A step-controllable load prediction method and system based on multi-granularity timing

Abstract

The application discloses a step length controllable load prediction method and system based on multi-granularity time sequence, and belongs to the technical field of load prediction.The application can cover different periodic modes of electric vehicle charging load through splicing historical charging load data, can extract and fuse time sequence information of multiple granularities through a multi-granularity cross-time sequence feature extraction algorithm, can model long-time sequence features through a spiral ring scanning state space processing algorithm and a parallel ring scanning state space processing algorithm, can mine long-distance jump correlations in the charging load sequence which are discontinuous and cross time periods, can enhance the prediction adaptability of the randomness and fluctuation of the electric vehicle charging load in the charging load prediction step length, and can adaptively fuse the multi-time interval load sequence and the long-time sequence features of the load prediction through a load prediction gating coefficient, so that the application realizes accurate prediction of the electric vehicle charging load prediction.

Description

Technical Field

[0001] This invention belongs to the field of load forecasting technology, and particularly relates to a step-controllable load forecasting method and system based on multi-granularity time series. Background Technology

[0002] Electric vehicle charging load exhibits significant temporal concentration and regional variability, and is influenced by multiple factors such as user charging behavior, weather conditions, and traffic conditions, resulting in high randomness, suddenness, and nonlinearity. This makes its prediction significantly more difficult than that of traditional electricity load.

[0003] Currently, load forecasting methods for electric vehicle (EV) charging mainly consist of statistical modeling and deep learning-based methods. Statistical modeling methods typically use autoregressive models and Monte Carlo methods, which are often based on linear assumptions and lack adaptability to highly volatile and non-stationary EV charging load data, leading to inaccurate load forecasts. Deep learning-based methods, typically using convolutional neural networks and long short-term memory networks, can extract features and capture time dependencies, but are limited by the multi-factor influence of EV charging load. They have limitations in long-term time-series dependency modeling, multi-granularity time-series feature extraction, and the fusion of information from different time granularities. This can result in incomplete time-series feature extraction and decreased accuracy in charging load forecasting under long prediction step sizes, further contributing to inaccurate EV charging load forecasts. Furthermore, current deep learning methods usually rely on a fixed prediction step size for load forecasting, which is inaccurate due to the randomness and volatility of EV charging load. Therefore, there is an urgent need for a multi-granularity time-series-based load forecasting method and system with controllable step sizes to address the shortcomings of existing technologies. Summary of the Invention

[0004] This invention aims to provide a step-controllable load forecasting method and system based on multi-granularity time series to solve the technical problem of inaccurate charging load forecasting for electric vehicles in the prior art. By using multi-granularity cross-time series feature extraction, spiral ring scanning state space processing algorithm and parallel ring scanning state space processing algorithm, it realizes long-time series dependency modeling, multi-granularity time series feature extraction and fusion of information at different time granularities, thereby improving the accuracy of charging load forecasting for electric vehicles.

[0005] To address the aforementioned technical problems, embodiments of the present invention provide a step-size controllable load forecasting method based on multi-granularity time series, comprising: Historical charging load data of electric vehicles is obtained, and the historical charging load data is spliced ​​together according to the interval time of the historical charging load data to construct a multi-time interval load sequence; Based on a preset multi-granularity cross-temporal feature extraction algorithm, multi-granularity temporal feature extraction within groups and between groups is performed on the multi-time interval load sequence to obtain multi-granularity load temporal features; The charging load prediction step size of the electric vehicle is obtained, and a prediction step size embedding vector is constructed based on the charging load prediction step size. Multi-granularity load time-series features are determined based on the multi-granularity load time-series features. Long-term time-series feature modeling is performed on the multi-granularity load time-series features using a preset spiral ring scan state-space processing algorithm and a preset parallel ring scan state-space processing algorithm, combined with the prediction step size embedding vector, to obtain the spiral ring scan load prediction long-term time-series features and the parallel ring scan load prediction long-term time-series features. The load prediction long-term time-series features are obtained based on the spiral ring scan load prediction long-term time-series features and the parallel ring scan load prediction long-term time-series features. The load forecasting gating coefficient is calculated based on the multi-time interval load sequence and the long-term time series characteristics of load forecasting; and the multi-time interval load sequence and the long-term time series characteristics of load forecasting are fused based on the load forecasting gating coefficient to determine the charging load forecasting result of the electric vehicle.

[0006] It is understood that this invention splices historical charging load data by using the interval time of historical charging load data, enabling multi-time interval load sequences to cover different periodic patterns of electric vehicle charging load. Through intra-group and inter-group multi-granularity time-series feature extraction, it can extract and fuse time-series information of multiple granularities, transforming multi-time interval load sequences into multi-granularity load time-series features containing multi-period priors, providing rich multi-granularity features for subsequent long-term feature modeling and load prediction. Then, a prediction step size embedding vector is constructed using the charging load prediction step size, and multi-granularity load time-series combined features are determined through multi-granularity load time-series features. A preset spiral loop scanning state space processing algorithm can progressively scan and model long-term time-series features in a spiral manner, and a preset parallel... The ring scan state-space processing algorithm can scan and model long-term features in a parallel manner, thereby more accurately capturing the changing patterns of electric vehicle charging load over long periods. Combined with the prediction step size embedding vector, it can not only uncover cross-period, discontinuous long-distance jump correlations in the charging load sequence, achieving efficient long-term feature modeling and improving adaptability to the multi-period and highly volatile characteristics of charging load, but also achieve controllable dynamic charging load prediction based on the charging load prediction step size, enhancing the predictive adaptability to the randomness and volatility of electric vehicle charging load. Subsequently, by using the load prediction gating coefficient, it can adaptively fuse multi-time interval load sequences and long-term features of load prediction, improving the prediction accuracy of different charging load prediction step sizes, and achieving accurate prediction of electric vehicle charging load.

[0007] As a preferred embodiment, the step of acquiring historical charging load data of electric vehicles and constructing a multi-time interval load sequence by splicing the historical charging load data according to the interval time of the historical charging load data includes: acquiring historical charging load data of electric vehicles; dividing the historical charging load data according to the interval time of the historical charging load data to determine several long-time interval load sequences and several short-time interval load sequences; splicing the long-time interval load sequences and short-time interval load sequences one by one to obtain several multi-time interval spliced ​​load sequences; constructing a load sequence matrix based on the multi-time interval spliced ​​load sequences, and using the load sequence matrix as the multi-time interval load sequence.

[0008] This preferred scheme determines several long-interval load sequences and several short-interval load sequences by using the interval time of historical charging load data, and then performs a splicing operation. This allows the multi-interval load sequences to cover different periodic patterns of electric vehicle charging load, fully considers the characteristics of electric vehicle charging load at different time scales, and can more comprehensively capture the load change patterns, thereby improving the accuracy of subsequent load forecasting.

[0009] As a preferred embodiment, the step of extracting multi-granularity time-series features from the multi-time interval load sequence based on a preset multi-granularity cross-temporal feature extraction algorithm to obtain multi-granularity load temporal features includes: performing dimensional permutation on the multi-time interval load sequence to obtain a multi-time interval load transposed sequence; obtaining several mask matrices; and performing intra-group multi-granularity time-series feature extraction on the multi-time interval load transposed sequence based on the mask matrices to obtain intra-group multi-granularity time-series features; and obtaining several dilated volumes. The system employs a multi-layered approach, whereby the dilated convolutional layer extracts inter-group multi-granularity temporal features from the multi-time interval load transpose sequence to obtain inter-group multi-granularity temporal features. The intra-group and inter-group multi-granularity temporal features are then added element-wise to obtain multi-granularity load transpose features. The multi-granularity load transpose features are then dimensionally permuted to obtain initial multi-granularity load temporal features. Finally, the initial multi-granularity load temporal features are weighted and combined using a preset first fully connected layer to obtain further multi-granularity load temporal features.

[0010] This preferred scheme extracts intra-group multi-granularity temporal features using a mask matrix, focusing on the relationships between different granular features within the sequence. It then extracts inter-group multi-granularity temporal features using dilated convolutional layers, capturing temporal features between different groups. These features are then summed, dimension-permutated, and weighted to obtain multi-granularity load temporal features. This multi-granularity load temporal feature extraction uncovers the temporal features of multi-time interval load sequences from multiple perspectives, comprehensively considering intra-group and inter-group feature relationships. This results in a more comprehensive and accurate reflection of the changing patterns of electric vehicle charging load, thereby improving the accuracy of subsequent electric vehicle charging load prediction.

[0011] As a preferred embodiment, the step of obtaining several mask matrices and extracting intra-group multi-granularity temporal features from the multi-time interval load transpose sequence based on the mask matrices to obtain intra-group multi-granularity temporal features includes: obtaining several mask matrices; performing element-wise multiplication of each mask matrix with the multi-time interval load transpose sequence to obtain several mask load temporal features; performing convolution operation on the mask load temporal features based on a preset first one-dimensional convolutional layer to obtain mask load temporal convolutional features corresponding to each mask load temporal feature; performing convolution operation on the multi-time interval load transpose sequence based on the first one-dimensional convolutional layer to obtain multi-time interval load temporal convolutional features corresponding to the multi-time interval load transpose sequence; performing element-wise addition of the mask load temporal convolutional features and the multi-time interval load temporal convolutional features to obtain intra-group multi-granularity load temporal convolutional features; and performing convolution operation on the intra-group multi-granularity load temporal convolutional features based on a preset second one-dimensional convolutional layer to obtain intra-group multi-granularity temporal features.

[0012] This preferred scheme extracts the temporal features of the masked load by performing element-wise multiplication of the mask matrix with the multi-time-interval load transpose sequence. Then, convolution and addition operations are performed to obtain multi-granularity temporal features within the group. The mask matrix effectively extracts temporal features of different granularities within the group, and the convolution operation enhances the expressive power of the features. This allows for a more accurate capture of the multi-granularity variation patterns of the multi-time-interval load transpose sequence within the group, improving the accuracy and comprehensiveness of the multi-granularity temporal features within the group, thereby enhancing the accuracy and comprehensiveness of subsequent electric vehicle charging load prediction.

