A method and system for electric vehicle charging load prediction based on a space-time-frequency multi-domain state space model

Abstract

The application discloses a kind of electric vehicle charging load prediction method and system based on space-time frequency multi-domain state space model.The method first uses discrete wavelet transform to decompose non-stationary load data into orthogonal low-frequency trend component and high-frequency characteristic component, the frequency domain component is fused with time embedding feature and space embedding feature, realize the multi-dimensional feature extraction of original charging load sequence;Subsequently, the extracted features are input into the hierarchical coding structure, and the core uses parallel bidirectional Mamba processing flow to synchronously capture the forward and backward long-term time dependence of the sequence.Thirdly, the encoded latent variable is input into the symmetric multi-layer decoder, the multi-scale information from the encoder is integrated using the skip connection, and the recursive feature recovery and refinement are performed through the bidirectional Mamba flow.Finally, the high-precision prediction result is output through the linear mapping layer.The application uses a linear computational complexity state space model to solve the problem of excessive computational overhead when the traditional Transformer model processes long sequences, as well as the non-linear, volatility and long-distance dependence problems that traditional models cannot handle.Experiments show that the method improves the prediction accuracy while greatly reducing the inference power consumption, and the system is suitable for deployment on the power grid scheduling terminal with lightweight requirements, helping to balance and schedule the load in the power network.

Description

Technical Field

[0001] This invention relates to the fields of power system load forecasting and artificial intelligence technology, and specifically to a method for forecasting electric vehicle charging load based on the improved state-space model (Mamba). Background Technology

[0002] With the acceleration of global energy transition and the advancement of transportation electrification, electric vehicles (EVs) have become a key component of network-based physical energy systems. However, the large-scale integration of EVs has placed unprecedented pressure on distribution networks. Because charging behavior is influenced by multiple stochastic factors such as user psychology, traffic conditions, weather conditions, and electricity pricing policies, charging loads exhibit strong nonlinearity, non-stationarity, and significant sudden "climbing" characteristics. Accurate long-sequence charging load forecasting has become a core prerequisite for achieving reliable grid dispatch, optimal resource allocation, and scientific planning of charging infrastructure.

[0003] Currently, mainstream electric vehicle load forecasting technologies have evolved from traditional statistical models to deep learning architectures. Early statistical models, such as ARIMA, while having certain advantages in handling stationary sequences, struggled to capture the complex nonlinear mappings in charging behavior. Subsequent recurrent neural networks, such as LSTM and GRU, improved time-series modeling capabilities, but were prone to gradient vanishing and computationally inefficient when dealing with extremely long sequences, making it difficult to meet the real-time processing needs of large-scale data. In recent years, while Transformer models, with their self-attention mechanism at their core, have made breakthroughs in capturing long-range dependencies, they are limited by the bottleneck that computational complexity increases quadratically with sequence length, making them difficult to deploy in resource-constrained industrial edge computing scenarios. Furthermore, they often neglect in-depth analysis and denoising of the frequency characteristics of load signals. Summary of the Invention

[0004] This invention aims to address the problems of severe noise interference and difficulty in capturing key trends in electric vehicle (EV) charging load forecasting due to the high non-stationarity, strong randomness, and multi-scale frequency characteristics of load signals; and also to address the limitations of existing Transformer-type deep learning models in processing long-sequence data. O ( N 2To address the bottlenecks of high computational complexity and memory overhead, as well as the technical shortcomings of standard Mamba and other state-space models such as Spatiotemporal Frequency Domain (SFLG) due to their unidirectional recursive inference mechanism leading to the accumulation of prediction errors in long sequences and insufficient utilization of bidirectional contextual information, this invention provides a spatiotemporal frequency domain prediction scheme that balances linear computational efficiency with high accuracy. This scheme first utilizes wavelet transform to construct an embedding layer, decoupling the original load signal into multi-scale frequency domain components. Then, it fuses these components with the time- and spatial embedding layers to achieve multi-dimensional feature fusion, thereby preserving the core evolutionary trend while achieving explicit denoising. In terms of core architecture, this invention abandons the high-overhead self-attention mechanism. By introducing a state-space model with linear computational complexity and designing a dual-stream encoder-decoder structure in parallel with dual-stream Mamba, it achieves accurate capture of bidirectional contextual information for ultra-long sequences. This hierarchical multi-scale modeling approach not only significantly improves computational efficiency in complex industrial scenarios but also greatly enhances the model's robustness and prediction accuracy to dynamic changes in charging load through deep integration of features in the time, space, and frequency domains.

