Charging pile demand prediction method based on frequency domain multi-head space-time attention and fluctuation propagation

Abstract

The invention relates to the technical field of charging pile demand prediction, in particular to a charging pile demand prediction method based on frequency domain multi-head space-time attention and fluctuation propagation, and the method comprises the steps: collecting charging pile demand space-time sequence data; frequency domain conversion is carried out on the actual charging pile demand space-time sequence data through fast Fourier transform; according to a multi-head space-time attention mechanism, performing dynamic fusion on output of multi-head attention; constructing a fluctuation propagation layer based on a layered high-order differential equation, a multi-head fusion output matrix and a random turbulence model; constructing a multi-task loss function; constructing a multi-level spatio-temporal representation learning framework; and training the multi-level space-time representation learning framework according to the to-be-tested charging pile demand space-time sequence data set and the multi-task loss function, and outputting a charging pile demand prediction model. According to the method, complex features in the space-time sequence data can be fully mined, the features are enhanced by using the frequency domain information, and the accuracy of a prediction result is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of charging pile demand forecasting technology, and in particular to a charging pile demand forecasting method based on frequency domain multi-head spatiotemporal attention and fluctuation propagation. Background Technology

[0002] As the global low-carbon transition strategy progresses, the electric vehicle industry is experiencing rapid development. As a key infrastructure for electric vehicle energy replenishment, the construction and operation efficiency of charging piles directly impacts user experience and the healthy development of the industry, with accurate charging demand forecasting being a core issue. Charging demand forecasting, by analyzing historical data and related factors, quantitatively estimates the future charging pile usage demand in a specific area, which is of great significance for charging infrastructure planning, resource optimization, and operation management. Traditional charging demand forecasting methods are mainly divided into three categories. Statistical model-based methods were widely used in the early stages, modeling the temporal patterns of charging demand through time series analysis. While computationally simple and easy to implement, they are limited by linear assumptions and struggle to capture nonlinear features and complex spatiotemporal dependencies. With the development of machine learning technology, methods such as support vector regression and random forests have been introduced, establishing nonlinear mapping relationships by learning feature patterns to improve accuracy. However, these methods are highly dependent on feature engineering and struggle to effectively utilize the inherent structure of spatiotemporal data, leading to unstable prediction performance in complex urban environments. In recent years, deep learning technology has provided new ideas for charging demand forecasting. Sequence models such as recurrent neural networks can automatically learn time dependencies, reducing the need for manual feature engineering, but they do not fully utilize spatial dimension information. Graph neural networks and spatiotemporal graph neural networks, while further integrating spatiotemporal modeling capabilities, still face numerous technical challenges: First, their receptive fields are limited, making it difficult to capture long-distance spatiotemporal dependencies; attention mechanisms are computationally complex and do not fully consider the topological characteristics of road networks. Second, they lack physical consistency; the pure data-driven model does not incorporate physical constraints, resulting in insufficient robustness and difficulty in physical interpretation when data is sparse or noisy. Third, they struggle to simultaneously characterize deterministic propagation patterns and stochastic turbulence characteristics, making it difficult to achieve a balance between the two. Fourth, modeling behaviors such as demand propagation and reflection in complex urban environments is challenging. Fifth, their low computational efficiency limits applications in large-scale urban scenarios and real-time prediction needs.

[0003] Chinese Patent Publication No. CN118411055A discloses a method and system for predicting the energy demand of electric vehicle charging pile networks. The method includes: acquiring energy consumption data of charging piles; training a long short-term memory neural network based on the energy consumption data of the charging piles to obtain an end-to-end model; training the end-to-end model based on an adaptive evolution strategy of the covariance matrix to obtain an energy demand prediction model for the electric vehicle charging pile network; and generating a prediction result for the future energy demand of the charging pile network based on the result of the energy demand prediction model. This solution can only predict the energy demand of charging piles based on the long short-term memory neural network, but it cannot fully explore the complex features in the spatiotemporal sequence data, it is difficult to enhance features using frequency domain information, and it has shortcomings in spatiotemporal feature fusion and fluctuation propagation modeling, which reduces the accuracy of the prediction results. Summary of the Invention

[0004] To address this, the present invention provides a charging pile demand prediction method based on frequency domain multi-head spatiotemporal attention and wave propagation, which overcomes the problems in existing technologies that cannot fully exploit the complex features in spatiotemporal sequence data, are difficult to enhance features using frequency domain information, and have shortcomings in spatiotemporal feature fusion and wave propagation modeling, thus reducing the accuracy of prediction results.

