A power distribution network source-load prediction method, system, device and storage medium based on multi-modal data

Abstract

The application discloses a power distribution network source and load prediction method, system and device based on multi-modal data and a storage medium, relates to the technical field of power dispatching, and comprises the following steps: acquiring multi-source data of a power distribution network, constructing a graph window sequence, merging state extraction and node state coding to obtain unified node state coding; a historical event trajectory database is introduced, similarity matching is performed to generate a candidate set, a path selection weight is calculated in combination with a weight matching mechanism, behavior prior vectors are generated in combination with behavior response information, and the node state coding is spliced and input into a feedforward neural network to output a prediction result; by comparing the behavior prior vectors with the prediction result, an abnormal node is identified, and a preset strategy is executed; an actual observation and a prediction value error are calculated and fed back to the database, and graph node attributes are synchronously updated; the application can realize multi-scale high-precision prediction and situation deduction of power distribution network source and load, enhance non-stationary scene modeling, fuse multi-modal data, and improve complex scene understanding and power distribution network management effect.

Description

Technical Field

[0001] This invention relates to the field of power dispatching technology, and in particular to a method, system, device and storage medium for predicting power distribution network source and load based on multimodal data. Background Technology

[0002] With the large-scale integration of distributed renewable energy sources, the rapid growth of electric vehicle users, and the accelerated construction of active distribution networks, power systems are exhibiting greater complexity and dynamism. Traditional load forecasting models relying on centralized generation and rule-based dispatching can no longer meet the high-precision and high-timeliness requirements of distribution network operation and management. In active distribution networks, the behavior of source-side (such as photovoltaic and wind power), load-side (residential loads and electric vehicles), and energy storage systems is highly uncertain and frequently affected by external environments (such as weather and traffic), resulting in multi-timescale, multi-source heterogeneous, and multi-type interactive coupling operation characteristics. Especially at the medium- and low-voltage distribution network level, due to its wide coverage, numerous access points, fine-grained regulation, and flexible topology (even with the ability to frequently reconfigure), unprecedented challenges are posed to forecasting models in terms of data adaptability, structural flexibility, and spatiotemporal representation capabilities.

[0003] Current mainstream source-load forecasting methods are mostly based on time series modeling, typically using deep learning methods such as LSTM, GRU, and CNN-LSTM, or using graph neural networks (GNNs) to handle the spatial relationship structure of the power grid in topology modeling. While these methods can achieve good performance under certain conditions, several key problems remain. First, most existing methods only process single-modal input data, such as relying solely on historical power sequences or meteorological parameters. They cannot integrate unstructured multimodal data such as text scheduling records, user behavior events, and image information generated during power grid operation into the model, resulting in poor model representation and understanding of complex scenarios. Second, the source-load fluctuations in distribution networks exhibit significant multi-timescale characteristics. For example, the charging and discharging behavior of electric vehicles fluctuates at the minute level, conventional loads change at the hour level, and new energy sources are affected by meteorological conditions at the daily / weekly level. Single time series modeling methods struggle to effectively characterize dynamic changes at different time granularities within a unified framework. Secondly, the operating structure of the distribution network is not static; it may undergo topology reconfiguration due to planned scheduling or sudden faults. This structural evolution information is currently difficult to effectively encode in prediction models, leading to significant inaccuracies in predictions when topology disturbances occur. Furthermore, most existing prediction models are "black boxes," lacking interpretability and causal analysis capabilities. Especially during extreme weather events or large-scale abnormal load fluctuations, dispatchers cannot make reasonable judgments and extrapolations based on the results provided by the prediction models. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is that the existing technology cannot simulate the structural change path of the power grid during the time evolution process.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for predicting the source and load of a distribution network based on multimodal data, comprising:

[0008] Acquire multi-source data from the power distribution network, map the multi-source data to the feature vectors of the nodes, construct an edge set, and generate a graph structure;

[0009] Based on the graph structure and node feature vectors, a graph window sequence is constructed. The graph window sequence is then subjected to state extraction and node state encoding merging to obtain a unified node state encoding.

[0010] By introducing a historical event trajectory database, the node status code is matched with the historical event trajectory database to generate a candidate set. Combined with a weighted matching mechanism, the path selection weight is calculated.

[0011] Based on path selection weights and behavioral response information from historical event trajectory databases, behavioral prior vectors are generated at all time scales. The node state encoding is concatenated with the behavioral prior vectors at the corresponding scales and then input into the feedforward neural network to output the prediction results.

[0012] By comparing the prior vector of behavior with the prediction result, nodes with differences exceeding a preset threshold are identified, and preset strategies are executed according to the abnormal node type. The error between the actual observation value and the prediction value is calculated and fed back to the historical event trajectory database, and the node attributes in the graph structure are updated synchronously.

[0013] As a preferred option for a distribution network source-load prediction method based on multimodal data, wherein:

[0014] The process of acquiring multi-source data from the distribution network, mapping the multi-source data to feature vectors of nodes, constructing an edge set, and generating a graph structure includes:

[0015] A multimodal distribution network state map is constructed based on multi-source data of the distribution network. The map structure includes a set of nodes, a set of edges, and a set of feature vectors of the nodes.

[0016] Multi-source data from the power distribution network is mapped to different types of nodes, and each type of node adopts a different input structure based on its data characteristics. Through a unified embedding strategy, the original input data of all nodes are mapped to a feature space of a unified dimension to obtain the feature vector of the node.

[0017] The beneficial effects of this preferred technical solution are as follows: the constructed graph structure can effectively integrate multi-source data of the distribution network; different types of nodes adopt different input structures to better adapt to the characteristics of multimodal data; the unified embedding strategy makes the node feature vectors in a unified dimensional space, which is convenient for subsequent analysis and processing, and improves the ability to understand and utilize complex multimodal data.