[0005] The specific technical solution is as follows: A spatiotemporal frequency prediction method for electric vehicle charging load based on a dual-stream Mamba structure, comprising the following steps: 1) Charging load data acquisition and preprocessing This invention employs a two-stage data cleaning process. First, missing values ​​are filled using linear interpolation. Then, outliers exceeding three standard deviations are pruned. To improve convergence, minimum-maximum normalization is used to map the data to the [0,1] interval. In the above formula, where xmax and xmin These represent the maximum and minimum load values, respectively.

[0006] 2) Frequency-Time-Spatial Multidimensional Feature Embedding Mechanism The original load sequence was decomposed into low-frequency components using a first-order Haar discrete wavelet transform. X LF and high frequency components X HF Low-frequency components are used to capture the long-term evolution trend and seasonal cycle of charging load, while high-frequency components are used to identify sudden charging behavior and random noise, thereby achieving explicit decoupling of time-series signals in the frequency domain.

[0007] In the above formula, This represents an approximation coefficient used to capture low-frequency trends. ( t ) represents the initial decomposition scalej 0 Horizontal displacement k The scaling function below; This represents the detail coefficients for capturing high-frequency fluctuations. Representing different scales i Translational displacement k The wavelet function below; h and g These represent the low-pass and high-pass filters associated with the Haar wavelet, respectively. The low-frequency components, high-frequency components, and the original sequence are concatenated to obtain the wavelet transform output. : In the above formula, Concat (·) represents a feature concatenation operation; Based on frequency characteristics, a time embedding module is introduced to identify periodic information such as daily and weekly data. Through learnable linear projection, the original multivariate time series data is transformed. Transform into rich latent representation : In the above formula, Represents learning weights, Indicates bias; Spatial embedding is also introduced to characterize the geographical associations and POI features of different charging stations: In the above formula i Indicates the number of features, ti Indicates the first i A time characteristic, This represents the corresponding embedding matrix, where n represents the number of feature categories, and o(·) represents the operation of converting a scalar into a one-hot vector; Finally, the above features are linearly transformed and concatenated to generate a frequency-time-space-frequency fusion feature tensor. X WST ： In the above formula, Linear (·) represents a linear transformation operation.

[0008] 3) Parallel bidirectional Mamba backbone network Symmetric parallel processing streams are introduced in each layer of the encoder and decoder: the forward Mamba stream extracts historical cumulative dependencies in chronological order, and the reverse Mamba stream extracts future context in reverse order. In the above formula, This represents positive Mamba operation output. This indicates the output of the reversed Mamba operation; The outputs of bidirectional Mamba are concatenated along the feature dimension to obtain the output of the double Mamba layer. Y dual : The outputs of the forward and inverse Mamba algorithms are summed element-wise, followed by convolution and layer normalization. Then, the residual connections are summed with the input features, and the result is normalized again to obtain the encoder output. R : In the above formula, Presentation layer normalization operation, This indicates a convolution operation.

[0009] 4) Layered encoding-decoding structure The frequency-temporal-spatial fused feature tensor is input into a hierarchical two-stream Mamba encoder, which consists of multiple stacked coding blocks. The encoded latent representation is denoted as R: In the above formula, This indicates a dual-stream Mamba encoder operation; The dual-stream Mamba decoder employs a progressive decoding strategy, combining latent representations from the encoder with dynamic dependency modeling. The encoder output is first enhanced through residual convolutional blocks to strengthen local features. Subsequently, layer normalization is applied to stabilize training, thereby generating latent representations. : Then Input is fed into a bidirectional Mamba stream to decode the forward and backward time dependencies: In the above formula, This indicates a dual Mamba decoder operation.

[0010] 5) Model Output Finally, the decoded features are mapped to the predicted sequence through a linear transformation: 6) Loss Function The training objective of EVMamba is to minimize the predicted charging load. Compared with the true value y i,j Mean squared error (MSE) between: In the above formula, θ Indicates trainable parameters, B Indicates batch size. H Indicates the predicted duration.

[0011] Compared with the prior art, the present invention is able to: This paper pioneers a Mamba-based prediction paradigm and applies it for the first time to electric vehicle charging load prediction. It employs a refined decoder instead of an autoregressive generation method to mitigate recursive error accumulation. While maintaining linear computational complexity (…),… O ( L At the same time, it avoids the limitations of Transformer and RNN.