[0005] To achieve the above objectives, this invention provides a charging pile demand prediction method based on frequency domain multi-head spatiotemporal attention and fluctuation propagation, comprising the following steps: S1. Collect spatiotemporal sequence data of charging pile demand, and preprocess the spatiotemporal sequence data of charging pile demand to obtain spatiotemporal sequence data of actual charging pile demand. S2. The spatiotemporal sequence data of actual charging pile demand is transformed into the frequency domain by fast Fourier transform to obtain the frequency domain representation. A self-calibrated spectral filter is constructed, and the frequency domain representation is enhanced by the self-calibrated spectral filter to obtain the frequency domain enhanced representation. S3. Construct a multi-head spatiotemporal attention mechanism based on frequency domain enhanced representation, and dynamically fuse the outputs of the multi-head attention according to the multi-head spatiotemporal attention mechanism to obtain the multi-head fusion output matrix; S4. Construct a wave propagation layer based on hierarchical high-order differential equations, multi-head fusion output matrix, and stochastic turbulence model; S5. Determine the prediction loss based on the prediction results of the wave propagation layer, and construct a multi-task loss function based on the prediction loss, physical consistency loss, wave equation constraints and regularization terms. S6. Construct a multi-level spatiotemporal representation learning framework through temporal multi-scale representation learning and spatial multi-scale representation learning; S7. Train the multi-level spatiotemporal representation learning framework based on the spatiotemporal sequence dataset of the charging pile demand and the multi-task loss function, and output the multi-level spatiotemporal representation learning framework that meets the preset accuracy as the charging pile demand prediction model.

[0006] Compared with the prior art, the beneficial effects of this application are as follows: By constructing a multi-head spatiotemporal attention mechanism and dynamically fusing the outputs of multiple attention heads, the shortcomings of existing technologies in spatiotemporal feature fusion are addressed. The multi-head spatiotemporal attention mechanism extracts key features from both the temporal dimension (capturing the changing patterns of demand over time) and the spatial dimension (analyzing the demand correlations among charging stations in different regions). Each attention head independently focuses on features at different spatiotemporal granularities. The output weights of each head are automatically adjusted based on feature importance, integrating spatiotemporal features of different granularities to capture spatiotemporal interactions. This avoids information loss from a single attention head, improves the comprehensiveness and discriminative power of feature representation, and enables the model to more accurately characterize the spatiotemporal correlations of charging demand.

[0007] By constructing a wave propagation layer using hierarchical high-order differential equations, a multi-source fusion output matrix, and a stochastic turbulence model, the shortcomings of existing technologies in wave propagation modeling are addressed. The hierarchical high-order differential equations can describe the dynamic propagation patterns of charging demand fluctuations in space and time, capturing high-order variation characteristics. The stochastic turbulence model introduces stochastic disturbance terms to simulate fluctuations in actual charging demand caused by uncertainties such as user behavior and the external environment. Combined with the multi-source fusion output matrix, the wave propagation layer can more accurately model the propagation process of demand fluctuations, capturing both long-term trends and responding to sudden fluctuations, thus improving the model's predictive ability for complex wave scenarios.

[0008] By constructing a multi-layered spatiotemporal representation learning framework through temporal and spatial multi-scale representation learning, this approach addresses the problem that existing technologies cannot fully exploit the complex features of spatiotemporal sequence data. Temporal multi-scale learning captures demand change patterns at different temporal granularities, focusing on both short-term fluctuations and long-term trends. Spatial multi-scale learning analyzes demand correlations across different spatial ranges, considering both the personalized needs of local charging stations and the synergistic effects of regional networks. This multi-layered framework integrates spatiotemporal features at different scales, forming hierarchical feature representations that comprehensively mine complex spatiotemporal correlations in the data. This enhances the model's feature expression capabilities, making prediction results more closely aligned with the spatiotemporal dynamics of actual charging demand.

[0009] Furthermore, in S2, the frequency domain representation The mathematical expression is: In the formula, Represents the Fast Fourier Transform. This represents the spatiotemporal sequence data of charging pile demand. This indicates a self-calibrating spectral filter. Indicates the parameters of the self-calibrating spectral filter; The mathematical expression for the self-calibrating spectral filter is: In the formula, This represents the transfer function between frequency component i and frequency component j. These represent learnable filtering parameters used to control the bandwidth or sharpness of the filter. These represent learnable filtering parameters used to adjust the gain or amplitude response of the filter. These represent learnable filtering parameters used to adjust the phase response or symmetry of the filter. Indicates the target frequency component. Represents the frequency components of the input signal. Represents the target frequency component and the frequency components of the input signal The frequency distance between them.

[0010] In this scheme, the spatiotemporal sequence data of charging pile demand is converted to the frequency domain by using Fast Fourier Transform, and a self-calibrating spectrum filter and its parameters are introduced. The filter performance can be adaptively adjusted according to the characteristics of the spatiotemporal sequence data of charging pile demand, so as to process the frequency domain signal more accurately.