[0013] As a preferred embodiment, the step of obtaining several dilated convolutional layers and extracting inter-group multi-granularity temporal features from the multi-time interval load transpose sequence based on the dilated convolutional layers to obtain inter-group multi-granularity temporal features includes: obtaining several dilated convolutional layers, wherein the dilated convolutional layers have different dilation rates; performing convolution operations on the multi-time interval load transpose sequence based on each dilated convolutional layer to obtain several load temporal dilated convolutional features; performing element-wise addition operations on all the load temporal dilated convolutional features to obtain inter-group multi-granularity temporal dilated convolutional features; and performing convolution operations on the inter-group multi-granularity temporal dilated convolutional features based on a preset second one-dimensional convolutional layer to obtain inter-group multi-granularity temporal features.

[0014] This preferred scheme uses dilated convolutional layers with different dilation rates to perform convolution operations on multi-time interval load transpose sequences, obtaining several load time-series dilated convolutional features. Then, further addition and convolution operations are used to extract multi-granularity time-series features between groups. By using dilated convolutional layers with different dilation rates, the temporal information of different ranges in multi-time interval load transpose sequences can be comprehensively captured, thereby more accurately reflecting the complex relationship of load changes between different groups. This improves the accuracy and completeness of multi-granularity time-series features between groups, and thus improves the accuracy and comprehensiveness of subsequent electric vehicle charging load prediction.

[0015] As a preferred embodiment, the steps of obtaining the charging load prediction step size of the electric vehicle and constructing a prediction step size embedding vector based on the charging load prediction step size; determining multi-granularity load time series combination features based on the multi-granularity load time series features; performing long-term feature modeling on the multi-granularity load time series combination features according to a preset spiral ring scan state space processing algorithm and a preset parallel ring scan state space processing algorithm, combined with the prediction step size embedding vector, to obtain spiral ring scan load prediction long-term features and parallel ring scan load prediction long-term features; obtaining load prediction long-term features based on the spiral ring scan load prediction long-term features and parallel ring scan load prediction long-term features, including: obtaining the charging load prediction step size of the electric vehicle and constructing a prediction step size embedding vector based on the charging load prediction step size; and performing weighted combination of the multi-granularity load time series features based on a preset second fully connected layer. A multi-granularity load time-series combination feature is obtained; based on a preset spiral ring scan state space processing algorithm and the prediction step size embedding vector, long-term time-series feature modeling is performed on the multi-granularity load time-series combination feature to obtain spiral ring scan load prediction long-term time-series feature; based on a preset parallel ring scan state space processing algorithm and the prediction step size embedding vector, long-term time-series feature modeling is performed on the multi-granularity load time-series combination feature to obtain parallel ring scan load prediction long-term time-series feature; element-wise addition operation is performed on the spiral ring scan load prediction long-term time-series feature and the parallel ring scan load prediction long-term time-series feature to obtain initial load prediction long-term time-series feature; convolution operation is performed on the initial load prediction long-term time-series feature based on a preset third one-dimensional convolutional layer to obtain first load prediction long-term time-series feature; weighted combination is performed on the first load prediction long-term time-series feature based on a preset third fully connected layer to obtain load prediction long-term time-series feature.

[0016] This preferred scheme constructs a prediction step size embedding vector based on the electric vehicle charging load prediction step size, and weights and combines multi-granularity load time-series features to obtain multi-granularity load time-series combined features. Then, using the spiral ring scan state-space processing algorithm and the parallel ring scan state-space processing algorithm, long-term time-series feature modeling is performed based on the prediction step size embedding vector. This ensures that the long-term time-series feature modeling fully considers the prediction step size factor in actual electric vehicle charging load prediction, realizing controllable and dynamic charging load prediction, enhancing the predictive adaptability to the randomness and volatility of electric vehicle charging load, and more accurately capturing the changing patterns of electric vehicle charging load over long time series, thereby improving the accuracy of subsequent electric vehicle charging load prediction.

[0017] As a preferred embodiment, the step of modeling long-time features of the multi-granularity load time-series combination features based on the preset spiral ring scanning state space processing algorithm and the prediction step size embedding vector to obtain the spiral ring scanning load prediction long-time features includes: weighting the multi-granularity load time-series combination features based on the second fully connected layer to obtain initial multi-granularity load time-series combination features; performing a convolution operation on the initial multi-granularity load time-series combination features based on a preset fourth one-dimensional convolutional layer to obtain first multi-granularity load time-series combination features; performing a nonlinear transformation on the first multi-granularity load time-series combination features based on a preset first activation function to obtain third multi-granularity load time-series combination features; and performing a nonlinear transformation on the first multi-granularity load time-series combination features based on a preset first activation function. An activation function performs a nonlinear transformation on the initial multi-granularity load time-series combination features to obtain a fourth multi-granularity load time-series combination feature. Based on a preset spiral ring scan selective state-space model and combined with the prediction step size embedding vector, a long-term feature model is performed on the third multi-granularity load time-series combination features to obtain the initial spiral ring scan load prediction long-term feature. An element-wise multiplication operation is performed on the initial spiral ring scan load prediction long-term feature and the fourth multi-granularity load time-series combination feature to obtain the first spiral ring scan load prediction long-term feature. Based on the second fully connected layer, the first spiral ring scan load prediction long-term feature is weighted and combined to obtain the spiral ring scan load prediction long-term feature.

[0018] This preferred scheme employs weighted combination, convolution, and nonlinear transformation operations. Then, it uses a spiral ring scan selective state-space model combined with a prediction step-size embedding vector to model the long-term features of the third multi-granularity load time-series combination. Finally, it obtains the spiral ring scan load prediction long-term features through multiplication and weighted combination operations. The spiral ring scan selective state-space model can progressively scan and model long-term features in a spiral manner, thus more accurately capturing the changing patterns of electric vehicle charging load over long periods. Combined with the prediction step-size embedding vector, the long-term feature model fully considers the prediction step-size factor in actual electric vehicle charging load prediction, achieving controllable and dynamic charging load prediction, thereby improving the accuracy of subsequent electric vehicle charging load prediction.

[0019] As a preferred embodiment, the step of modeling long-time features of the multi-granularity load time-series combination features based on the preset parallel loop scanning state space processing algorithm and the prediction step size embedding vector to obtain the long-time features of parallel loop scanning load prediction includes: weighting the multi-granularity load time-series combination features based on the second fully connected layer to obtain initial multi-granularity load time-series combination features; performing a convolution operation on the initial multi-granularity load time-series combination features based on a preset fourth one-dimensional convolutional layer to obtain first multi-granularity load time-series combination features; performing a nonlinear transformation on the first multi-granularity load time-series combination features based on a preset first activation function to obtain third multi-granularity load time-series combination features; and performing a nonlinear transformation on the first multi-granularity load time-series combination features based on a preset first activation function. An activation function performs a nonlinear transformation on the initial multi-granularity load time-series combination features to obtain a fourth multi-granularity load time-series combination feature. Based on a preset parallel loop scan selective state-space model and combined with the prediction step size embedding vector, a long-term feature model is performed on the third multi-granularity load time-series combination features to obtain the initial parallel loop scan load prediction long-term feature. An element-wise multiplication operation is performed on the initial parallel loop scan load prediction long-term feature and the fourth multi-granularity load time-series combination feature to obtain the first parallel loop scan load prediction long-term feature. Based on the second fully connected layer, the first parallel loop scan load prediction long-term feature is weighted and combined to obtain the parallel loop scan load prediction long-term feature.

[0020] This preferred scheme employs weighted combination, convolution, and nonlinear transformation operations. Then, it uses a spiral ring scan selective state-space model combined with a prediction step-size embedding vector to model the long-term features of the third multi-granularity load time-series combination. Next, multiplication and weighted combination operations are used to obtain the spiral ring scan load prediction long-term features. The parallel ring scan selective state-space model enables parallel and progressive scanning and modeling of long-term features, thus more accurately capturing the changing patterns of electric vehicle charging load over long periods. Combined with the prediction step-size embedding vector, the long-term feature modeling fully considers the prediction step-size factor in actual electric vehicle charging load prediction, achieving controllable and dynamic charging load prediction, thereby improving the accuracy of subsequent electric vehicle charging load prediction.

[0021] As a preferred embodiment, the step of calculating the load prediction gating coefficient based on the multi-time interval load sequence and the long-term load prediction features, and fusing the multi-time interval load sequence and the long-term load prediction features based on the load prediction gating coefficient to determine the charging load prediction result of the electric vehicle, includes: performing a convolution operation on the multi-time interval load sequence based on a preset third one-dimensional convolutional layer to obtain multi-granularity load time-series convolutional features; performing a weighted combination of the multi-granularity load time-series convolutional features based on a preset third fully connected layer to obtain multi-granularity load time-series residual features; concatenating the multi-granularity load time-series residual features with the long-term load prediction features to obtain load prediction gating features; performing a weighted combination of the load prediction gating features based on the third fully connected layer to obtain load prediction gating combination features; performing a nonlinear transformation on the load prediction gating combination features based on a preset second activation function to obtain load prediction gating coefficients; and performing a weighted fusion of the multi-time interval load sequence and the long-term load prediction features based on the load prediction gating coefficients to obtain the charging load prediction result of the electric vehicle.

[0022] This preferred scheme first performs convolution and weighted combination of multi-time interval load sequences using fully connected layers to obtain multi-granularity load time-series residual features. These residual features are then concatenated with long-term load prediction feature channels and subjected to weighted combination and nonlinear transformation to obtain load prediction gating coefficients. Subsequently, the multi-time interval load sequences and long-term load prediction features are weighted and fused based on these gating coefficients to obtain the charging load prediction result. By calculating the gating coefficients, the weights of the multi-time interval load sequences and long-term load prediction features in the fusion process can be dynamically and adaptively adjusted. Features can be flexibly combined according to different load change scenarios, enabling the charging load prediction result to accurately reflect actual load changes and improving the accuracy and adaptability of load prediction.