[0012] A frequency-time-space multi-domain feature embedding method is proposed. First, discrete wavelet transform is used to decompose the original load data into mutually orthogonal frequency components. Then, the features in the time domain and spatial domain are fused together to effectively filter out high-frequency noise and retain key trends, thereby enhancing the feature extraction capability in volatile industrial environments.

[0013] A parallel dual-Mamba backbone network was designed to achieve bidirectional context modeling. By simultaneously performing forward and backward processing, long-range dependencies are captured. Furthermore, the feature representation capability is enhanced, ensuring excellent prediction accuracy in both short-term fluctuations and long-term planning scenarios.

[0014] The number of model parameters is reduced by 50% compared to Transformer, making it ideal for deployment on resource-constrained grid edge controllers. Attached Figure Description

[0015] To more clearly illustrate the prediction scheme in the implementation of the present invention, the following is a brief introduction to the drawings used in the prior art and embodiments. The following drawings are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 The present invention provides a method for predicting electric vehicle charging load based on a spatiotemporal frequency domain state-space model and an overall block diagram thereof.

[0017] Figure 2 The flowchart of the frequency-time-space multidimensional embedding module provided by the present invention.

[0018] Figure 3The graph shows the 336-hour load prediction results of EVMamba and MetaProbformer on the Palo dataset provided by this invention.

[0019] Figure 4 A comparison chart of the computational efficiency of EVMamba with other models provided for this invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific examples. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention, with each act of creative effort, are within the scope of protection of this invention.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0022] The electric vehicle charging load prediction method based on a spatiotemporal frequency domain state-space model described in this invention is as follows: Figure 1 Follow these steps in sequence: Step 1) Data Acquisition and Preprocessing The original transaction messages of EV charging piles are collected in real time through the charging station management system, and core fields including charging start / end time and single charging amount (kWh) are extracted. Discrete charging events are transformed into a sequence of average hourly charging loads (kW). For sensor faults in industrial scenarios, linear interpolation is used to repair missing points, and the three-standard-deviation criterion (3...) is applied. Smoothly prune outlier values.

[0023] To accelerate convergence, the load values ​​are mapped to the [0, 1] interval using a minimum-maximum normalization method: In the above formula, where xmax and xmin These represent the maximum and minimum load values, respectively. Input-output pairs are generated using a sliding window with a step size of 1 hour. Finally, the dataset is divided into training, validation, and test sets in chronological order, with a ratio of 7:1:2.

[0024] Step 2) Construct a frequency-time-space multidimensional embedding layer like Figure 2As shown, the original load sequence is first decomposed into low-frequency components using a first-order Haar discrete wavelet transform. X LF and high frequency components X HF Low-frequency components are used to capture the long-term evolution trend and seasonal cycle of charging load, while high-frequency components are used to identify sudden charging behavior and random noise, thereby achieving explicit decoupling of time-series signals in the frequency domain; In the above formula, This represents an approximation coefficient used to capture low-frequency trends. ( t ) represents the initial decomposition scale j 0 Translational displacement k The scaling function below; This represents the detail coefficients for capturing high-frequency fluctuations. Representing different scales i Translational displacement k The wavelet function below; h and g These represent the low-pass and high-pass filters associated with the Haar wavelet, respectively. Low-frequency components are separated, high-frequency components are concatenated with the original sequence, and the wavelet transform output is obtained. : In the above formula, Concat (·) represents a feature concatenation operation; Based on frequency characteristics, a time embedding module is introduced to identify periodic information such as daily and weekly data. Through learnable linear projection, the original multivariate time series data is transformed. Transform into rich latent representation : In the above formula, Represents learning weights, Indicates bias; Spatial embedding is also introduced to characterize the geographical associations and POI features of different charging stations: In the above formula i Indicates the number of features, ti Indicates the first i A time characteristic, This represents its corresponding embedding matrix. n This indicates the number of feature categories, and o(·) represents the operation of converting a scalar into a one-hot vector; Finally, the above features are linearly transformed and concatenated to generate a frequency-time-space-frequency fusion feature tensor. X WST： In the above formula, Linear (·) represents a linear transformation operation.