[0018] As a preferred option for a distribution network source-load prediction method based on multimodal data, wherein:

[0019] The process involves introducing a historical event trajectory database, matching node status codes with the database for similarity to generate a candidate set, and then calculating path selection weights using a weighted matching mechanism.

[0020] Select a matching metric and perform similarity matching between the current node status code and the node status vector in the historical event trajectory database; based on the similarity ranking, select a preset number of similar historical samples to form a candidate set.

[0021] As a preferred option for a distribution network source-load prediction method based on multimodal data, wherein:

[0022] The process involves constructing a graph window sequence based on the graph structure and node feature vectors, then extracting the state of the graph window sequence and merging the node state codes to obtain a unified node state code, including:

[0023] Based on different preset time scales, the feature vectors of the graph structure and nodes are grouped to establish a graph window sequence;

[0024] For each timescale of the graph window sequence, a time-series graph convolutional network with the same structure and shared parameters is used to extract the state, thereby obtaining the node state encoding at different timescales.

[0025] By introducing a node category-sensitive aggregation operation, node state codes at different time scales are merged to obtain a unified node state code.

[0026] As a preferred option for a distribution network source-load prediction method based on multimodal data, wherein:

[0027] The process of introducing a historical event trajectory database, matching node status codes with the historical event trajectory database to generate a candidate set, and calculating path selection weights using a weighted matching mechanism also includes:

[0028] A weighted matching mechanism is introduced to align the current topological trend with the structural evolution paths in the historical trajectory database. This measures the similarity between the current topological trend and the structural evolution chains in the historical paths. By combining the state similarity matching results and the topological alignment results, the path selection weight is calculated.

[0029] The beneficial effects of this preferred technical solution are as follows: similarity matching can filter samples similar to the current state from historical data, providing a reference for prediction; the weight matching mechanism comprehensively considers state similarity and topological changes, making the path selection weight more reasonable and improving the accuracy and reliability of prediction.

[0030] As a preferred option for a distribution network source-load prediction method based on multimodal data, wherein:

[0031] Based on path selection weights and behavioral response information from the historical event trajectory database, behavioral prior vectors are generated at all time scales. The node state encoding is concatenated with the corresponding scale's behavioral prior vector and input into the feedforward neural network. The output prediction results include:

[0032] Based on path selection weights and behavioral response information from historical event trajectory databases, trajectory behavior prior vectors for structural behavior fusion are generated at all preset time scales.

[0033] The node state code at each time scale is concatenated with the behavior prior vector at the corresponding scale to form the input data for prediction.

[0034] The concatenated data is input into a lightweight feedforward neural network for prediction, and the prediction results of each node are output at each preset time scale, finally resulting in a set of prediction results across multiple time scales.

[0035] The beneficial effects of this preferred technical solution are as follows: the generated behavioral prior vector integrates structural and behavioral information, and after being concatenated with the node state code, it can provide richer input information for the neural network. The lightweight feedforward neural network can efficiently process data, realize high-precision prediction at multiple time scales, and meet the management needs of the distribution network at different time scales.

[0036] As a preferred option for a distribution network source-load prediction method based on multimodal data, wherein:

[0037] The comparison behavior involves prior vectors and prediction results to identify nodes whose differences exceed a preset threshold. A preset strategy is then executed based on the type of abnormal node. The error between the actual observed value and the predicted value is calculated and fed back to the historical event trajectory database. Simultaneously, node attributes in the graph structure are updated, including:

[0038] By comparing the prior behavioral vector with the prediction result, nodes with differences exceeding a preset threshold are identified and judged as potential abnormal nodes; based on the type of abnormal node, the corresponding preset strategy is executed.

[0039] The error between the actual observed value and the predicted value is calculated, and the error information is fed back to the historical event trajectory database. New samples containing the difference between the actual behavior and the predicted behavior in the current state and the control path after the current execution are added to the historical event trajectory database, and the node attributes in the graph structure are updated to form the graph at the next moment.

[0040] The beneficial effects of this preferred technical solution are: timely identification of abnormal nodes and execution of strategies can ensure the stable operation of the distribution network; error feedback and data updates can continuously optimize the historical event trajectory database and graph structure, improve the accuracy of subsequent predictions, and form a virtuous cycle of prediction-feedback-optimization.

[0041] Secondly, the present invention provides a power distribution network source-load prediction system based on multimodal data, comprising:

[0042] The distribution network graph construction module is used to acquire multi-source data of the distribution network, map the multi-source data into feature vectors of nodes, construct edge sets, and generate a graph structure.

[0043] The graph state coding module is used to construct a graph window sequence based on the graph structure and node feature vectors, extract the state of the graph window sequence and merge the node state codes to obtain a unified node state code.

[0044] The historical trajectory matching and weight calculation module is used to introduce the historical event trajectory database, match the node status code with the historical event trajectory database to generate a candidate set, and calculate the path selection weight by combining the weight matching mechanism.

[0045] The behavior prediction output module is used to generate behavior prior vectors at all time scales based on path selection weights and behavior response information in the historical event trajectory database. The node state encoding is concatenated with the behavior prior vectors of the corresponding scale and then input into the feedforward neural network to output the prediction results.

[0046] The anomaly handling and data update module is used to compare the prior vector of behavior with the prediction result, identify nodes whose differences exceed a preset threshold, execute preset strategies according to the type of abnormal node, calculate the error between the actual observation value and the prediction value and feed it back to the historical event trajectory database, and synchronously update the node attributes in the graph structure.