[0023] Accordingly, this invention provides a step-size controllable load forecasting system based on multi-granularity time series, including: a multi-time interval load sequence construction module, a multi-granularity load time series feature extraction module, a long time series feature modeling module, and a charging load forecasting module; The multi-time interval load sequence construction module is used to acquire historical charging load data of electric vehicles, and to splice the historical charging load data according to the interval time of the historical charging load data to construct a multi-time interval load sequence. The multi-granularity load time-series feature extraction module is used to extract intra-group multi-granularity time-series features and inter-group multi-granularity time-series features from the multi-time interval load sequence based on a preset multi-granularity cross-time-series feature extraction algorithm, thereby obtaining multi-granularity load time-series features. The long-time-series feature modeling module is used to obtain the charging load prediction step size of the electric vehicle, construct a prediction step size embedding vector based on the charging load prediction step size, determine multi-granularity load time-series combination features based on the multi-granularity load time-series features, perform long-time-series feature modeling on the multi-granularity load time-series combination features according to the preset spiral ring scan state space processing algorithm and the preset parallel ring scan state space processing algorithm, combined with the prediction step size embedding vector, to obtain the spiral ring scan load prediction long-time-series features and the parallel ring scan load prediction long-time-series features, and obtain the load prediction long-time-series features based on the spiral ring scan load prediction long-time-series features and the parallel ring scan load prediction long-time-series features. The charging load prediction module is used to calculate the load prediction gating coefficient based on the multi-time interval load sequence and the long-term time series characteristics of load prediction; and to fuse the multi-time interval load sequence and the long-term time series characteristics of load prediction based on the load prediction gating coefficient to determine the charging load prediction result of the electric vehicle.

[0024] Understandably, this system stitches together historical charging load data by using the intervals between historical charging load data, enabling multi-time interval load sequences to cover different cyclical patterns of electric vehicle charging loads. Through intra-group and inter-group multi-granularity time-series feature extraction, it can extract and fuse time-series information of multiple granularities, transforming multi-time interval load sequences into multi-granularity load time-series features containing multi-cycle priors. This provides rich multi-granularity features for subsequent long-term feature modeling and load prediction. Then, it constructs a prediction step-size embedding vector based on the charging load prediction step size, determines multi-granularity load time-series combined features through multi-granularity load time-series features, and uses a preset spiral loop scanning state-space processing algorithm to progressively scan and model long-term time-series features in a spiral manner, using preset parallel... The ring scan state-space processing algorithm can scan and model long-term features in a parallel manner, thereby more accurately capturing the changing patterns of electric vehicle charging load over long periods. Combined with the prediction step size embedding vector, it can not only uncover cross-period, discontinuous long-distance jump correlations in the charging load sequence, achieving efficient long-term feature modeling and improving adaptability to the multi-period and highly volatile characteristics of charging load, but also achieve controllable dynamic charging load prediction based on the charging load prediction step size, enhancing the predictive adaptability to the randomness and volatility of electric vehicle charging load. Subsequently, by using the load prediction gating coefficient, it can adaptively fuse multi-time interval load sequences and long-term features of load prediction, improving the prediction accuracy of different charging load prediction step sizes, and achieving accurate prediction of electric vehicle charging load. Attached Figure Description

[0025] Figure 1 A flowchart illustrating the steps of a step-controllable load forecasting method based on multi-granularity time series provided in this embodiment of the invention; Figure 2 This is a schematic diagram of a multi-granularity cross-temporal feature extractor provided in an embodiment of the present invention; Figure 3 A schematic diagram of a step-size controllable hybrid ring scan state space processor provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a dual-source dynamic weight fusion device provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of a step-controllable load forecasting system based on multi-granularity time series, provided in an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Example 1 Please refer to Figure 1 , Figure 1 The flowchart of a step-size controllable load forecasting method based on multi-granularity time series provided in this embodiment of the invention includes steps S101 to S104.

[0028] Step S101: Obtain historical charging load data of electric vehicles, and splice the historical charging load data according to the interval time of the historical charging load data to construct a multi-time interval load sequence.

[0029] In this embodiment, the step of acquiring historical charging load data of electric vehicles and constructing a multi-time interval load sequence by splicing the historical charging load data according to the interval time of the historical charging load data includes: acquiring historical charging load data of electric vehicles; dividing the historical charging load data according to the interval time of the historical charging load data to determine several long-time interval load sequences and several short-time interval load sequences; splicing the long-time interval load sequences and short-time interval load sequences one by one to obtain several multi-time interval spliced ​​load sequences; constructing a load sequence matrix based on the multi-time interval spliced ​​load sequences, and using the load sequence matrix as the multi-time interval load sequence.

[0030] In one optional embodiment, historical charging load data of electric vehicles is obtained. This embodiment extracts historical charging load data from two different charging stations: two public charging stations and one workplace charging station. Then, a long time interval is set to 10 minutes, and a short time interval is... Based on the intervals of historical charging load data, several short-interval load sequences are extracted every few minutes. and several long time interval load sequences Furthermore, the short-interval load sequence and the long-interval load sequence satisfy the following: ; Indices representing short-interval load sequences and long-interval load sequences; Indicates an integer multiple of the existence of two short-interval loads (this can be set by the experimenter according to actual needs); then sets the combination of short-interval load sequences. Furthermore, by concatenating the long-interval load sequence and the short-interval load sequence one by one, it can be represented as: ; in, This indicates a splicing operation. This can be understood as being in Long time interval load sequence at time point ; Therefore, the load sequence matrix, i.e., the multi-time interval load sequence for: ; in, Represents a matrix format; The dimension is ,Right now .

[0031] This embodiment determines several long-time interval load sequences and several short-time interval load sequences by using the interval time of historical charging load data, and then performs a splicing operation so that the multi-time interval load sequences can cover different periodic patterns of electric vehicle charging load. It fully considers the characteristics of electric vehicle charging load at different time scales, can more comprehensively capture the load change pattern, and thus improves the accuracy of subsequent load forecasting.

[0032] It should be noted that the step-size controllable load prediction method based on multi-granularity time series provided in this embodiment is implemented based on a network model. Therefore, before constructing the multi-time interval load sequence, it also includes dividing the historical charging load data into training set, validation set, and test set, and then constructing the multi-time interval load sequences of the training set, validation set, and test set respectively. Since the training set, validation set, and test set are common techniques in network models, this embodiment will not describe them in detail here. Furthermore, this embodiment describes the relevant concepts of the structure involved in the subsequent network model. A one-dimensional convolutional layer (Conv1D) is a layer structure in convolutional neural networks specifically used to process sequential data. Its core is the convolution operation, which calculates the output by sliding a specific number of convolutional kernels on the input sequence and performing local dot products. Each convolutional kernel has a fixed size. The number of convolutional kernels directly determines the dimension of the output feature map of the layer. It is widely used in time series prediction, speech recognition, and natural language processing. A fully connected layer is a basic layer structure in neural networks. In this layer, each neuron is connected to all neurons in the previous layer, and its core operation is weighted combination. This layer takes each feature unit in the input vector, sums it with learnable weight parameters, and adds a bias term to form a completely new output feature. Essentially, this process combines all input features into new, higher-level features through linear weighting. The number of output features is a pre-defined hyperparameter that determines the dimensionality of the new features generated after weighted combination. A selective state space model (SSL) is a deep learning architecture designed for efficiently processing long sequences of data. Its core innovation lies in introducing a selection mechanism, enabling the model to dynamically decide how to update its internal state and convey information based on the current input, thereby achieving data-dependent, context-aware reasoning. The Spiral Ring Scanning Selective State Space Model, also known as the Selective State Space Model with Spiral Ring Scanning, introduces a more advanced spiral ring scanning path on the basis of the Selective State Space Model to perform more refined long-term feature modeling. Its unique spiral scanning method also traverses and fuses sequences at multiple scales, thereby enabling the parallel capture of complex patterns ranging from local short-term interactions to global long-term dependencies.The parallel ring scan selective state-space model, also known as the selective state space model with parallel ring scanning, achieves ultra-long time-series feature modeling through a combination of selective mechanisms and parallel ring scans. Element-wise addition is a basic operation that adds corresponding values ​​in two matrices, vectors, or tensors with identical shapes, generating a new result that retains the original shape. Element-wise multiplication is defined as multiplying corresponding elements in two matrices, vectors, or tensors with identical shapes, generating a new data structure that retains the original shape.

[0033] In an optional embodiment, for training the network model, this embodiment uses the MSE loss function as the loss function during the training process, and utilizes the backpropagation of deep learning to narrow the gap between the predicted load value and the actual load value, thereby optimizing the network model.

[0034] Step S102: Based on the preset multi-granularity cross-time series feature extraction algorithm, perform intra-group multi-granularity time series feature extraction and inter-group multi-granularity time series feature extraction on the multi-time interval load sequence to obtain multi-granularity load time series features.

[0035] In this embodiment, the step of extracting multi-granularity cross-temporal feature based on a preset multi-granularity cross-temporal feature extraction algorithm to perform intra-group multi-granularity temporal feature extraction and inter-group multi-granularity temporal feature extraction on the multi-time interval load sequence to obtain multi-granularity load temporal features includes: performing dimensional permutation on the multi-time interval load sequence to obtain a multi-time interval load transposed sequence; obtaining several mask matrices, and performing intra-group multi-granularity temporal feature extraction on the multi-time interval load transposed sequence based on the mask matrices to obtain intra-group multi-granularity temporal features; and obtaining several dilated volumes. The system employs a multi-layered approach, whereby the dilated convolutional layer extracts inter-group multi-granularity temporal features from the multi-time interval load transpose sequence to obtain inter-group multi-granularity temporal features. The intra-group and inter-group multi-granularity temporal features are then added element-wise to obtain multi-granularity load transpose features. The multi-granularity load transpose features are then dimensionally permuted to obtain initial multi-granularity load temporal features. Finally, the initial multi-granularity load temporal features are weighted and combined using a preset first fully connected layer to obtain further multi-granularity load temporal features.

[0036] This embodiment extracts intra-group multi-granularity temporal features using a mask matrix, focusing on the relationships between different granular features within the sequence. It then extracts inter-group multi-granularity temporal features using dilated convolutional layers, capturing temporal features between different groups. These features are then added, dimension-permutated, and weighted to obtain multi-granularity load temporal features. This multi-granularity load temporal feature extraction uncovers the temporal features of multi-time interval load sequences from multiple perspectives, comprehensively considering intra-group and inter-group feature relationships. This results in a more comprehensive and accurate reflection of the changing patterns of electric vehicle charging load, thereby improving the accuracy of subsequent electric vehicle charging load prediction.

