weighted aggregation on the feature sequence to generate path aggregation features; Extract the temporal context and regional context corresponding to the path, construct a context feature vector, and input it into the context gating function to generate an activation vector; Multiply the activation vector and the aggregated feature vector element-wise to obtain the context-enhanced feature vector; In each time step, the 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 sequence and path number, and finally a three-dimensional structured feature tensor is constructed.

5. The method for predicting the overload rate of distribution transformers based on multi-source data fusion according to claim 4 is characterized in that: The knowledge graph is embedded in the graph neural network, and the graph embedding and feature tensor are input into the time series prediction model to build a joint prediction network, as follows: Perform iterative propagation and embedding learning on the knowledge graph based on the graph neural network model to generate an embedding vector for 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 through global average pooling to generate a global embedding vector, and the global embedding vector is linearly transformed to align it with the time dimension of the feature tensor; Concatenate the aligned global embedding vector and the feature tensor 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.

6. The method for predicting the overload rate of distribution transformers based on multi-source data fusion according to claim 5 is characterized in that: The time series prediction model for transformer overload rate prediction is constructed based on the joint input tensor, as follows: The Transformer model is used to extract the semantic relationship and feature association of key paths from the joint input tensor of each path in the compressed time window, and generate a path feature vector; The path feature vector is expanded according to the time series to form a path feature sequence at the time step level, and is input into the gated recurrent unit network for training to obtain a time series prediction model; The output results of the time series prediction model are input into the multi-layer perceptron, and the overload rate of the target transformer in the next t time steps is predicted by stepwise regression to obtain the overload prediction value sequence; The mean square error between the predicted value sequence of the overload rate and the actual predicted value sequence of the overload rate is used as the prediction loss function; According to the output value of the prediction loss function, the parameters in the time series prediction model are iteratively updated through the back propagation algorithm with the goal of minimizing the prediction loss function.

7. The method for predicting the overload rate of distribution transformers based on multi-source data fusion according to claim 6 is characterized in that: The causal inference method is used to simulate interventions on input variables and evaluate the impact on the predicted results, as follows: Based on the knowledge graph, we identify characteristic nodes that have a reachable causal path with the target predictor variable. We then combine the path score with historical covariate change data to construct a path dependency matrix and select variables with causal path scores higher than the set threshold as candidate intervention variables. Obtain the frequency of occurrence of each candidate variable in all causal paths, assign variable weights based on the path importance score, and screen variables with scores above the threshold to form the final set of intervention variables; Set the intervention operation for each intervention variable to generate the intervention sample input tensor; The feature tensors before and after the intervention are input into the time series prediction model respectively, the transformer overload rate prediction results under the two inputs are recorded, and the difference between the prediction results is calculated as the impact intensity value of the prediction result; If the intervention variable appears in multiple paths, the impact intensity on each path is calculated separately, and these paths are weighted averaged according to the path importance weight to form a normalized impact intensity vector; The impact intensity values ​​of the predicted results under simulated interventions are corrected.

8. The method for predicting the heavy load rate of distribution transformers based on multi-source data fusion according to claim 7 is characterized in that: The impact intensity value of the predicted results under the simulated intervention is corrected 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 has broken paths, its impact strength will be weakened proportionally to obtain a preliminary revised impact value. Multiple rounds of predictions are performed on samples before and after the intervention using a neural network based on MC Dropout. If the intervention operation of the variable causes large model fluctuations, the impact intensity is weighted and compressed according to the prediction uncertainty. For each candidate variable, set multiple intervention disturbance amplitudes and construct corresponding counterfactual input samples; Set multiple intervention disturbance amplitudes for each variable, generate counterfactual samples, and construct intervention response curves; If the response curve has high nonlinear fluctuations, the impact intensity under multiple perturbations is averaged and smoothed to obtain the local impact intensity value; Obtain the set of all causal paths involved in each variable in the graph, calculate the stability score of each path, perform weighted fusion of the local impact intensity value on each path and the corresponding stability score, and finally calculate the normalized impact intensity value of the variable.

9. The method for predicting the heavy load rate of distribution transformers based on multi-source data fusion according to claim 8, characterized in that: According to the impact 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, as follows: Based on the modified variable influence intensity vector, the corresponding intervention variable in the joint input feature tensor is weighted and adjusted to generate a modified joint input tensor; Input the modified joint input tensor into the trained time series prediction model and output a sequence of predicted values ​​of the overload rate in the future time period; Based on the knowledge graph, retrieve all reachable causal paths corresponding to each intervention variable and record the stability score of each path; Based on the stability scores of the nodes in the path and the impact intensity of the simulated intervention, the paths are weighted and scored, and the contribution of each path to the prediction results is output. The paths are sorted by contribution and several key causal paths are screened. Extract the intervention magnitude, predicted response value change, and path score of each node in the key causal path to form a path-level causal impact explanation vector; Integrate the predicted sequence with the key path causal explanation vector mapping to establish the causal path relationship between the intervention variable and the predicted change; According to 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 the corresponding causal impact explanation vector, a prediction result set with causal explanation ability is formed.

10. A system using the method for predicting the overload rate of a distribution transformer based on multi-source data fusion according to any one of claims 1 to 9, characterized in that: It includes data acquisition module, extraction module, construction module, evaluation module and output module, and there are connections between modules; The data acquisition module is used to obtain 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; A construction module is used to embed the knowledge graph through a graph neural network, and input the graph embedding and feature tensor into the time series prediction model to build a joint prediction network; Evaluation module, used to simulate interventions on input variables based on causal reasoning methods and evaluate the intensity of the impact on the predicted results; The output module is used to output the predicted value of the heavy load rate of the target transformer in the future time period according to the impact intensity of the simulated intervention, and generate the main causal path leading to the change of the predicted value.

Images