[0047] Thirdly, the present invention provides an electronic device, comprising:

[0048] Memory and processor;

[0049] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the one or more programs are executed by the one or more processors, the one or more processors implement the power distribution network source-load prediction method based on multimodal data as described in this invention.

[0050] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the aforementioned power distribution network source-load prediction method based on multimodal data.

[0051] The beneficial effects of this invention: The multi-timescale source-load prediction method for distribution networks based on multimodal data of this invention has significant advantages in practical applications. Current distribution network operations are complex, and traditional source-load prediction methods suffer from problems such as single data modality, difficulty in handling multi-timescale fluctuations, difficulty in encoding topology evolution, and lack of interpretability. This invention introduces a unified information fusion framework and evolutionary expression mechanism, enabling high-precision, multi-scale prediction of dynamic changes in distribution network sources and loads and deduction of operational status. It can identify and utilize typical evolutionary path templates, effectively enhancing the modeling capability for non-stationary operation scenarios of the power grid, and accurately predicting even when the topology changes frequently, avoiding significant inaccuracies in prediction results. The constructed multi-timescale hierarchical prediction mechanism can respectively handle rapid changes at the minute level, medium-term fluctuations at the hour level, and trend-type changes at the daily / weekly level, accurately predicting source-load dynamics with matched modeling granularity. Simultaneously, this method integrates unstructured multimodal data, such as text scheduling records and user behavior events, improving the model's ability to express and understand complex scenarios. In terms of anomaly handling, preset strategies are executed based on the type of anomaly node. For example, power curtailment decisions are issued for load nodes, peak shaving and charging shifting decisions are issued for electric vehicle nodes, and advance energy storage scheduling decisions are issued for new energy nodes to ensure stable grid operation. Historical event trajectory databases and graph structure node attributes are updated through error feedback to continuously optimize prediction performance. These improvements better meet the high precision and timeliness requirements of distribution network operation and management, enhancing the rationality and reliability of grid dispatching. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is an overall flowchart of the power distribution network source-load prediction method based on multimodal data provided by the present invention. Detailed Implementation

[0054] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0055] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a distribution network source-load prediction method based on multimodal data, including:

[0056] S1: Obtain multi-source data of the distribution network, map the multi-source data to the feature vectors of the nodes, construct the edge set, and generate the graph structure;

[0057] S2: Based on the graph structure and node feature vectors, construct a graph window sequence, extract the state of the graph window sequence and merge the node state codes to obtain a unified node state code;

[0058] S3: Introduce a historical event trajectory database, match the node status code with the historical event trajectory database to generate a candidate set, and combine the weight matching mechanism to calculate the path selection weight;

[0059] S4: Based on the path selection weights and behavioral response information in the historical event trajectory database, generate behavioral prior vectors at all time scales. Then, concatenate the node state encoding with the behavioral prior vectors at the corresponding scales and input them into the feedforward neural network to output the prediction results.

[0060] S5: Compare the prior vector of behavior with the prediction result, identify nodes whose differences exceed the preset threshold, execute the preset strategy according to the abnormal node type, calculate the error between the actual observation value and the prediction value and feed it back to the historical event trajectory database, and update the node attributes in the graph structure synchronously.

[0061] It should be noted that steps S1-S5 effectively utilize and deeply mine multi-source data from the distribution network. From initially constructing the distribution network map structure, to obtaining unified node status codes, then combining historical event trajectory databases for similarity matching and weight calculation, and finally completing behavior prediction, the prediction results are used for anomaly identification and processing and data updates, forming a complete distribution network operation status analysis and prediction system. This system helps to promptly identify abnormal nodes in the distribution network operation, implement targeted strategies, and continuously optimize the historical database and map structure through error feedback, improving the accuracy and reliability of predictions and providing strong support for the stable and efficient operation of the distribution network.

[0062] Example 2, refer to Figure 1As one embodiment of the present invention, based on the previous embodiment, a distribution network source-load prediction method based on multimodal data is provided, including:

[0063] In this embodiment, the steps S1 described above, including acquiring multi-source data of the distribution network, mapping the multi-source data to feature vectors of nodes, constructing an edge set, and generating a graph structure, include:

[0064] Distribution network multi-source data includes distribution network topology data, user load data, electric vehicle charging behavior data, photovoltaic and weather data, and dispatch text logs; among them, distribution network topology data is used to construct structural edges.

[0065] The graph structure includes a set of nodes, a set of edges, and a set of feature vectors for the nodes. The set of edges includes physical connection edges, behavioral structure edges, and textual behavioral semantic edges.

[0066] Specifically, a structured multimodal distribution network state map is constructed. This serves as a unified input structure. The graph integrates distribution network operation data from various sources, including physical topology connections, real-time load, electric vehicle behavior, renewable energy output, and dispatch command text. It represents the physical dependencies and semantic relationships between different entities through a node-edge structure. Each node type employs a specific embedding strategy, mapping the raw data to a unified-dimensional feature space to achieve unified modeling of multimodal information. The final generated multimodal distribution network state graph... It supports dynamic updates and multi-timescale modeling, which forms the basis for subsequent modules such as state representation, event memory, and inference prediction.

[0067] Obtain distribution network topology data from GIS or DAS systems, providing physical connection information between buses, loads, and branches, to construct structural edges in the graph, represented as: .in, This represents a structural edge from node i to node j. This represents the set of edges in the distribution network state diagram at time t.

[0068] User load data is obtained from smart meters, sampled every 15 minutes, and the current power value is extracted. and its change Build load nodes .

[0069] The charging pile system collects electric vehicle charging behavior data, including the current SOC (State of Charge). ), current charging power ( ) and expected disconnection time ( ), construct electric vehicle nodes .