[0037] In this embodiment, the step of obtaining several mask matrices and extracting intra-group multi-granularity temporal features from the multi-time interval load transpose sequence based on the mask matrices to obtain intra-group multi-granularity temporal features includes: obtaining several mask matrices; performing element-wise multiplication of each mask matrix with the multi-time interval load transpose sequence to obtain several mask load temporal features; performing convolution operation on the mask load temporal features based on a preset first one-dimensional convolutional layer to obtain mask load temporal convolutional features corresponding to each mask load temporal feature; performing convolution operation on the multi-time interval load transpose sequence based on the first one-dimensional convolutional layer to obtain multi-time interval load temporal convolutional features corresponding to the multi-time interval load transpose sequence; performing element-wise addition of the mask load temporal convolutional features and the multi-time interval load temporal convolutional features to obtain intra-group multi-granularity load temporal convolutional features; and performing convolution operation on the intra-group multi-granularity load temporal convolutional features based on a preset second one-dimensional convolutional layer to obtain intra-group multi-granularity temporal features.

[0038] This embodiment extracts the temporal features of the masked load by performing element-wise multiplication of the mask matrix with the multi-time-interval load transpose sequence. Then, convolution and addition operations are performed to obtain multi-granularity temporal features within the group. The mask matrix effectively extracts temporal features of different granularities within the group, and the convolution operation enhances the expressive power of the features. This allows for a more accurate capture of the multi-granularity variation patterns of the multi-time-interval load transpose sequence within the group, improving the accuracy and comprehensiveness of the multi-granularity temporal features within the group, thereby enhancing the accuracy and comprehensiveness of subsequent electric vehicle charging load prediction.

[0039] In this embodiment, the step of obtaining several dilated convolutional layers and extracting inter-group multi-granularity temporal features from the multi-time interval load transpose sequence based on the dilated convolutional layers to obtain inter-group multi-granularity temporal features includes: obtaining several dilated convolutional layers, wherein the dilated convolutional layers have different dilation rates; performing convolution operations on the multi-time interval load transpose sequence based on each dilated convolutional layer to obtain several load temporal dilated convolutional features; performing element-wise addition operations on all the load temporal dilated convolutional features to obtain inter-group multi-granularity temporal dilated convolutional features; and performing convolution operations on the inter-group multi-granularity temporal dilated convolutional features based on a preset second one-dimensional convolutional layer to obtain inter-group multi-granularity temporal features.

[0040] This embodiment uses dilated convolutional layers with different hole rates to perform convolution operations on multi-time interval load transpose sequences, obtaining several load time-series dilated convolutional features. Then, further addition and convolution operations are used to extract multi-granularity time-series features between groups. By using dilated convolutional layers with different hole rates, the temporal information of different ranges in multi-time interval load transpose sequences can be comprehensively captured, thereby more accurately reflecting the complex relationship of load changes between different groups. This improves the accuracy and completeness of multi-granularity time-series features between groups, and thus improves the accuracy and comprehensiveness of subsequent electric vehicle charging load prediction.

[0041] In one alternative embodiment, please refer to Figure 2 , Figure 2 This is a schematic diagram of a multi-granularity cross-temporal feature extractor provided in an embodiment of the present invention. Figure 2 As shown, this embodiment sets up a multi-granularity cross-temporal feature extractor in the network model. Multi-granularity cross-temporal feature extractor This includes intra-group multi-granularity temporal feature extraction branches and inter-group multi-granularity temporal feature extraction branches; in this embodiment, step S102 is implemented through intra-group multi-granularity temporal feature extraction branches and inter-group multi-granularity temporal feature extraction branches; specifically, for multi-time interval load sequences... Dimensional permutation is performed to obtain a multi-time interval load transpose sequence. Then, multi-granularity temporal feature extraction within the group is performed through a branch. Specifically, several mask matrices are obtained. In this embodiment, two mask matrices are set, namely the first mask matrix. Second mask matrix In the first mask matrix Second mask matrix In the first mask matrix, the element value is either 0 or 1, elements in the same row have the same value, and the elements in the last row all have the value 1; The first row contains all 1s, the second row contains all 0s, and so on; the second mask matrix The first row contains all 1s, the second and third rows contain all 0s; the fourth row contains all 1s; the fifth and sixth rows contain all 0s, and so on, forming the second mask matrix. Rows with an element value of 1 are separated by two rows; then the first mask matrix is... Second mask matrix , respectively with multi-time interval load transpose sequence Perform element-wise multiplication to obtain the masked payload time-series characteristics. and Then, the number of convolutional kernels in the first one-dimensional convolutional layer is set to... The kernel size is 3; the temporal features of the masked payload are processed through the first one-dimensional convolutional layer. and Perform convolution operations separately to obtain the temporal features of the masked payload. and The corresponding masked load temporal convolutional features; and the multi-time interval load transpose sequence through the first one-dimensional convolutional layer. Convolution operations are performed to obtain multi-time interval load temporal convolutional features; then, the masked load temporal convolutional features and the multi-time interval load temporal convolutional features are added element-wise to obtain intra-group multi-granularity load temporal convolutional features; finally, the number of convolutional kernels in the preset second one-dimensional convolutional layer is set to a certain value. The kernel size is 3. Multi-granularity temporal features within the group are obtained by performing convolution operations on the multi-granularity load temporal convolutional features within the group. Then, several dilated convolutional layers are obtained. In this embodiment, three dilated convolutional layers are set, namely the first dilated convolutional layer with a dilation rate of 1. The second porous convolutional layer with a porosity of 2 and a third voided convolutional layer with a void ratio of 3 The number of kernels in the three dilated convolutional layers is 1. The kernel size is 3 for all convolutional layers; then the convolutional layers are passed through the first dilated convolutional layer. Convolutional operations are performed on multi-time-interval load transpose sequences to obtain load temporal dilated convolutional features. Through the second hole convolution layer Convolutional operations are performed on multi-time-interval load transpose sequences to obtain load temporal dilated convolutional features. Through the third hole convolution layer Convolutional operations are performed on multi-time-interval load transpose sequences to obtain load temporal dilated convolutional features. Then, the temporal dilated convolution features of the load were analyzed. , and Element-wise addition is performed to obtain inter-group multi-granularity temporal dilated convolutional features; then, a second one-dimensional convolutional layer is used to convolve the intra-group multi-granularity load temporal features to obtain intra-group multi-granularity temporal features; then, element-wise addition is performed on the intra-group and inter-group multi-granularity temporal features to obtain multi-granularity load temporal transpose features; finally, the dimensionality of the multi-granularity load temporal transpose features is permuted to obtain the initial multi-granularity load temporal features; the output feature of the first fully connected layer is set to... The initial multi-granularity load time series features are weighted and combined through the first fully connected layer to obtain multi-granularity load time series features.

[0042] Step S103: Obtain the charging load prediction step size of the electric vehicle, and construct a prediction step size embedding vector based on the charging load prediction step size; determine the multi-granularity load time series combination features based on the multi-granularity load time series features; perform long-term feature modeling on the multi-granularity load time series combination features according to the preset spiral ring scan state space processing algorithm and the preset parallel ring scan state space processing algorithm, combined with the prediction step size embedding vector, to obtain the spiral ring scan load prediction long-term feature and the parallel ring scan load prediction long-term feature; obtain the load prediction long-term feature based on the spiral ring scan load prediction long-term feature and the parallel ring scan load prediction long-term feature.

[0043] In this embodiment, the steps of obtaining the charging load prediction step size of the electric vehicle and constructing a prediction step size embedding vector based on the charging load prediction step size; determining multi-granularity load time series combination features based on the multi-granularity load time series features; performing long-term feature modeling on the multi-granularity load time series combination features according to a preset spiral ring scan state space processing algorithm and a preset parallel ring scan state space processing algorithm, combined with the prediction step size embedding vector, to obtain spiral ring scan load prediction long-term features and parallel ring scan load prediction long-term features; obtaining load prediction long-term features based on the spiral ring scan load prediction long-term features and parallel ring scan load prediction long-term features, including: obtaining the charging load prediction step size of the electric vehicle and constructing a prediction step size embedding vector based on the charging load prediction step size; and performing weighted combination of the multi-granularity load time series features based on a preset second fully connected layer. A multi-granularity load time-series combination feature is obtained; based on a preset spiral ring scan state space processing algorithm and the prediction step size embedding vector, long-term time-series feature modeling is performed on the multi-granularity load time-series combination feature to obtain spiral ring scan load prediction long-term time-series feature; based on a preset parallel ring scan state space processing algorithm and the prediction step size embedding vector, long-term time-series feature modeling is performed on the multi-granularity load time-series combination feature to obtain parallel ring scan load prediction long-term time-series feature; element-wise addition operation is performed on the spiral ring scan load prediction long-term time-series feature and the parallel ring scan load prediction long-term time-series feature to obtain initial load prediction long-term time-series feature; convolution operation is performed on the initial load prediction long-term time-series feature based on a preset third one-dimensional convolutional layer to obtain first load prediction long-term time-series feature; weighted combination is performed on the first load prediction long-term time-series feature based on a preset third fully connected layer to obtain load prediction long-term time-series feature.

[0044] This embodiment constructs a prediction step size embedding vector based on the electric vehicle charging load prediction step size, and weights and combines multi-granularity load time-series features to obtain multi-granularity load time-series combined features. Then, using the spiral ring scan state-space processing algorithm and the parallel ring scan state-space processing algorithm, long-term time-series feature modeling is performed based on the prediction step size embedding vector. This ensures that the long-term time-series feature modeling fully considers the prediction step size factor in actual electric vehicle charging load prediction, realizing controllable and dynamic charging load prediction, enhancing the predictive adaptability to the randomness and volatility of electric vehicle charging load, and more accurately capturing the changing patterns of electric vehicle charging load over long time series, thereby improving the accuracy of subsequent electric vehicle charging load prediction.

