Search

CN-121998180-A - Electric automobile charging demand prediction method, system and electronic equipment

CN121998180ACN 121998180 ACN121998180 ACN 121998180ACN-121998180-A

Abstract

The invention provides a method and a system for predicting charging requirements of an electric automobile and electronic equipment, and relates to the technical field of deep learning prediction. The method comprises the steps of obtaining historical observation data and corresponding time characteristic data of an electric vehicle charging station, constructing MIFM a prediction model, enabling an initial input information coding module to receive the initial data and conduct parallel embedding processing, constructing unified input representation comprising node specific self-adaptive embedding, enabling a space-time information dynamic fusion module to obtain the unified input representation, capturing the correlation between explicit geographic spatial dependence and implicit function in parallel through a dual-channel dynamic graph learning mechanism, dynamically integrating multiple layers of space-time characteristics through a self-adaptive Kalman filtering fusion strategy to generate an enhanced fusion signal, enabling a space-time dependence modeling and prediction module to learn the spatial relationship representation based on the enhanced fusion signal through a stacked graph attention network, capturing time dynamics through a time sequence decoding module, and generating a charging demand prediction result of a future time step.

Inventors

  • MEI QICHENG
  • CHEN ZIJIAN

Assignees

  • 三峡大学

Dates

Publication Date
20260508
Application Date
20260116

Claims (10)