[0070] Retrieve photovoltaic output and weather description from the photovoltaic equipment and weather API, and record the current output. Combined with weather text descriptions (such as "cloudy" or "strong radiation"), photovoltaic nodes are constructed. .

[0071] Obtain scheduling text logs from the scheduling instructions of the DMS system, in natural language format (e.g., "Eastern area load limit 10% at 17:00"), and construct scheduling semantic nodes. .

[0072] First, the multi-source data of the distribution network is mapped to different types of nodes. Each type of node uses a different input structure based on its data characteristics. For example, the input for a load node is... The input of the electric vehicle node is The input to the text node is the scheduling text sentence.

[0073] All node inputs Through unified mapping function Transformed into an embedded representation of the same dimension: ;

[0074] in, For nodes The feature vectors are all set to a dimension of 64.

[0075] The specific implementation is as follows: For numerical nodes (load, electric vehicle, photovoltaic, etc.), a two-layer feedforward network is used (FC1: input dimension → 128, ReLU; FC2: 128 → 64).

[0076] For text nodes, use BERT-base to extract sentence vectors (768 dimensions), and then use a linear transformation to project them to 64 dimensions.

[0077] After completing the node representation, construct the edge set. ,side The construction method is as follows:

[0078] Physical connection edges: determined by the topology, such as transformers and loads, busbars and branches;

[0079] Behavioral-structural edges: such as an electric vehicle connecting to its transformer;

[0080] Text-behavior semantic edge: such as “Eastern Area Load Limitation” log connected to “Eastern Area Load Node”.

[0081] Finally, a multimodal distribution network state diagram is formed. , is represented as: ;

[0082] in, For a moment The set of nodes; Let be a set of edges, representing structural connections and semantic dependencies; This is the set of embedded features for all nodes.

[0083] For example: If the scheduling log at a certain moment reads "South Area load limit after 5 PM", then This text, after being embedded using BERT, is connected to all load nodes in the "Southern Region". If the electric vehicle nodes in this area are at this time It has a physical connection with the load node and will also form a three-layer dependency: scheduling → load → electric vehicle, which demonstrates the advantages of coupled modeling of structure-semantics-behavior.

[0084] It should be noted that this step involves constructing a graph structure that integrates structural data (topology, load) and semantic data (dispatch text) in the distribution network. This enables the construction of a unified information carrier. The graph nodes are based on real business objects, and the edges represent real physical or semantic impact paths. Each type of node adopts a dedicated embedding strategy, which enables the graph to have temporal dynamism, cross-modal fusion, and structural integrity. This is a prerequisite for subsequent multi-timescale state representation, path memory, and inference prediction.

[0085] In another possible implementation, the embedding strategy can be an autoencoder-based embedding strategy, which consists of an encoder and a decoder. The encoder compresses the original feature vectors of the nodes into a low-dimensional latent space, while the decoder attempts to reconstruct the original feature vectors from the representation in the latent space.

[0086] An autoencoder is trained using node features from multi-source data of the distribution network as input. The parameters of the encoder and decoder are learned by minimizing the reconstruction error (e.g., mean squared error). For example, for the feature vector x of a load node... i The training objective is to minimize ||x i −D(E(x i ))∥ 2 Where E represents the encoder and D represents the decoder.

[0087] After training, the encoder's output is the node's embedding representation. A linear transformation can be used to project the embedding representations of different node types into a unified 64-dimensional space.

[0088] In another possible implementation, the embedding strategy can also be a pre-trained model-based embedding strategy. For text-type nodes (such as scheduling semantic nodes), in addition to BERT-base, other pre-trained language models, such as RoBERTa and XLNet, can be used. These models are pre-trained on large-scale text data and are able to learn rich linguistic and semantic information.

[0089] The pre-trained model is fine-tuned on dispatch text data of the distribution network to better adapt it to the textual characteristics of the distribution network domain. During fine-tuning, a linear layer can be added to the output layer of the model to project the output into a 64-dimensional space.

[0090] For numerical nodes (such as load nodes, electric vehicle nodes, and photovoltaic nodes), a pre-trained numerical feature embedding model can be used, or the numerical features can be standardized and then embedded into a 64-dimensional space using a multilayer perceptron (MLP). Finally, the embedded representations of textual nodes and numerical nodes are fused to construct a unified graph structure.

[0091] In this embodiment, step S2 above involves constructing a graph window sequence based on the graph structure and node feature vectors, and then performing state extraction and node state encoding merging on the graph window sequence to obtain a unified node state encoding, including:

[0092] The map window sequence is established by grouping according to different preset time scales;

[0093] The state of the graph window sequence is extracted using a temporal graph convolutional network to obtain the node state code, and a node category-sensitive aggregation operation is introduced to merge the node state codes to obtain a unified node state code; the unified node state code is obtained through time-aware graph convolution and gating dynamic modeling.

[0094] It should be noted that this step aims to visualize the spectral structure. With node embedding features This transforms the data into a structural representation suitable for multi-timescale state evolution modeling. Considering the significant differences in the operational behavior of loads, electric vehicles, and renewable energy sources in the distribution network at different time granularities (e.g., electric vehicles exhibit high-frequency fluctuations at the minute level, loads show regular changes at the hour level, and photovoltaic output shows trend changes at the day level), this step designs a dynamic graph encoding method for multiple time scales. This method extracts graph structural state representations for different time granularities for subsequent prediction and deduction. The key innovation lies in introducing a time-segmented graph convolution mechanism and combining it with adjacency weights and time decay factors to achieve joint structure-time modeling, avoiding the problem of existing methods confusing features at different scales within a unified time model.