[0045] In this embodiment, the step of modeling long-term features of the multi-granularity load time-series combination features based on the preset spiral ring scanning state space processing algorithm and the prediction step size embedding vector to obtain the spiral ring scanning load prediction long-term features includes: weighting the multi-granularity load time-series combination features based on the second fully connected layer to obtain the initial multi-granularity load time-series combination features; performing a convolution operation on the initial multi-granularity load time-series combination features based on the preset fourth one-dimensional convolutional layer to obtain the first multi-granularity load time-series combination features; performing a nonlinear transformation on the first multi-granularity load time-series combination features based on the preset first activation function to obtain the third multi-granularity load time-series combination features; and performing a nonlinear transformation on the first multi-granularity load time-series combination features based on the preset first activation function. An activation function performs a nonlinear transformation on the initial multi-granularity load time-series combination features to obtain a fourth multi-granularity load time-series combination feature. Based on a preset spiral ring scan selective state-space model and combined with the prediction step size embedding vector, a long-term feature model is performed on the third multi-granularity load time-series combination features to obtain the initial spiral ring scan load prediction long-term feature. An element-wise multiplication operation is performed on the initial spiral ring scan load prediction long-term feature and the fourth multi-granularity load time-series combination feature to obtain the first spiral ring scan load prediction long-term feature. Based on the second fully connected layer, the first spiral ring scan load prediction long-term feature is weighted and combined to obtain the spiral ring scan load prediction long-term feature.

[0046] This embodiment employs weighted combination, convolution, and nonlinear transformation operations. Then, it uses a spiral ring scan selective state-space model combined with a prediction step-size embedding vector to model the long-term features of the third multi-granularity load time-series combination. Finally, it obtains the spiral ring scan load prediction long-term features through multiplication and weighted combination operations. The spiral ring scan selective state-space model can progressively scan and model long-term features in a spiral manner, thereby more accurately capturing the changing patterns of electric vehicle charging load over long periods. Combined with the prediction step-size embedding vector, the long-term feature modeling fully considers the prediction step-size factor in actual electric vehicle charging load prediction, achieving controllable dynamic charging load prediction and thus improving the accuracy of subsequent electric vehicle charging load prediction.

[0047] In this embodiment, the step of modeling long-term features of the multi-granularity load time-series combination features based on the preset parallel loop scan state space processing algorithm and the prediction step size embedding vector to obtain the parallel loop scan load prediction long-term features includes: weighting the multi-granularity load time-series combination features based on the second fully connected layer to obtain the initial multi-granularity load time-series combination features; performing a convolution operation on the initial multi-granularity load time-series combination features based on the preset fourth one-dimensional convolutional layer to obtain the first multi-granularity load time-series combination features; performing a nonlinear transformation on the first multi-granularity load time-series combination features based on the preset first activation function to obtain the third multi-granularity load time-series combination features; and performing a nonlinear transformation on the first multi-granularity load time-series combination features based on the preset first activation function. An activation function performs a nonlinear transformation on the initial multi-granularity load time-series combination features to obtain a fourth multi-granularity load time-series combination feature. Based on a preset parallel loop scan selective state-space model and combined with the prediction step size embedding vector, a long-term feature model is performed on the third multi-granularity load time-series combination features to obtain the initial parallel loop scan load prediction long-term feature. An element-wise multiplication operation is performed on the initial parallel loop scan load prediction long-term feature and the fourth multi-granularity load time-series combination feature to obtain the first parallel loop scan load prediction long-term feature. Based on the second fully connected layer, the first parallel loop scan load prediction long-term feature is weighted and combined to obtain the parallel loop scan load prediction long-term feature.

[0048] This embodiment employs weighted combination, convolution, and nonlinear transformation operations. Then, it uses a spiral ring scan selective state-space model combined with a prediction step-size embedding vector to model the long-term features of the third multi-granularity load time-series combination. Next, multiplication and weighted combination operations are used to obtain the spiral ring scan load prediction long-term features. The parallel ring scan selective state-space model enables parallel and progressive scanning and modeling of long-term features, thus more accurately capturing the changing patterns of electric vehicle charging load over long periods. Combined with the prediction step-size embedding vector, the long-term feature modeling fully considers the prediction step-size factor in actual electric vehicle charging load prediction, achieving controllable and dynamic charging load prediction, thereby improving the accuracy of subsequent electric vehicle charging load prediction.

[0049] In one alternative embodiment, please refer to Figure 3 , Figure 3This is a schematic diagram of a step-size controllable hybrid ring scan state space processor provided in an embodiment of the present invention. The step-size controllable hybrid ring scan state space processor consists of a second fully connected layer, a state space processing layer based on spiral ring scan, a state space processing layer based on parallel ring scan, a third one-dimensional convolutional layer, and a third fully connected layer. The state space processing layer based on spiral ring scan is used to obtain long-term features of spiral ring scan load prediction; the state space processing layer based on parallel ring scan is used to obtain long-term features of parallel ring scan load prediction. In the state space processing layers based on spiral ring scan and parallel ring scan, except that the selective state space models of spiral ring scan and parallel ring scan are inconsistent, the other layer connections are the same, and the feature dimension of each time step is 5. In this embodiment, the long-term feature modeling in step S103 is achieved through the step-size controllable hybrid ring scan state space processor to obtain the long-term features of load prediction; specifically, the step size of the charging load prediction for electric vehicles is obtained. Then, a learnable neural network embedding layer is used to map the step size values ​​into low-dimensional continuous prediction step size embedding vectors. Then, the output characteristic of the second fully connected layer is set to 5; the multi-granularity load time series characteristics are processed through the second fully connected layer. A weighted combination is performed to obtain multi-granularity load time series combination features. Then, the multi-granularity load time series combination features are weighted and combined again through a second fully connected layer to obtain initial multi-granularity load time series combination features. The fourth one-dimensional convolutional layer is set to have 5 kernels and a kernel size of 3. The initial multi-granularity load time series combination features are convolved through the fourth one-dimensional convolutional layer to obtain first multi-granularity load time series combination features. The first activation function is set to the SiLu activation function. The first multi-granularity load time series combination features are nonlinearly transformed using the SiLu activation function to obtain third multi-granularity load time series combination features. Finally, the initial multi-granularity load time series combination features are nonlinearly transformed using the SiLu activation function to obtain fourth multi-granularity load time series combination features. Then, the third multi-granularity load time series combination feature and the prediction step size embedding vector are jointly input into the spiral ring scan selective state space model, so that the spiral ring scan selective state space model can perform long-term feature modeling on the third multi-granularity load time series combination feature to obtain the initial spiral ring scan load prediction long-term feature; the initial spiral ring scan load prediction long-term feature and the fourth multi-granularity load time series combination feature are multiplied element-wise to obtain the first spiral ring scan load prediction long-term feature; the first spiral ring scan load prediction long-term feature is weighted and combined based on the second fully connected layer to obtain the spiral ring scan load prediction long-term feature; Then, the third multi-granularity load time series combination feature and the prediction step size embedding vector are jointly input into the parallel loop scan selective state space model, so that the spiral loop scan selective state space model can model the long-term features of the third multi-granularity load time series combination feature to obtain the initial spiral loop scan load prediction long-term features; then, the initial parallel loop scan load prediction long-term features and the fourth multi-granularity load time series combination feature are multiplied element-wise to obtain the first parallel loop scan load prediction long-term features; the first parallel loop scan load prediction long-term features are weighted and combined based on the second fully connected layer to obtain the parallel loop scan load prediction long-term features; then, the spiral loop scan load prediction long-term features and the parallel loop scan load prediction long-term features are added element-wise to obtain the initial load prediction long-term features; Set the number of convolution kernels in the third one-dimensional convolutional layer to... The initial load prediction long-term features are convolved using a third one-dimensional convolutional layer with a kernel size of 3 to obtain the first load prediction long-term features. The output feature of the third fully connected layer is set to 1, and the first load prediction long-term features are weighted and combined using the third fully connected layer to obtain the final load prediction long-term features. ; Furthermore, since both the spiral ring scan selective state space model and the parallel ring scan selective state space model are based on the selective state space model, their operational processes follow the following formulas (1) to (3). (1); (2); (3); In the public notices (1) to (3), , and It is a discrete matrix that has been preserved to zero order. To predict the step-size embedding vector, Indicates the charging load prediction step size Embed the adjusted parameter matrix; This represents the third multi-granularity load temporal combination feature of the input spiral ring scan selective state-space model or parallel ring scan selective state-space model; and These represent the original state and the current updated state, respectively. This indicates the long-term time-series characteristics of load prediction for initial spiral loop scanning or initial parallel loop scanning. The difference between the spiral loop scan selective state-space model and the parallel loop scan selective state-space model lies in their scanning methods; this embodiment further explains the scanning methods of these two models; assuming the third multi-granularity load time-series combination feature is input to the spiral loop scan selective state-space model or the parallel loop scan selective state-space model. For the spiral ring scanning selective state-space model, its scanning order is as follows: arrive , arrive , arrive This spiral loop pattern continues until the feature scan at each time step is completed; for the parallel loop scanning selective state-space model, if... If it is even, then let satisfy Therefore, the scanning order is as follows: arrive , arrive , arrive , arrive This parallel loop pattern continues until the feature scan at each time step is complete; if If it is an odd number, then let satisfy Therefore, the scanning order is as follows: arrive , arrive , arrive , arrive Scanning in this parallel loop pattern, finally by arrive Scan complete.

[0050] Step S104: Calculate the load forecasting gating coefficient based on the multi-time interval load sequence and the long-term time series characteristics of load forecasting; and fuse the multi-time interval load sequence and the long-term time series characteristics of load forecasting based on the load forecasting gating coefficient to determine the charging load forecasting result of the electric vehicle.