[0025] Step 3) Bidirectional parallel Mamba encoding process A symmetrical parallel processing stream is introduced in each layer of the encoder: the forward Mamba stream extracts historical cumulative dependencies in chronological order, and the reverse Mamba stream extracts future context in reverse order. In the above formula, This represents a positive Mamba operation output. This indicates the output of the reversed Mamba operation; The outputs of bidirectional Mamba are concatenated along the feature dimension to obtain the output of the Mamba layer. Y dual : The outputs of the forward and inverse Mamba algorithms are summed element-wise, followed by convolution and layer normalization. Then, the residual connections are summed with the input features, and the result is normalized again to obtain the encoder output. R : In the above formula, Presentation layer normalization operation, This indicates a convolution operation.

[0026] For ease of calculation, let To represent the operation of a dual-stream Mamba encoder, the encoder output R can be expressed as: Step 4) Dual-stream Mamba layered decoding process The dual-stream Mamba decoder employs a progressive decoding strategy, combining latent representations from the encoder with dynamic dependency modeling. The encoder output is first enhanced through residual convolutional blocks to strengthen local features. Subsequently, layer normalization is applied to stabilize training, thereby generating latent representations. : In the above formula, Presentation layer normalization operation; Then Input is fed into a bidirectional Mamba stream to decode the forward and backward time dependencies: In the above formula, This indicates a dual Mamba decoder operation.

[0027] Step 5) Output Prediction and Model Optimization The decoded features are mapped to the predicted sequence through a linear transformation: The training objective of EVMamba is to minimize the predicted charging load. Compared with the true value Mean squared error (MSE) between: In the above formula, θ Indicates trainable parameters, B Indicates batch size. H Indicates the predicted duration.

[0028] To verify the effectiveness and superiority of the electric vehicle charging load prediction method based on a spatiotemporal frequency domain state-space model described in this invention, we evaluated it on two real-world electric vehicle charging station load datasets: data from Palo Alto, USA, and Perth, Australia. The datasets record the start and end timestamps of all charging sessions and the total energy consumption (kWh), covering charging load records for 4, 31, 12, and 24 months, with 3113 and 17520 data points, respectively. Specifically, EVMamba was compared and validated with traditional statistical models, machine learning models, and commonly used deep learning models. Table 1 shows the load forecasting accuracy evaluation metrics of the compared models in two cases. EVMamba demonstrated consistent superiority and strong robustness across all evaluation dimensions. In the PALO dataset, EVMamba achieved excellent results with MSE of 4.471 kW², RMSE of 2.114%, MAPE of 20.935 kW, and MAE of 1.401 kW. Compared to the best baseline method MetaProbformer, its MSE was reduced by approximately 10.72%, and its MAPE, which reflects relative error control capability, was significantly reduced by 38.41%. The results of the Perth dataset further validated this leading position, with an MSE of 158.678 kW² and a MAPE of 4.026 kW, representing improvements of approximately 6.7% and 23.21% respectively compared to MetaProbformer. Compared to similar SMamba models, EVMamba demonstrates particularly significant performance improvements on both datasets, especially on the PALO dataset where improvements range from 8.6% to 55.6%. Experimental results show that while some models, such as TimeMachine, show some competitiveness in the MAE metric, EVMamba, with its unique spatiotemporal frequency domain modeling architecture, comprehensively surpasses traditional statistical methods and current state-of-the-art deep learning models in terms of squared error of fit (MSE, RMSE) and noise robustness (MAPE), which measure large deviation fitting ability.

[0029] Table 1 To evaluate the impact performance of each component of the model, a comprehensive ablation study was conducted, as shown in Table 2, which demonstrates the performance of EVMamba with and without (w / o) key components.

[0030] Table 2 Figure 4 The visualization compares EVMamba and MetaProbformer on a randomly selected 336-hour charging load prediction window. Although the performance of both models is basically consistent with the actual values, EVMamba shows better consistency, especially near the peak and trough values.

[0031] In addition, the computational efficiency of EVMamba was evaluated using three metrics: accuracy, GFLOPS, and parameter count. Figure 3As shown, EVMamba achieves an optimal balance. It achieves the lowest mean square error (MSE) with minimal hardware overhead. With only 5.35 million parameters, EVMamba is 50% smaller than the Transformer model, indicating that it is suitable for deployment in lightweight grid edge controllers.