[0095] The graph structure and node embedding features are constructed into a dynamic graph sequence grouped by time scale, which is used to extract structural state representations at three time scales.

[0096] Considering the objective of simultaneously modeling source-load state changes at the minute, hour, and day levels, three sets of spectral window sequences are constructed, defined as follows:

[0097] Minute-level chart sequence The time step is 1 minute;

[0098] Hourly chart sequence The time step is 1 hour;

[0099] Celestial Level Chart Sequence The time step is 1 day.

[0100] in, Indicates in Distribution network status diagram at any given time. It is the index of the sequence. It is the length of the minute-level chart sequence. It is the length of the hourly chart sequence. It is the length of the celestial chart sequence.

[0101] For each time scale, a temporal graph convolutional network with the same structure but shared parameters is used for state extraction. Taking the hourly level as an example, graph convolution + gated update unit is used to implement structural temporal modeling, and the state is expressed as follows: ;

[0102] in, For the first A graph structure for each time step; The node features corresponding to this graph; This represents a graph convolution operation, using two layers of graph convolution, each 64-dimensional. The adjacency matrix is ​​based on the edge set. structure; This is a gated recurrent unit with an input dimension of 64 and a hidden dimension of 64, used to model the dynamic evolution of the graph state over time; the output... This represents the structure-time joint representation of each node at the hourly scale.

[0103] This operation is built at the minute and day levels respectively. and The only difference is the time step and the sequence length; the graph network structure remains consistent, and weight sharing ensures that the number of parameters is controllable, thus improving system stability.

[0104] Considering that different types of nodes in the graph (such as load nodes, electric vehicle nodes, and text nodes) have different physical or behavioral response time scales, a node category-sensitive aggregation operation is introduced after state extraction to merge the time scale representations of nodes into a unified state code. , represented as: ;

[0105] in, Output the state at three time scales; This is a linear mapping matrix (not shared), used for dimensionality reduction mapping at different time scales. Specifically, State output on a minute-level timescale The linear mapping matrix, Output status on an hourly timescale The linear mapping matrix, State output on a timescale of days The linear mapping matrix, , This indicates the number of nodes in the graph.

[0106] It should be noted that this step not only integrates the relationships between multimodal nodes at the structural level, but also introduces a time-scale-sensitive graph representation mechanism for the first time, avoiding the information loss problem caused by compressing all temporal behaviors to a single granularity. Through time-aware graph convolution and gated dynamic modeling mechanisms, the system can accurately represent the operational status of various power grid elements at different time granularities in the structure-behavior joint space, which has significant advantages in supporting joint source-load prediction and evolutionary extrapolation.

[0107] In another possible implementation, when constructing the spectrum window sequence, key events in the distribution network can be used as trigger points. For example, when events such as renewable energy access, significant load changes, or equipment failures occur, spectrum data within a certain time range is extracted before and after the event's occurrence time to form a spectrum window. Different time ranges can be set for different types of events. For example, a 1-hour time range can be set before and after a renewable energy access event, and a 2-hour time range can be set before and after a equipment failure event.

[0108] These graph windows are sorted according to the order in which the events occur, forming a graph window sequence. This constructed graph window sequence can focus on the impact of key events on the distribution network, and is more conducive to analyzing and predicting changes in the system state after the events occur.

[0109] In another possible implementation, when constructing the spectrum window sequence, the periodic patterns of distribution network operation data, such as daily, weekly, and monthly cycles, can also be analyzed. Each cycle is divided into several equal time segments. For example, for a daily cycle, a day can be divided into 24 one-hour time segments.

[0110] A spectral window is constructed based on each time segment, with the spectral data within each time segment forming a spectral window. These spectral windows are then combined into a sequence according to time order. This method is suitable for distribution networks with obvious periodic patterns in operation and can better capture similar behavioral patterns of the system across different periods.

[0111] In this embodiment, step S3 above introduces a historical event trajectory database, performs similarity matching between node status codes and the historical event trajectory database to generate a candidate set, and calculates path selection weights using a weighted matching mechanism, including:

[0112] The historical event trajectory database includes node state vectors, behavioral response information, and the structural evolution path of the event before it begins. Actual events refer to events that actually occur in real-world scenarios, such as faults or load surges in a power distribution network.

[0113] In similarity matching, this embodiment preferably uses weighted cosine similarity as the matching index; a preset number of similar historical samples are selected to form a candidate set based on similarity ranking; a weighted matching mechanism is introduced to align the changing trend of the graph structure with the structural evolution path in the historical trajectory database and calculate the path selection weight.

[0114] The path selection weight uses a topology alignment scoring function to measure the similarity between the current topology change trend and the structural evolution chain in the historical path, which is achieved by a hybrid index of edit distance and topology similarity.

[0115] Specifically, a historical event trajectory database is introduced. ,in, Indicates the first The node state vector before the start of a historical event; This indicates the source load change behavior (i.e., the behavioral outcome) after the state evolves. It represents the structural evolution path of the event and records the sequence of events such as topology changes and anomalies.

[0116] First, encode the current node state. With historical event status coding Similarity matching is performed to filter historical event trajectories that are closest to the current map state. Weighted cosine similarity is used as the matching metric, defined as: ;

[0117] in, The current node The state representation, For the first The state of each historical sample corresponding to a node; It is a weighting factor defined according to the node type. For example, the weight of a load node is set to 1.0, an electric vehicle node to 0.8, and a scheduling text node to 0.5, indicating the strength of their influence on the evolution of the event. Represents the dot product. It is a 2-norm.

[0118] By calculating similarity, the top [number] [items] are selected. A candidate set consists of similar historical samples. .