[0051] In this embodiment, the step of calculating the load prediction gating coefficient based on the multi-time interval load sequence and the long-term load prediction features, and fusing the multi-time interval load sequence and the long-term load prediction features based on the load prediction gating coefficient to determine the charging load prediction result of the electric vehicle, includes: performing a convolution operation on the multi-time interval load sequence based on a preset third one-dimensional convolutional layer to obtain multi-granularity load time-series convolutional features; performing a weighted combination of the multi-granularity load time-series convolutional features based on a preset third fully connected layer to obtain multi-granularity load time-series residual features; concatenating the multi-granularity load time-series residual features with the long-term load prediction features to obtain load prediction gating features; performing a weighted combination of the load prediction gating features based on the third fully connected layer to obtain load prediction gating combination features; performing a nonlinear transformation on the load prediction gating combination features based on a preset second activation function to obtain load prediction gating coefficients; and performing a weighted fusion of the multi-time interval load sequence and the long-term load prediction features based on the load prediction gating coefficients to obtain the charging load prediction result of the electric vehicle.

[0052] In one alternative embodiment, please refer to Figure 4 , Figure 4 This is a schematic diagram of a dual-source dynamic weighted fusion device provided in an embodiment of the present invention. In this embodiment, the load forecasting threshold coefficient is calculated in step S104 using the dual-source dynamic weighted fusion device to obtain long-term load forecasting features. Based on the load forecasting threshold coefficient, the multi-time interval load sequence and the long-term load forecasting features are fused to determine the charging load forecasting result for electric vehicles. Specifically, as shown... Figure 4 As shown, the multi-time interval load sequence is processed by a third one-dimensional convolutional layer. Convolution operations are performed to obtain multi-granularity load temporal convolutional features; then, the multi-granularity load temporal convolutional features are weighted and combined through a third fully connected layer to obtain multi-granularity load temporal residual features. Then, the multi-granularity load time-series residual characteristics are... With the long-term characteristics of the load forecast Channel splicing is performed to obtain load forecasting gating features; then, the load forecasting gating features are weighted and combined through a third fully connected layer to obtain load forecasting gating combined features; the second activation function is set to the sigmoid activation function, and the load forecasting gating combined features are nonlinearly transformed using the sigmoid activation function to obtain the load forecasting gating coefficients. Then, based on the load forecasting gating coefficient, the multi-time interval load sequence and the long-term time series features of load forecasting are weighted and fused to obtain the charging load forecasting result of the electric vehicle. Multi-granularity load time-series residual characteristics The calculation process is shown in formula (4), load forecasting threshold coefficient The calculation process is shown in formula (5); the charging load prediction results The calculation process is shown in formula (6); (4); (5); (6); In formulas (4) to (6), It is a multi-time interval load sequence. and All of these are weight matrices, and their values ​​can be customized by technical personnel according to actual needs. and All of these are weights, and their values ​​can be customized by technical personnel according to actual needs. This represents the temporal residual characteristics of multi-granularity loads; Indicates channel splicing; This represents the sigmoid activation function; For load forecasting threshold coefficient; This refers to the charging load forecast results; This indicates element-wise multiplication; This indicates the long-term characteristics of load forecasting.

[0053] This embodiment first performs convolution and weighted combination of multi-time interval load sequences using fully connected layers to obtain multi-granularity load time-series residual features. These residual features are then concatenated with long-term load prediction feature channels and subjected to weighted combination and nonlinear transformation to obtain load prediction gating coefficients. Subsequently, the multi-time interval load sequences and long-term load prediction features are weighted and fused based on the gating coefficients to obtain the charging load prediction result. By calculating the gating coefficients, the weights of the multi-time interval load sequences and long-term load prediction features in the fusion process can be dynamically and adaptively adjusted. Features can be flexibly combined according to different load change situations, so that the charging load prediction result can accurately reflect the actual load changes, improving the accuracy and adaptability of load prediction.

[0054] This embodiment stitches together historical charging load data by using time intervals, enabling multi-time interval load sequences to cover different cyclical patterns of electric vehicle charging loads. Through intra-group and inter-group multi-granularity time-series feature extraction, it extracts and fuses time-series information at multiple granularities, transforming the multi-time interval load sequences into multi-granularity load time-series features containing multi-cycle priors. This provides rich multi-granularity features for subsequent long-term feature modeling and load prediction. Then, it constructs a prediction step-size embedding vector using the charging load prediction step size, determines multi-granularity load time-series combination features using these features, and uses a preset spiral loop scanning state-space processing algorithm to progressively scan and model long-term time-series features in a spiral manner, while also using a preset parallel loop scanning algorithm. The state-space processing algorithm can scan and model long-term features in a parallel manner, thereby more accurately capturing the changing patterns of electric vehicle charging load over long periods. Combined with the prediction step size embedding vector, it can not only uncover cross-period, discontinuous long-distance jump correlations in the charging load sequence, achieving efficient long-term feature modeling and improving adaptability to the multi-period and highly volatile characteristics of charging load, but also achieve controllable dynamic charging load prediction based on the charging load prediction step size, enhancing the predictability adaptability to the randomness and volatility of electric vehicle charging load. Subsequently, by using the load prediction gating coefficient, it can adaptively fuse multi-time interval load sequences and long-term features of load prediction, improving the prediction accuracy of different charging load prediction step sizes, and achieving accurate prediction of electric vehicle charging load.

[0055] Example 2 Please refer to Figure 5 , Figure 5 A schematic diagram of a step-size controllable load forecasting system based on multi-granularity time series provided in an embodiment of the present invention includes: a multi-time interval load sequence construction module 201, a multi-granularity load time series feature extraction module 202, a long time series feature modeling module 203, and a charging load forecasting module 204; The multi-time interval load sequence construction module 201 is used to acquire historical charging load data of electric vehicles, and to splice the historical charging load data according to the interval time of the historical charging load data to construct a multi-time interval load sequence.

[0056] The multi-granularity load time-series feature extraction module 202 is used to extract intra-group multi-granularity time-series features and inter-group multi-granularity time-series features from the multi-time interval load sequence based on a preset multi-granularity cross-time-series feature extraction algorithm, thereby obtaining multi-granularity load time-series features.

[0057] The long-time-series feature modeling module 203 is used to obtain the charging load prediction step size of the electric vehicle, construct a prediction step size embedding vector based on the charging load prediction step size; determine the multi-granularity load time-series combination features based on the multi-granularity load time-series features; perform long-time-series feature modeling on the multi-granularity load time-series combination features according to the preset spiral ring scan state space processing algorithm and the preset parallel ring scan state space processing algorithm, combined with the prediction step size embedding vector, to obtain the spiral ring scan load prediction long-time-series features and the parallel ring scan load prediction long-time-series features; and obtain the load prediction long-time-series features based on the spiral ring scan load prediction long-time-series features and the parallel ring scan load prediction long-time-series features.

[0058] The charging load prediction module 204 is used to calculate the load prediction gating coefficient based on the multi-time interval load sequence and the long-term time series characteristics of load prediction; and to fuse the multi-time interval load sequence and the long-term time series characteristics of load prediction based on the load prediction gating coefficient to determine the charging load prediction result of the electric vehicle.

[0059] In this embodiment, the multi-time interval load sequence construction module 201 includes: a multi-time interval load sequence construction unit; the multi-time interval load sequence construction unit is used to acquire historical charging load data of electric vehicles; based on the interval time of the historical charging load data, the historical charging load data is divided to determine several long-time interval load sequences and several short-time interval load sequences; the long-time interval load sequences and short-time interval load sequences are concatenated one by one to obtain several multi-time interval concatenated load sequences; based on the multi-time interval concatenated load sequences, a load sequence matrix is ​​constructed, and the load sequence matrix is ​​used as the multi-time interval load sequence.

[0060] In this embodiment, the multi-granularity load time-series feature extraction module 202 includes: a multi-granularity load time-series feature extraction unit; the multi-granularity load time-series feature extraction unit is used to perform dimensional permutation on the multi-time interval load sequence to obtain a multi-time interval load transposed sequence; acquire several mask matrices, and perform intra-group multi-granularity time-series feature extraction on the multi-time interval load transposed sequence based on the mask matrices to obtain intra-group multi-granularity time-series features; acquire several dilated convolutional layers, and perform inter-group multi-granularity time-series feature extraction on the multi-time interval load transposed sequence based on the dilated convolutional layers to obtain inter-group multi-granularity time-series features; perform element-wise addition operation on the intra-group multi-granularity time-series features and inter-group multi-granularity time-series features to obtain multi-granularity load time-series transposed features; perform dimensional permutation on the multi-granularity load time-series transposed features to obtain initial multi-granularity load time-series features; and perform weighted combination of the initial multi-granularity load time-series features based on a preset first fully connected layer to obtain multi-granularity load time-series features.

[0061] In this embodiment, the multi-granularity load time-series feature extraction unit includes: an intra-group multi-granularity time-series feature extraction subunit; the intra-group multi-granularity time-series feature extraction subunit is used to obtain several mask matrices; perform element-wise multiplication of each mask matrix with the multi-time-interval load transpose sequence to obtain several mask load time-series features; perform convolution operation on the mask load time-series features based on a preset first one-dimensional convolutional layer to obtain mask load time-series convolutional features corresponding to each mask load time-series feature; perform convolution operation on the multi-time-interval load transpose sequence based on the first one-dimensional convolutional layer to obtain multi-time-interval load time-series convolutional features corresponding to the multi-time-interval load transpose sequence; perform element-wise addition of the mask load time-series convolutional features and the multi-time-interval load time-series convolutional features to obtain intra-group multi-granularity load time-series convolutional features; perform convolution operation on the intra-group multi-granularity load time-series convolutional features based on a preset second one-dimensional convolutional layer to obtain intra-group multi-granularity time-series features.

[0062] In this embodiment, the multi-granularity load temporal feature extraction unit includes: an inter-group multi-granularity temporal feature extraction subunit; the inter-group multi-granularity temporal feature extraction subunit is used to obtain several dilated convolutional layers, wherein the dilated convolutional layers have different dilation rates; based on each dilated convolutional layer, convolution operations are performed on the multi-time interval load transpose sequence to obtain several load temporal dilated convolutional features; all the load temporal dilated convolutional features are added element-wise to obtain inter-group multi-granularity temporal dilated convolutional features; based on a preset second one-dimensional convolutional layer, convolution operations are performed on the inter-group multi-granularity temporal dilated convolutional features to obtain inter-group multi-granularity temporal features.