  1. 1. The electric automobile charging demand prediction method is characterized by comprising the following steps of: Acquiring historical observation data and corresponding time characteristic data of an electric vehicle charging station; Constructing MIFM a prediction model, inputting the historical observation data and the corresponding time feature data, and obtaining a prediction result, wherein the MIFM prediction model is a depth mixed model based on a graph attention network and a transducer architecture, the whole adopts an encoder-decoder type space-time prediction architecture, the front end processes initial features through a multi-layer perceptron and an embedded layer, the middle end adopts a parallel dynamic graph learning structure, and the rear end respectively carries out space modeling and time modeling by combining a stacked graph attention network and a time sequence decoding module based on the transducer encoder, and the method specifically comprises the following steps of: The method comprises the steps of receiving initial data by an initial input information coding module, carrying out parallel embedding processing to construct unified input representation containing node specific self-adaption embedding, acquiring the unified input representation by a space-time information dynamic fusion module, capturing the correlation between explicit geospatial dependence and implicit function in parallel through a dual-channel dynamic graph learning mechanism, dynamically integrating multiple layers of space-time characteristics by utilizing a self-adaption Kalman filtering fusion strategy to generate an enhanced fusion signal, learning a spatial relationship representation by a space-time dependence modeling and prediction module by utilizing a stacked graph attention network based on the enhanced fusion signal, capturing time dynamics by a time sequence decoding module based on a transform encoder, and generating a charging demand prediction result of a future time step.
  2. 2. The method for predicting the charging demand of an electric vehicle according to claim 1, further comprising training MIFM the prediction model, and retaining an optimal MIFM prediction model: And a regularization constraint module based on a related information principle is introduced, a regularization loss function comprising structural entropy constraint and divergence constraint is constructed, learning of attention weight and connection strength is guided by combining with a prediction loss function, and model parameters are optimized.
  3. 3. The method for predicting the charging demand of an electric vehicle according to claim 1, wherein the parallel embedding of the historical observation data and the time feature data by using the initial input information encoding module comprises: mapping the occupancy rate data of the original charging pile into high-dimensional basic characteristic embedding through the full connection layer; mapping time of day (ToD) and day of week (DoW) into ToD and DoW embeddings, respectively, using a learnable embedding matrix; introducing a learnable adaptive embedding parameter to capture a node-specific aperiodic time pattern; After the basic features are embedded, the ToD embedded, the Dow embedded and the self-adaptive embedded parameters are spliced, the unified input representation is obtained through the processing of a full connection layer and an activation function, wherein the formula is as follows: ; Wherein the method comprises the steps of The occupancy rate data of the original charging pile is obtained; embedding basic features; Time of day and day of week, respectively; embedding for node specific adaptation; is a comprehensive feature tensor; -representing for said unified input; representing a fully connected layer; Representing a stitching operation.
  4. 4. The method for predicting charging demand of an electric vehicle according to claim 1, wherein capturing explicit geospatial dependencies and implicit functional dependencies in parallel through a two-channel dynamic graph learning mechanism comprises: the dynamic filter graph neural network unit is utilized to learn the dynamic filter matrix, and the dynamic adjacency weight matrix is generated by combining the static adjacency matrix, wherein the formula is as follows: ; Wherein, the The weight coefficient is represented by a number of weight coefficients, Representing the original static adjacency matrix, The linear transformation layer is represented by a layer, Representing a learnable filtering matrix; The implicit function correlation among nodes is inferred based on a multi-head self-attention mechanism by utilizing a data-driven computing unit, and implicit function correlation characteristics are generated, and the method is realized by the following formula: ; Wherein, the Representing the similarity matrix of the current calculation, The representation of nodes processed through the multi-headed attention mechanism, Representing an attention dimension; to pass through the learnable parameters The generated gating value; a similarity weight matrix representing the previous training round; The implicit correlation weight matrix after adjustment; Is an intermediate feature subjected to an activation process; is a learnable influencing factor.
  5. 5. The method for predicting the charging demand of an electric vehicle according to claim 1, wherein the dynamic integration of the multi-layered spatio-temporal features using an adaptive kalman filter fusion strategy comprises: Extracting time sequence features by using a parallel time sequence signal extractor, and respectively combining with the dynamic adjacent weight matrix and the implicit function related features to obtain space-time fusion features, wherein the space-time fusion features are realized by the following formula: ; Wherein, the Represent a learnable impact factor; Computing variance tensors of the spatio-temporal fusion features As a measure of observation uncertainty and calculate an accuracy weight tensor based on the variance tensor ; And carrying out weighted fusion on the characteristics from different layers by utilizing the precision weight tensor to obtain the enhanced fusion signal, wherein the enhanced fusion signal is realized by the following formula: ; Wherein, the In correspondence with the characteristic observation value(s), Is a learnable parameter.
  6. 6. The method of claim 1, wherein the learning the spatial relationship representation based on the enhanced fusion signal using the stacked graph attention network and capturing the time dynamics via the timing decoding module generates the charging demand prediction result of the future time step, comprises: Inputting the enhanced fusion signals into a K-layer stacked graph attention network, wherein each layer utilizes a multi-head attention mechanism to dynamically calculate attention scores among nodes based on the characteristic representation of the previous layer, acquires final attention weights by fusing the multi-head attention scores through learnable linear transformation and carrying out normalization processing, further utilizes the attention weights to carry out weighted aggregation on the node characteristics, combines residual connection to form the output of the layer, and extracts space enhanced characteristics containing complex space dependence layer by layer; Inputting the space enhancement features to an L-layer time sequence decoding module based on a transducer encoder architecture, wherein each layer processes the features through a multi-head self-attention sub-layer and a bitwise feedforward network sub-layer which are sequentially connected, and residual connection and layer normalization are applied after each sub-layer so as to capture long-term and short-term time dependence; And obtaining the output characteristics of the last layer of the time sequence decoding module, and mapping the output characteristics to a predicted target space through a full-connection layer to obtain a predicted value of the charging demand of a target future time step.
  7. 7. The method of claim 6, wherein obtaining the predicted charging demand value for the target future time step is achieved by the following formula: ; ; ; in the formula, Is the first Layer node And Attention weight in between; is an attention score; Is the first Layer drawing annotates output features of the force network; Is an attention weight matrix; Is a weight coefficient; Is that Final output characteristics of the layer timing decoding module; And predicting a result for the charging demand.
  8. 8. The method for predicting the charging demand of the electric vehicle according to claim 2, wherein the introducing a regularization constraint module based on a related information principle includes: Constructing a graph attention weight regularization term, and reserving a key space relation by minimizing the structural entropy of an attention weight matrix and maximizing the log likelihood of edges existing in an original graph; Constructing an adaptive weight regularization term, and maintaining consistency of the dependency relationship by minimizing entropy of a node correlation structure and minimizing KL divergence between the learned correlation representation and a graph structure representation based on physical connection; And gradually transitioning the simplified regularization loss to the full regularization loss by adopting a smooth transition mechanism, wherein the formula is as follows: ; Wherein, the A dynamic weight coefficient that increases from 0 to 1 with training rounds; to simplify regularization loss; For attention weighting matrix for graph Is a structural entropy constraint term of (1); for weighting matrices for implicit dependencies Is used for the adaptive weight regularization term of the (a).
  9. 9. An electric vehicle charging demand prediction system, which is adapted to the electric vehicle charging demand prediction method of any one of claims 1 to 8, comprising: the data acquisition module is used for acquiring historical observation data and corresponding time characteristic data of the electric vehicle charging station; The module construction and prediction module is used for constructing MIFM a prediction model, inputting the historical observation data and the corresponding time feature data, and obtaining a prediction result, wherein the MIFM prediction model is a deep mixed model based on a graph attention network and a transducer architecture, the whole adopts a space-time prediction architecture of an encoder-decoder type, the front end processes initial features through a multi-layer perceptron and an embedded layer, the middle end adopts a parallel dynamic graph learning structure, and the rear end combines a stacked graph attention network and a time sequence decoding module based on the transducer encoder to respectively carry out space modeling and time modeling, and the method specifically comprises the following steps: The method comprises the steps of receiving initial data by an initial input information coding module, carrying out parallel embedding processing to construct unified input representation containing node specific self-adaption embedding, acquiring the unified input representation by a space-time information dynamic fusion module, capturing the correlation between explicit geospatial dependence and implicit function in parallel through a dual-channel dynamic graph learning mechanism, dynamically integrating multiple layers of space-time characteristics by utilizing a self-adaption Kalman filtering fusion strategy to generate an enhanced fusion signal, learning a spatial relationship representation by a space-time dependence modeling and prediction module by utilizing a stacked graph attention network based on the enhanced fusion signal, capturing time dynamics by a time sequence decoding module based on a transform encoder, and generating a charging demand prediction result of a future time step.
  10. 10. An electronic device comprising a memory and a processor; The memory is used for storing a computer program; the processor is configured to implement the electric vehicle charging demand prediction method according to any one of claims 1 to 8 when executing the computer program.