Claims

1. A method and system for predicting electric vehicle charging load based on a spatiotemporal frequency domain state-space model, characterized in that, Includes the following steps: (1) Obtain historical electric vehicle charging load sequence data and preprocess it to obtain a normalized input sequence. X ; (2) Input sequence X The input sequence is first decomposed into high-frequency components by discrete wavelet transform into frequency-time-space embedding layer. X HF and low frequency components X LF Then, the frequency domain components are embedded with the time features. E t and spatial embedding features E s By fusing the data, multi-domain joint features can be generated. X WST ; (3) X WST The input is fed into a dual-stream Mamba multilayer encoder, which uses parallel forward and reverse Mamba processing streams to capture long-term dependencies from forward and reverse time order, respectively, and extracts multi-resolution feature patterns through multilayer downsampling. (4) Input the encoded latent variables into the symmetric multilayer decoder, integrate the multi-scale information from the encoder using skip connections, and perform recursive feature recovery and refinement through bidirectional Mamba streams; (5) The decoded features are mapped back to the time series domain through a linear mapping layer, and the output has linear computational complexity ( O ( L Predicted charging load sequence .

2. The method according to claim 1, wherein the charging load data preprocessing in step 1) is characterized in that: A two-stage data cleaning process is adopted. First, missing values ​​are filled in using linear interpolation. Then, outliers exceeding three standard deviations are pruned. Finally, min-max normalization is used to map the data to the [0,1] interval to improve data convergence. In the above formula, where xmax and xmin These represent the maximum and minimum load values, respectively.

3. The method according to claim 1, wherein the time-space-frequency feature embedding layer in step 2) is characterized in that: The charging load multi-domain feature extraction is performed using a frequency-temporal-spatial multidimensional embedding method. The specific steps are as follows: (1) Using Haar discrete wavelet transform as the basis function, the scale is decomposed. j 0 Setting it to 1 decomposes the fluctuating raw load data into mutually orthogonal high-frequency components. X HF with low-frequency components X LF : In the above formula, This represents an approximation coefficient used to capture low-frequency trends. ( t ) represents the initial decomposition scale j0 Translational displacement k The scaling function below; This represents the detail coefficients for capturing high-frequency fluctuations. Representing different scales i Translational displacement k The wavelet function below; h and g These represent the low-pass and high-pass filters associated with the Haar wavelet, respectively. (2) The low-frequency component, high-frequency component and the original sequence are concatenated to obtain the output of the wavelet transform. : (3) Integrating frequency domain features with time embedding features Et and spatial embedding features Es By fusing the data, multi-domain joint features can be generated. XWST : 。 4. The method according to claim 1, wherein the dual-stream Mamba multilayer encoder in step 3) is characterized in that, Generate feature representations that take into account contextual information based on bidirectional Mamba streams, including: Each coding unit includes a forward Mamba operator. and a reverse Mamba operator : The outputs of the forward and inverse operators are summed element-wise, followed by convolution and layer normalization. Then, the residual connections are summed with the input features, and layer normalization is applied again to obtain the encoder output. R : In the above formula, Presentation layer normalization operation, This indicates a convolution operation.

5. The method according to claim 1, wherein the decoder in step (4) is a symmetric multilayer decoder, characterized in that, To avoid the accumulation of recursive errors, non-autoregressive generation methods are used instead of traditional autoregressive generation methods, including: In the future, the encoder output will first undergo feature enhancement through residual convolutional blocks, followed by application layer normalization for training, thereby generating a latent representation. : Will The input is fed into a bidirectional Mamba stream to the encoder to decode the forward and backward time dependencies: In the above formula, This indicates a dual-stream Mamba decoder operation.

6. The method according to claim 1, wherein the charging load prediction output in step (5) is characterized in that, The decoded features are mapped to the output sequence through a linear transformation: 。 7. The method according to claim 1, characterized in that: During the model training phase, mean squared error (MSE) was used as the loss function, and the Adam optimizer was used in conjunction with a cosine annealing strategy for iterative optimization to improve the model's prediction performance during peak power grid periods. In the above formula, This indicates the predicted charging load value. Let θ represent the true value and θ represent the trainable parameters. B Indicates batch size. H Indicates the predicted duration.

8. An electric vehicle charging load prediction system implementing the method of any one of claims 1 to 6, characterized in that, include: (1) Data acquisition and preprocessing module, used to normalize historical charging load and handle outliers; (2) Time-space-frequency feature extraction module, used to extract the frequency domain, time domain and spatial domain features of the sequence; (3) The core prediction module includes the hierarchical dual-stream Mamba encoder and the symmetric decoder; (4) Prediction output module, used to generate prediction curves and push them to the power grid dispatch terminal in real time.

Images