[0119] To further improve the adaptability of the callback trajectory, a structure-sensitive weight matching mechanism is introduced. This mechanism aligns the changing trend of the current graph structure with the evolution direction of historical trajectories to calculate the final path selection weight. ;

[0120] in, For the first The matching weights of the candidate paths; Used to control the degree of influence of state matching when calculating the weights of the final path selection. Used to control the degree of influence of structural alignment when calculating the final path selection weight; The topology alignment scoring function measures the similarity between the current topological change trend and the structural evolution chain in the historical path, and is implemented using a hybrid index of edit distance and topology similarity. Ensure that the output is a normalized weight, which will be used as a prior reference in the subsequent prediction module.

[0121] It should be noted that this step not only utilizes the current state encoding for vector space similarity matching, but also determines whether "historical events are truly representative in terms of structural evolution" by aligning the topological evolution paths between graph structures. This avoids recalling paths unrelated to structure solely based on state "similarity," enhancing the structural adaptability of memory recall and the rationality of prediction scenarios. Furthermore, to ensure the practical usability of path recall, each historical trajectory includes not only the state change results... It also includes the structural evolution path. This includes information such as specific topology adjustment actions, text commands, and response times.

[0122] In another possible implementation, the path selection weights can be computed by building a deep learning model: a recurrent neural network such as a Long Short-Term Memory (LSTM) network or a Gated Recurrent Unit (GRU) can be used, or a Transformer architecture can be employed. The current node state encoding and information such as node state vectors and structural evolution paths from a historical event trajectory database can be used as input features. For example, the node state encoding and structural evolution paths can be converted into sequence data and input into the LSTM.

[0123] The goal of training using historical data is to predict path selection weights. The actual path selections of historical events can be used as labels, and the model is trained by minimizing the loss function (such as mean squared error loss) between the predicted and actual weights.

[0124] By inputting the current node state encoding and historical event trajectories into a trained model, the model output becomes the path selection weights. This method can automatically learn the complex relationship between node states, structural evolution paths, and path selection weights, potentially offering greater accuracy than traditional similarity matching methods.

[0125] In another possible implementation, the path selection weights can also be calculated by means of path selection weights based on fuzzy comprehensive evaluation;

[0126] Specifically, the evaluation factor set is determined, including factors such as node state similarity, structural evolution path similarity, and behavioral response information similarity. For example, node state similarity can be measured by Euclidean distance, Manhattan distance, etc.; structural evolution path similarity can be measured by graph editing distance, etc.; and behavioral response information similarity can be quantified according to the type and degree of behavioral response.

[0127] The weights of each evaluation factor were determined using methods such as expert scoring and the analytic hierarchy process (AHP). For example, based on experience, experts determined that the importance of node state similarity in path selection was 0.4, the importance of structural evolution path similarity was 0.5, and the importance of behavioral response information similarity was 0.1.

[0128] For each candidate path, a fuzzy evaluation is given based on its performance on various evaluation factors. For example, node state similarity can be divided into three levels: "high," "medium," and "low," with corresponding membership degrees assigned.

[0129] The fuzzy evaluation results of each evaluation factor are weighted and summed with their respective weights to obtain a comprehensive evaluation score for each candidate path. Finally, the comprehensive evaluation score is normalized to obtain the path selection weight. This method can comprehensively consider the influence of multiple factors on path selection and can handle the uncertainty of evaluation information.

[0130] In another possible implementation, when selecting the matching metric, the Manhattan distance can also be selected as the matching metric, and the Manhattan distance between the current node state code Z and the j-th historical event state code Z(j) can be calculated. Then, based on the Manhattan distance sorted from smallest to largest, the top K historical samples with the smallest distances were selected to form a candidate set.

[0131] In another possible implementation, the Jaccard similarity coefficient can be selected as the matching index. The node state codes are treated as a set, and the Jaccard similarity coefficient between the current node state code Z and the j-th historical event state code Z(j) is calculated. Sort the historical samples by Jaccard similarity coefficient from largest to smallest, and select the top K samples with the highest similarity coefficients to form a candidate set.

[0132] In this embodiment, in step S4 above, based on path selection weights and behavioral response information in the historical event trajectory database, behavioral prior vectors are generated at all time scales. The node state encoding is concatenated with the behavioral prior vector at the corresponding scale and then input into the feedforward neural network. The output prediction results include:

[0133] The behavioral prior vector represents the response value of the trajectory at each node on the corresponding time scale;

[0134] First, based on the behavior response information in the historical path... Matching weights Generate a trajectory behavior prior vector for structural behavior fusion at each time scale. The calculation is as follows: ;

[0135] in, Representing the trajectory The response value of each node on a minute / hour / day scale (e.g., the rise and fall of node power, in kW). It is a matching weight calculated based on state-structure similarity; It is a weighted behavior template that integrates K trajectories.

[0136] Current state Behavioral priors at the corresponding scale After concatenation, the data is input into a lightweight feedforward network for prediction, and the calculation is as follows: ;

[0137] in, By concatenating the current node state with the historical behavior template, a new node is formed. Input; It is a two-layer feedforward neural network, with the following structure:

[0138] The first layer is Activated as ReLU;

[0139] The second layer is Output the predicted value for each node;

[0140] This represents the predicted power value for each node at the next time step.

[0141] Execution is performed on three scales: minutes, hours, and days, with outputs... .

[0142] Obtain multi-timescale prediction results This represents the power prediction of each node at future times; behavioral template. (One copy each for minutes, hours, and days), as historical priors used in the current forecast.