Description

Electric automobile charging demand prediction method, system and electronic equipment Technical Field The invention relates to the technical field of deep learning prediction, in particular to a method and a system for predicting charging requirements of an electric automobile and electronic equipment. Background With the promotion of urban propulsion and environment sustainable development demands, the popularization rate of electric vehicles is continuously increased, charging demand prediction becomes a key link for optimizing resource allocation, relieving peak queuing pressure, avoiding overload of charging piles and reducing the influence of power grid load, and the electric vehicle charging system is essentially a complex space-time prediction problem, and has the core challenge of accurately representing complicated spatial dependence among geographically dispersed charging stations and simultaneously coping with prediction uncertainty caused by dynamic change of data. The prior art has the following defects that firstly, models such as a long short term memory network (LSTM), a Transformer and the like are relied on in early stage, spatial correlation among stations is ignored, prediction accuracy is insufficient under a complex space interaction scene of a city, secondly, a Graph Neural Network (GNN) related model (such as STID, ASTGRN, graphWaveNet and the like) which is introduced later still has the problem that spatial relation fusion and dynamic adaptation capability are insufficient, most of the models adopt a simple linear fusion mode, the guiding effect of time information is ignored, the graph structure is fixed after training, and the correlation intensity among nodes cannot be dynamically adjusted according to real-time input in an reasoning stage. 3. The weight learning lacks an effective constraint and balance mechanism, and the existing methods (such as PGCN and G-STAN) introduce partial dynamic mechanisms, but do not have independent constraint modules, so that the learned relation strength is easy to be unstable when the data contains noise or the pattern is deviated, and the information redundancy or generalization is insufficient. In summary, how to dynamically infer a graph structure and apply effective constraint to a learning process on the basis of effectively integrating multi-source space-time information, so as to improve prediction accuracy and robustness, and become a key problem to be solved in the field of current electric vehicle charging demand prediction. Disclosure of Invention The invention mainly aims to provide a method, a system and electronic equipment for predicting the charging requirement of an electric automobile, and solves the problems of insufficient space correlation mining, single data, lack of constraint in weight learning and insufficient prediction precision in the prior art. In order to solve the technical problems, the technical scheme adopted by the invention is that the method for predicting the charging requirement of the electric automobile comprises the following steps: Acquiring historical observation data and corresponding time characteristic data of an electric vehicle charging station; Constructing MIFM a prediction model, inputting the historical observation data and the corresponding time feature data, and obtaining a prediction result, wherein the MIFM prediction model is a depth mixed model based on a graph attention network and a transducer architecture, the whole adopts an encoder-decoder type space-time prediction architecture, the front end processes initial features through a multi-layer perceptron and an embedded layer, the middle end adopts a parallel dynamic graph learning structure, and the rear end respectively carries out space modeling and time modeling by combining a stacked graph attention network and a time sequence decoding module based on the transducer encoder, and the method specifically comprises the following steps of: The method comprises the steps of receiving initial data by an initial input information coding module, carrying out parallel embedding processing to construct unified input representation containing node specific self-adaption embedding, acquiring the unified input representation by a space-time information dynamic fusion module, capturing the correlation between explicit geospatial dependence and implicit function in parallel through a dual-channel dynamic graph learning mechanism, dynamically integrating multiple layers of space-time characteristics by utilizing a self-adaption Kalman filtering fusion strategy to generate an enhanced fusion signal, learning a spatial relationship representation by a space-time dependence modeling and prediction module by utilizing a stacked graph attention network based on the enhanced fusion signal, capturing time dynamics by a time sequence decoding module based on a transform encoder, and generating a charging demand prediction result of a future time step.