[0063] In this embodiment, the long-time-series feature modeling module 203 includes: a long-time-series feature modeling unit; the long-time-series feature modeling unit is used to obtain the charging load prediction step size of the electric vehicle, and construct a prediction step size embedding vector based on the charging load prediction step size; weighted combination of the multi-granularity load time-series features based on a preset second fully connected layer to obtain multi-granularity load time-series combined features; long-time-series feature modeling of the multi-granularity load time-series combined features according to a preset spiral ring scan state space processing algorithm and the prediction step size embedding vector to obtain spiral ring scan load prediction long-time-series features; and long-time-series feature modeling of the multi-granularity load time-series combined features according to a preset parallel... The ring scan state space processing algorithm and the prediction step size embedding vector are used to model the long-term features of the multi-granularity load time series combination features to obtain the parallel ring scan load prediction long-term features; the spiral ring scan load prediction long-term features and the parallel ring scan load prediction long-term features are added element-wise to obtain the initial load prediction long-term features; the initial load prediction long-term features are convolved based on a preset third one-dimensional convolutional layer to obtain the first load prediction long-term features; the first load prediction long-term features are weighted and combined based on a preset third fully connected layer to obtain the load prediction long-term features.

[0064] In this embodiment, the long-time-series feature modeling unit includes: a spiral loop long-time-series feature modeling subunit; the spiral loop long-time-series feature modeling subunit is used to perform weighted combination of the multi-granularity load time-series combination features based on the second fully connected layer to obtain an initial multi-granularity load time-series combination feature; perform convolution operation on the initial multi-granularity load time-series combination feature based on a preset fourth one-dimensional convolutional layer to obtain a first multi-granularity load time-series combination feature; perform nonlinear transformation on the first multi-granularity load time-series combination feature based on a preset first activation function to obtain a third multi-granularity load time-series combination feature; and perform nonlinear transformation on the initial multi-granularity load time-series combination feature based on the preset first activation function. The sequence combination features are nonlinearly transformed to obtain the fourth multi-granularity load time-series combination feature; based on the preset spiral ring scan selective state-space model and combined with the prediction step size embedding vector, the third multi-granularity load time-series combination feature is modeled for long-term features to obtain the initial spiral ring scan load prediction long-term features; the initial spiral ring scan load prediction long-term features and the fourth multi-granularity load time-series combination feature are multiplied element-wise to obtain the first spiral ring scan load prediction long-term features; the first spiral ring scan load prediction long-term features are weighted and combined based on the second fully connected layer to obtain the spiral ring scan load prediction long-term features.

[0065] In this embodiment, the long-time-series feature modeling unit includes: a parallel-loop long-time-series feature modeling subunit; the parallel-loop long-time-series feature modeling subunit is used to perform weighted combination of the multi-granularity load time-series combination features based on the second fully connected layer to obtain initial multi-granularity load time-series combination features; perform convolution operation on the initial multi-granularity load time-series combination features based on a preset fourth one-dimensional convolutional layer to obtain first multi-granularity load time-series combination features; perform nonlinear transformation on the first multi-granularity load time-series combination features based on a preset first activation function to obtain third multi-granularity load time-series combination features; and perform nonlinear transformation on the initial multi-granularity load time-series combination features based on the preset first activation function. The sequence combination features are nonlinearly transformed to obtain the fourth multi-granularity load time-series combination feature; based on the preset parallel loop scan selective state-space model and combined with the prediction step size embedding vector, the third multi-granularity load time-series combination feature is modeled for long-term features to obtain the initial parallel loop scan load prediction long-term features; the initial parallel loop scan load prediction long-term features and the fourth multi-granularity load time-series combination feature are multiplied element-wise to obtain the first parallel loop scan load prediction long-term features; the first parallel loop scan load prediction long-term features are weighted and combined based on the second fully connected layer to obtain the parallel loop scan load prediction long-term features.

[0066] In this embodiment, the charging load prediction module 204 includes: a charging load prediction unit; the charging load prediction unit is used to perform convolution operations on the multi-time interval load sequence based on a preset third one-dimensional convolutional layer to obtain multi-granularity load temporal convolutional features; to perform weighted combination of the multi-granularity load temporal convolutional features based on a preset third fully connected layer to obtain multi-granularity load temporal residual features; to perform channel concatenation of the multi-granularity load temporal residual features and the load prediction long-time-series features to obtain load prediction gating features; to perform weighted combination of the load prediction gating features based on the third fully connected layer to obtain load prediction gating combination features; to perform nonlinear transformation on the load prediction gating combination features based on a preset second activation function to obtain load prediction gating coefficients; and to perform weighted fusion of the multi-time interval load sequence and the load prediction long-time-series features based on the load prediction gating coefficients to obtain the charging load prediction result of the electric vehicle.

[0067] This embodiment stitches together historical charging load data by using time intervals, enabling multi-time interval load sequences to cover different cyclical patterns of electric vehicle charging loads. Through intra-group and inter-group multi-granularity time-series feature extraction, it extracts and fuses time-series information at multiple granularities, transforming the multi-time interval load sequences into multi-granularity load time-series features containing multi-cycle priors. This provides rich multi-granularity features for subsequent long-term feature modeling and load prediction. Then, it constructs a prediction step-size embedding vector using the charging load prediction step size, determines multi-granularity load time-series combination features using these features, and uses a preset spiral loop scanning state-space processing algorithm to progressively scan and model long-term time-series features in a spiral manner, while also using a preset parallel loop scanning algorithm. The state-space processing algorithm can scan and model long-term features in a parallel manner, thereby more accurately capturing the changing patterns of electric vehicle charging load over long periods. Combined with the prediction step size embedding vector, it can not only uncover cross-period, discontinuous long-distance jump correlations in the charging load sequence, achieving efficient long-term feature modeling and improving adaptability to the multi-period and highly volatile characteristics of charging load, but also achieve controllable dynamic charging load prediction based on the charging load prediction step size, enhancing the predictability adaptability to the randomness and volatility of electric vehicle charging load. Subsequently, by using the load prediction gating coefficient, it can adaptively fuse multi-time interval load sequences and long-term features of load prediction, improving the prediction accuracy of different charging load prediction step sizes, and achieving accurate prediction of electric vehicle charging load.

[0068] In summary, this invention splices historical charging load data by using the interval time of historical charging load data, enabling multi-time interval load sequences to cover different periodic patterns of electric vehicle charging load. Through intra-group and inter-group multi-granularity time-series feature extraction, it can extract and fuse time-series information of multiple granularities, transforming the multi-time interval load sequences into multi-granularity load time-series features containing multi-period priors, providing rich multi-granularity features for subsequent long-term feature modeling and load prediction. Then, it constructs a prediction step-size embedding vector using the charging load prediction step size, determines multi-granularity load time-series combination features through multi-granularity load time-series features, and uses a preset spiral loop scanning state space processing algorithm to progressively scan and model long-term time-series features in a spiral manner, using a preset parallel... The ring scan state-space processing algorithm can scan and model long-term features in a parallel manner, thereby more accurately capturing the changing patterns of electric vehicle charging load over long periods. Combined with the prediction step size embedding vector, it can not only uncover cross-period, discontinuous long-distance jump correlations in the charging load sequence, achieving efficient long-term feature modeling and improving adaptability to the multi-period and highly volatile characteristics of charging load, but also achieve controllable dynamic charging load prediction based on the charging load prediction step size, enhancing the predictive adaptability to the randomness and volatility of electric vehicle charging load. Subsequently, by using the load prediction gating coefficient, it can adaptively fuse multi-time interval load sequences and long-term features of load prediction, improving the prediction accuracy of different charging load prediction step sizes, and achieving accurate prediction of electric vehicle charging load.

[0069] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A step-size controllable load forecasting method based on multi-granularity time series, characterized in that, include: Historical charging load data of electric vehicles is obtained, and the historical charging load data is spliced ​​together according to the interval time of the historical charging load data to construct a multi-time interval load sequence; Based on a preset multi-granularity cross-temporal feature extraction algorithm, multi-granularity temporal feature extraction within groups and between groups is performed on the multi-time interval load sequence to obtain multi-granularity load temporal features; Obtain the charging load prediction step size of the electric vehicle, construct a prediction step size embedding vector based on the charging load prediction step size, and determine the multi-granularity load time series combination features based on the multi-granularity load time series features; Based on the preset spiral ring scanning state space processing algorithm and the preset parallel ring scanning state space processing algorithm, and combined with the prediction step size embedding vector, the multi-granularity load time series combination features are modeled for long-time features to obtain the spiral ring scanning load prediction long-time features and the parallel ring scanning load prediction long-time features. The long-term characteristics of load prediction are obtained based on the long-term characteristics of the spiral ring scan load prediction and the long-term characteristics of the parallel ring scan load prediction. The load forecasting gating coefficient is calculated based on the multi-time interval load sequence and the long-term time series characteristics of load forecasting; and the multi-time interval load sequence and the long-term time series characteristics of load forecasting are fused based on the load forecasting gating coefficient to determine the charging load forecasting result of the electric vehicle.

2. The step-size controllable load forecasting method based on multi-granularity time series as described in claim 1, characterized in that, The process of acquiring historical charging load data for electric vehicles, and concatenating the historical charging load data according to the time intervals to construct a multi-time interval load sequence, includes: Obtain historical charging load data for electric vehicles; Based on the interval time of the historical charging load data, the historical charging load data is divided to determine several long time interval load sequences and several short time interval load sequences. By splicing the long-time interval load sequence and the short-time interval load sequence one by one, several multi-time interval spliced ​​load sequences are obtained. Based on the multi-time interval spliced ​​load sequence, a load sequence matrix is ​​constructed, and the load sequence matrix is ​​used as the multi-time interval load sequence.