[0143] It should be noted that this step introduces structural memory into the prediction process for the first time, building a bridge between prediction and evolutionary deduction, and providing the foundational data of structure, behavior, and outcome for the next step of generating operational strategies and closed-loop optimization. This design is applicable to highly dynamic scenarios in distribution networks caused by factors such as weather, dispatching, and electric vehicle access, improving the model's responsiveness and interpretability to sudden states.

[0144] In this embodiment, step S5 above compares the prior vector of behavior with the prediction result, identifies nodes whose differences exceed a preset threshold, executes a preset strategy based on the abnormal node type, calculates the error between the actual observed value and the predicted value and feeds it back to the historical event trajectory database, and synchronously updates the node attributes in the graph structure, including:

[0145] Preset strategies include:

[0146] If the abnormal indicator exceeds the preset threshold and the node is a load node, a power rationing decision will be output.

[0147] If the node is an electric vehicle node, output peak shaving and charging shifting decisions;

[0148] If the node is a new energy node and the predicted decline is greater than the historical maximum fluctuation, output an advance energy storage scheduling decision.

[0149] Specifically, firstly, by comparing the predicted values... With behavioral templates This identifies nodes with potentially atypical behaviors, i.e., areas where the predicted behavior deviates significantly from the historical trajectory. Anomaly indicators are defined as follows: ;

[0150] in, Represents a node The intensity of the anomaly at this scale; It is a predicted value. It is the historical trajectory behavior template value; if Exceeding the threshold If the value is twice the average, the node is identified as a "potentially abnormal node," triggering the generation of a scheduling policy.

[0151] Based on the type of abnormal node (overload, concentrated charging of electric vehicles, drastic fluctuations in new energy sources, etc.), the following recommended strategy logic should be implemented:

[0152] like and If it is a load node, then output a power limiting suggestion;

[0153] like If the node is an electric vehicle, then a peak-shaving and charging shifting suggestion will be output;

[0154] like If the node is a new energy source and the predicted decline is greater than the historical maximum fluctuation, then the suggestion of "advance energy storage scheduling" will be output.

[0155] in, Represents a node Anomaly intensity on an hourly timescale.

[0156] The above strategy is structured and encoded as "map attribute update instructions", for example:

[0157] node Attributes Set as requiring regulation;

[0158] side Attributes An increase indicates that a topology refactoring may be required.

[0159] After the strategy is generated, the actual observations will be... The error is calculated by comparing the data with the predicted value, and this error information is fed back to the historical trajectory memory. The update method is as follows: ;

[0160] in, This represents the expanded historical event trajectory database. This represents the difference between the actual behavior and the predicted behavior in the current state; This is the current control path after execution, such as whether to activate power rationing or whether to call up energy storage; this new sample is added to the event trajectory memory bank as a reference trajectory for future structural evolution.

[0161] Ultimately, the graph attribute structure The node attributes are updated to form the graph at the next time step. The graph changes do not involve topological reconnection, but only update node / edge attributes, such as marking logical fields like "whether it is under control" and "whether it is abnormal", to ensure that the system's physical structure is consistent with the actual operating mechanism.

[0162] It should be noted that this step completes the closed-loop logic from "prediction results" to "operational suggestion generation" and "system feedback update". Compared with the traditional system's strategy of generating strategies based solely on predicted values, this system, for the first time, integrates trajectory templates with the current state to identify anomalies, drives strategy generation through structured rules, and feeds back the execution results to the graph node attributes and historical memory bank, thus realizing a distribution network prediction-deduction-control system with evolutionary capabilities and self-adjustment.

[0163] Example 3: The above is an illustrative scheme of the distribution network source-load prediction method based on multimodal data in this embodiment. It should be noted that the technical solution of the distribution network source-load prediction system based on multimodal data and the technical solution of the distribution network source-load prediction method based on multimodal data described above belong to the same concept. Details not described in detail in the technical solution of the distribution network source-load prediction system based on multimodal data in this embodiment can be found in the description of the technical solution of the distribution network source-load prediction method based on multimodal data described above.

[0164] This embodiment also provides a power distribution network source-load prediction system based on multimodal data, including:

[0165] The distribution network graph construction module is used to acquire multi-source data of the distribution network, map the multi-source data into feature vectors of nodes, construct edge sets, and generate a graph structure.

[0166] The graph state coding module is used to construct a graph window sequence based on the graph structure and node feature vectors, extract the state of the graph window sequence and merge the node state codes to obtain a unified node state code;

[0167] The historical trajectory matching and weight calculation module is used to introduce the historical event trajectory database, match the node status code with the historical event trajectory database to generate a candidate set, and calculate the path selection weight by combining the weight matching mechanism.

[0168] The behavior prediction output module is used to generate behavior prior vectors at all time scales based on path selection weights and behavior response information in the historical event trajectory database. The node state encoding is concatenated with the behavior prior vectors at the corresponding scales and then input into the feedforward neural network to output the prediction results.

[0169] The anomaly handling and data update module is used to compare the prior vector of behavior with the prediction result, identify nodes whose differences exceed a preset threshold, execute preset strategies according to the type of abnormal node, calculate the error between the actual observation value and the prediction value and feed it back to the historical event trajectory database, and synchronously update the node attributes in the graph structure.

[0170] This embodiment also provides an electronic device applicable to power distribution network source-load prediction methods based on multimodal data, including:

[0171] The system includes a memory and a processor. The memory stores computer-executable instructions, and the processor executes these instructions to implement the power distribution network source-load prediction method based on multimodal data as proposed in the above embodiments.

[0172] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the power distribution network source-load prediction method based on multimodal data as proposed in the above embodiments.