3. The step-size controllable load forecasting method based on multi-granularity time series as described in claim 1, characterized in that, The algorithm based on preset multi-granularity cross-temporal feature extraction performs intra-group multi-granularity temporal feature extraction and inter-group multi-granularity temporal feature extraction on the multi-time interval load sequence to obtain multi-granularity load temporal features, including: The multi-time interval load sequence is subjected to dimensional permutation to obtain the multi-time interval load transpose sequence; Several mask matrices are obtained, and the multi-time interval load transpose sequence is subjected to intra-group multi-granularity time series feature extraction based on the mask matrices to obtain intra-group multi-granularity time series features. Several dilated convolutional layers are obtained, and inter-group multi-granularity temporal features are extracted from the multi-time interval load transpose sequence based on the dilated convolutional layers to obtain inter-group multi-granularity temporal features. The multi-granularity time series features within the group and the multi-granularity time series features between the groups are added element by element to obtain the multi-granularity load time series transpose features. The multi-granularity load time-series transpose features are subjected to dimensional permutation to obtain initial multi-granularity load time-series features; The initial multi-granularity load time series features are weighted and combined based on a preset first fully connected layer to obtain multi-granularity load time series features.

4. The step-size controllable load forecasting method based on multi-granularity time series as described in claim 3, characterized in that, The process involves obtaining several mask matrices and extracting intra-group multi-granularity temporal features from the multi-time interval load transpose sequence based on these mask matrices, resulting in intra-group multi-granularity temporal features, including: Obtain several mask matrices; perform element-wise multiplication of each mask matrix with the multi-time interval load transpose sequence to obtain several mask load time series features; Based on a preset first one-dimensional convolutional layer, the masked load time-series features are convolved to obtain the masked load time-series convolutional features corresponding to each of the masked load time-series features; Based on the first one-dimensional convolutional layer, a convolutional operation is performed on the multi-time interval load transpose sequence to obtain the multi-time interval load temporal convolutional feature corresponding to the multi-time interval load transpose sequence; The masked load temporal convolutional features and the multi-time interval load temporal convolutional features are added element-wise to obtain the multi-granularity load temporal convolutional features within the group. Based on a preset second one-dimensional convolutional layer, the multi-granularity load temporal convolutional features within the group are convolutionally processed to obtain multi-granularity temporal features within the group.

5. A step-size controllable load forecasting method based on multi-granularity time series as described in claim 3 or 4, characterized in that, The step involves obtaining several dilated convolutional layers, and then extracting inter-group multi-granularity temporal features from the multi-time interval load transpose sequence based on these dilated convolutional layers, resulting in inter-group multi-granularity temporal features, including: Obtain several dilated convolutional layers, wherein the dilated convolutional layers have different porosities; Based on each of the dilated convolutional layers, convolution operations are performed on the multi-time interval load transpose sequence to obtain several load temporal dilated convolutional features. All the aforementioned load temporal dilated convolutional features are added element-wise to obtain inter-group multi-granularity temporal dilated convolutional features. Based on a preset second one-dimensional convolutional layer, the inter-group multi-granularity temporal dilated convolutional features are convolved to obtain inter-group multi-granularity temporal features.

6. The step-size controllable load forecasting method based on multi-granularity time series as described in claim 1, characterized in that, The process involves obtaining the charging load prediction step size of the electric vehicle, constructing a prediction step size embedding vector based on the charging load prediction step size, and determining multi-granularity load time series combination features based on the multi-granularity load time series features. Based on the preset spiral ring scanning state space processing algorithm and the preset parallel ring scanning state space processing algorithm, and combined with the prediction step size embedding vector, the multi-granularity load time series combination features are modeled for long-time features to obtain the spiral ring scanning load prediction long-time features and the parallel ring scanning load prediction long-time features. Based on the aforementioned long-term characteristics of load prediction using spiral ring scanning and parallel ring scanning, long-term characteristics of load prediction are obtained, including: Obtain the charging load prediction step size of the electric vehicle, and construct a prediction step size embedding vector based on the charging load prediction step size; The multi-granularity load time series features are weighted and combined based on a preset second fully connected layer to obtain multi-granularity load time series combined features; Based on the preset spiral ring scanning state space processing algorithm and the prediction step size embedding vector, the long-time feature model of the multi-granularity load time series combination features is performed to obtain the long-time feature of spiral ring scanning load prediction. Based on the preset parallel loop scanning state space processing algorithm and the prediction step size embedding vector, the long time series feature modeling of the multi-granularity load time series combination feature is performed to obtain the long time series feature of parallel loop scanning load prediction. The initial long-term load prediction characteristics are obtained by adding the spiral ring scan load prediction long-term characteristics element by element to the parallel ring scan load prediction long-term characteristics. The initial load prediction long-time features are convolved based on a preset third one-dimensional convolutional layer to obtain the first load prediction long-time features. The first load forecast long-time features are weighted and combined based on a preset third fully connected layer to obtain the load forecast long-time features.

7. The step-size controllable load forecasting method based on multi-granularity time series as described in claim 6, characterized in that, The step of modeling long-time features of the multi-granularity load time series combination features based on the preset spiral ring scan state space processing algorithm and the prediction step size embedding vector, to obtain the long-time features of spiral ring scan load prediction, includes: The multi-granularity load time series combination features are weighted and combined based on the second fully connected layer to obtain the initial multi-granularity load time series combination features; Based on the preset fourth one-dimensional convolutional layer, the initial multi-granularity load time series combination features are convolved to obtain the first multi-granularity load time series combination features; The first multi-granularity load time series combination feature is nonlinearly transformed based on a preset first activation function to obtain the third multi-granularity load time series combination feature; The initial multi-granularity load time series combination features are nonlinearly transformed based on a preset first activation function to obtain the fourth multi-granularity load time series combination features; Based on the preset spiral ring scan selective state space model, and combined with the prediction step size embedding vector, the long-time feature model of the third multi-granularity load time series combination feature is performed to obtain the initial spiral ring scan load prediction long-time feature. The first spiral ring scan load prediction long-time characteristics are obtained by performing an element-wise multiplication operation on the initial spiral ring scan load prediction long-time characteristics and the fourth multi-granularity load time combination characteristics. Based on the second fully connected layer, the long-term features of the first spiral ring scan load prediction are weighted and combined to obtain the long-term features of the spiral ring scan load prediction.

8. The step-size controllable load forecasting method based on multi-granularity time series as described in claim 6, characterized in that, The multi-granularity load time-series combined features are modeled using a preset parallel loop scanning state-space processing algorithm and the predicted step-size embedding vector. The long-term time-series characteristics of parallel loop scan load prediction were obtained, including: The multi-granularity load time series combination features are weighted and combined based on the second fully connected layer to obtain the initial multi-granularity load time series combination features; Based on the preset fourth one-dimensional convolutional layer, the initial multi-granularity load time series combination features are convolved to obtain the first multi-granularity load time series combination features; The first multi-granularity load time series combination feature is nonlinearly transformed based on a preset first activation function to obtain the third multi-granularity load time series combination feature; The initial multi-granularity load time series combination features are nonlinearly transformed based on a preset first activation function to obtain the fourth multi-granularity load time series combination features; Based on the preset parallel loop scan selective state space model, and combined with the prediction step size embedding vector, the long time series feature model of the third multi-granularity load time series combination feature is performed to obtain the initial parallel loop scan load prediction long time series feature. The first parallel loop scan load prediction long-time series feature is obtained by performing an element-wise multiplication operation on the initial parallel loop scan load prediction long-time series feature and the fourth multi-granularity load time series combination feature. Based on the second fully connected layer, the long-term time-series features of the first parallel loop scan load prediction are weighted and combined to obtain the long-term time-series features of the parallel loop scan load prediction.

9. The step-size controllable load forecasting method based on multi-granularity time series as described in claim 1, characterized in that, The load forecasting gating coefficient is calculated based on the multi-time interval load sequence and the long-time-series characteristics of load forecasting. Based on the load forecasting gating coefficient, the multi-time interval load sequence and the long-term time-series characteristics of load forecasting are fused to determine the charging load forecasting result of the electric vehicle, including: Based on a preset third one-dimensional convolutional layer, the multi-time interval load sequence is convolved to obtain multi-granularity load temporal convolutional features. The multi-granularity load temporal convolutional features are weighted and combined based on a preset third fully connected layer to obtain multi-granularity load temporal residual features; The multi-granularity load time-series residual features and the load prediction long-series features are concatenated to obtain load prediction gating features. The load prediction gating features are weighted and combined based on the third fully connected layer to obtain the load prediction gating combined features. Based on a preset second activation function, the load prediction gating combination features are nonlinearly transformed to obtain the load prediction gating coefficients. Based on the load prediction gating coefficient, the multi-time interval load sequence and the long-term time series features of load prediction are weighted and fused to obtain the charging load prediction result of the electric vehicle.

10. A step-size controllable load forecasting system based on multi-granularity time series, characterized in that, include: The module includes a multi-time interval load sequence construction module, a multi-granularity load time series feature extraction module, a long time series feature modeling module, and a charging load prediction module. The multi-time interval load sequence construction module is used to acquire historical charging load data of electric vehicles, and to splice the historical charging load data according to the interval time of the historical charging load data to construct a multi-time interval load sequence. The multi-granularity load time-series feature extraction module is used to extract intra-group multi-granularity time-series features and inter-group multi-granularity time-series features from the multi-time interval load sequence based on a preset multi-granularity cross-time-series feature extraction algorithm, thereby obtaining multi-granularity load time-series features. The long-time-series feature modeling module is used to obtain the charging load prediction step size of the electric vehicle, construct a prediction step size embedding vector based on the charging load prediction step size, determine multi-granularity load time-series combination features based on the multi-granularity load time-series features, perform long-time-series feature modeling on the multi-granularity load time-series combination features according to the preset spiral ring scan state space processing algorithm and the preset parallel ring scan state space processing algorithm, combined with the prediction step size embedding vector, to obtain the spiral ring scan load prediction long-time-series features and the parallel ring scan load prediction long-time-series features, and obtain the load prediction long-time-series features based on the spiral ring scan load prediction long-time-series features and the parallel ring scan load prediction long-time-series features. The charging load prediction module is used to calculate the load prediction gating coefficient based on the multi-time interval load sequence and the long-term time series characteristics of load prediction; and to fuse the multi-time interval load sequence and the long-term time series characteristics of load prediction based on the load prediction gating coefficient to determine the charging load prediction result of the electric vehicle.

Images