[0173] The storage medium proposed in this embodiment belongs to the same inventive concept as the power distribution network source-load prediction method based on multimodal data proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0174] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the source and load of a distribution network based on multimodal data, characterized in that, include: Acquire multi-source data from the power distribution network, map the multi-source data to the feature vectors of the nodes, construct an edge set, and generate a graph structure; Based on the graph structure and node feature vectors, a graph window sequence is constructed. The graph window sequence is then subjected to state extraction and node state encoding merging to obtain a unified node state encoding. By introducing a historical event trajectory database, the node status code is matched with the historical event trajectory database to generate a candidate set. Combined with a weighted matching mechanism, the path selection weight is calculated. Based on path selection weights and behavioral response information from historical event trajectory databases, behavioral prior vectors are generated at all time scales. The node state encoding is concatenated with the behavioral prior vectors at the corresponding scales and then input into the feedforward neural network to output the prediction results. By comparing the prior behavioral vector with the prediction result, nodes with differences exceeding a preset threshold are identified, and preset strategies are executed based on the abnormal node type. Calculate the error between the actual observed value and the predicted value and feed it back to the historical event trajectory database, and update the node attributes in the graph structure synchronously; The process involves constructing a graph window sequence based on the graph structure and node feature vectors, then extracting the state of the graph window sequence and merging the node state codes to obtain a unified node state code, including: Based on different preset time scales, the feature vectors of the graph structure and nodes are grouped to establish a graph window sequence; For each timescale of the graph window sequence, a time-series graph convolutional network with the same structure and shared parameters is used to extract the state, thereby obtaining the node state encoding at different timescales. By introducing a node category-sensitive aggregation operation, node state codes at different time scales are merged to obtain a unified node state code. A weighted matching mechanism is introduced to align the current topological trend with the structural evolution paths in the historical trajectory database, measure the similarity between the current topological trend and the structural evolution chains in the historical paths, and calculate the path selection weight by combining the state similarity matching results and the topological alignment results. Based on path selection weights and behavioral response information from the historical event trajectory database, behavioral prior vectors are generated at all time scales. The node state encoding is concatenated with the corresponding scale's behavioral prior vector and input into the feedforward neural network. The output prediction results include: Based on path selection weights and behavioral response information from historical event trajectory databases, trajectory behavior prior vectors for structural behavior fusion are generated at all preset time scales. The node state code at each time scale is concatenated with the behavior prior vector at the corresponding scale to form the input data for prediction. The concatenated data is input into a lightweight feedforward neural network for prediction, and the prediction results of each node are output at each preset time scale, finally resulting in a set of prediction results across multiple time scales.

2. The distribution network source-load prediction method based on multimodal data as described in claim 1, characterized in that, The process of acquiring multi-source data from the distribution network, mapping the multi-source data to feature vectors of nodes, constructing an edge set, and generating a graph structure includes: A multimodal distribution network state map is constructed based on multi-source data of the distribution network. The map structure includes a set of nodes, a set of edges, and a set of feature vectors of the nodes. Multi-source data from the power distribution network is mapped to different types of nodes, and each type of node adopts a different input structure based on its data characteristics. Through a unified embedding strategy, the original input data of all nodes are mapped to a feature space of a unified dimension to obtain the feature vector of the node.

3. The distribution network source-load prediction method based on multimodal data as described in claim 2, characterized in that, The process involves introducing a historical event trajectory database, matching node status codes with the database for similarity to generate a candidate set, and then calculating path selection weights using a weighted matching mechanism. Select a matching metric and perform similarity matching between the current node status code and the node status vector in the historical event trajectory database; based on the similarity ranking, select a preset number of similar historical samples to form a candidate set.

4. The distribution network source-load prediction method based on multimodal data as described in claim 3, characterized in that, The comparison behavior prior vector and prediction result are used to identify nodes whose differences exceed a preset threshold, and a preset strategy is executed according to the abnormal node type. The error between actual observations and predicted values ​​is calculated and fed back to the historical event trajectory database, synchronously updating the node attributes in the atlas structure, including: By comparing the prior behavioral vector with the prediction result, nodes with differences exceeding a preset threshold are identified and judged as potential abnormal nodes; based on the type of abnormal node, the corresponding preset strategy is executed. The error between the actual observed value and the predicted value is calculated, and the error information is fed back to the historical event trajectory database. New samples containing the difference between the actual behavior and the predicted behavior in the current state and the control path after the current execution are added to the historical event trajectory database, and the node attributes in the graph structure are updated to form the graph at the next moment.

5. A power distribution network source-load prediction system based on multimodal data, using the method described in any one of claims 1 to 4, characterized in that, include: The distribution network graph construction module is used to acquire multi-source data of the distribution network, map the multi-source data into feature vectors of nodes, construct edge sets, and generate a graph structure. The graph state coding module is used to construct a graph window sequence based on the graph structure and node feature vectors, extract the state of the graph window sequence and merge the node state codes to obtain a unified node state code; The historical trajectory matching and weight calculation module is used to introduce the historical event trajectory database, match the node status code with the historical event trajectory database to generate a candidate set, and calculate the path selection weight by combining the weight matching mechanism. The behavior prediction output module is used to generate behavior prior vectors at all time scales based on path selection weights and behavior response information in the historical event trajectory database. The node state encoding is concatenated with the behavior prior vectors at the corresponding scales and then input into the feedforward neural network to output the prediction results. The anomaly handling and data update module is used to compare the behavior prior vector with the prediction result, identify nodes whose differences exceed a preset threshold, and execute preset strategies according to the anomaly node type. The error between the actual observed value and the predicted value is calculated and fed back to the historical event trajectory database, and the node attributes in the graph structure are updated synchronously.

6. An electronic device, characterized in that, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, It stores computer-executable instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 4.

Images