[0011] Furthermore, S3 includes the following steps: S31. Divide the frequency domain enhancement representation into H heads, determine the preset frequency range that is adapted to each head, and divide each head into the corresponding frequency domain components according to the frequency range adapted to each head. S32. Construct adaptive query, key, and value matrices in the frequency domain components corresponding to each head, and perform spatiotemporal domain transformation on the adaptive query, key, and value matrices according to the inverse Fourier transform to obtain the spatiotemporal domain query matrix, spatiotemporal domain key matrix, and spatiotemporal domain value matrix. S33. Construct a topology-sensitive distance matrix based on the location, distance, and topological relationship between each charging station. The mathematical expression is: In the formula, Represents the topology-sensitive distance matrix The element in the i-th row and j-th column, This represents the Euclidean distance between node i and node j. This represents the directional angle between node i and node j. Represents the distance parameter. Indicates the angle parameter. Indicates topology-sensitive terms, , Represents the normalized diffusion Laplace matrix. These represent the topological parameters used to control the diffusion rate. This represents the topological parameters used to adjust the power of the diffusion process. This represents the diffusion distance between node i and node j. This indicates that the direction depends on the modulation function. , Indicates the direction-dependent strength. Indicates the reference direction angle. S34. Construct a topology-aware attention mask based on the urban road network structure. The mathematical expression of the topology-aware attention mask is: , This represents the element in the i-th row and j-th column of the topology-aware attention mask. Indicates the initial mask. , This represents the shortest path length between node i and node j. This represents the distance threshold for the h-th head. Indicates an indicator function, Indicates the flow impact coefficient. This represents the traffic flow between node i and node j; S35. Query the spatiotemporal domain matrix based on the h-th head. The spatiotemporal key matrix of the h-th head The topological sensitivity distance matrix of the h-th head Topological attention mask of the h-th head A multi-head spatiotemporal attention mechanism is constructed, and its mathematical expression is as follows: , This represents the attention weight of the h-th head. Represents the transpose of the spatiotemporal domain key matrix. Indicates the scaling factor. Indicates the activation function; S36. Calculate the attention output of each head based on the multi-head spatiotemporal attention mechanism to obtain the multi-head attention output. Then, perform concatenation and linear merging operations on the multi-head attention outputs to generate a multi-head fusion output matrix. The mathematical expression for the attention output of the h-th head is: , The mathematical expression for the attention output of the h-th head, the connection operation, and the linear merging is: In the formula, This represents the multi-head fusion output matrix. Indicates a connection operation. This indicates the output for all H headers. Perform a connection operation. Represents the dynamic fusion matrix. , Represents the parameterized dynamic function. Indicates contextual information, This represents the output transformation matrix of the h-th head. This represents the set of output transformation matrices for all H heads.

[0012] In this scheme, by dividing the frequency domain enhanced representation into multiple heads and adapting them to a preset frequency range, it can process information in different frequency domains more precisely; by constructing an adaptive query, key, and value matrix and converting it to the spatiotemporal domain, it can better capture spatiotemporal features; by constructing a topology-sensitive distance matrix based on charging station locations, it considers the complex topological relationships between nodes; by constructing a topology-aware attention mask and combining it with road networks and traffic flow, it enhances the adaptability to real-world scenarios; finally, it generates a multi-head fusion output matrix, which effectively integrates the outputs of different heads, fully considers spatiotemporal and topological factors, and improves the accuracy and effectiveness of data processing and analysis.

[0013] Furthermore, S4 includes the following steps: S41. Construct hierarchical higher-order differential equations; S42. By analyzing the reflection and attenuation behavior of urban boundaries using boundary normal vectors and attenuation coefficients, the urban boundary condition processing mechanism is obtained. The mathematical expression of the urban boundary condition processing mechanism is: In the formula, This represents the boundary condition value at spatial location x and current time t. The reflection matrix represents the spatial location x and the current time t. This represents the fluctuation value at spatial location x and current time t. Represents the computational domain The boundary; S43. Based on the urban boundary condition processing mechanism and the multi-head fusion output matrix, an external condition vector is constructed. Then, based on the hidden state, the external condition vector, and the dynamic function, a hierarchical high-order differential equation state update equation is constructed. The mathematical expression of the state update equation is: In the formula, This indicates a hidden state. Represents hidden state rate of change over time Represents the external condition vector. Represents the parameterized dynamic function; S44. Based on the Gaussian process and diffusion tensor, a stochastic turbulence model is obtained by modeling the stochastic characteristics of charging pile demand changes. The mathematical expression of the stochastic turbulence model is: In the formula, This represents the turbulence term at spatial location x and current time t. Denotes the divergence operator, Represents the diffusion tensor. This represents the gradient of the fluctuation value at spatial location x and current time t. Indicates the diffusion term. Represents random noise; S45. Based on the stochastic turbulence model, the state update equation of the layered high-order differential equation is updated to obtain the wave propagation layer. The mathematical expression for the update process of the wave propagation layer is: In the formula, This indicates the current state of the wave propagation layer. This indicates the state of the wave propagation layer at the next moment. Indicates the time step.

[0014] In this scheme, by constructing hierarchical high-order differential equations, the complex dynamic changes in charging pile demand can be more accurately characterized; by analyzing the reflection and attenuation behavior of urban boundaries using boundary normal vectors and attenuation coefficients, a boundary condition processing mechanism is obtained, enhancing the consideration of urban boundary factors; by constructing external condition vectors and state update equations, combined with multi-head fusion output matrices, the rationality of state updates is improved; by modeling stochastic characteristics based on Gaussian processes and diffusion tensors, a stochastic turbulence model is obtained, which better reflects actual stochastic fluctuations; by using this model to update the state update equations, a fluctuation propagation layer is obtained, which can more accurately simulate the fluctuation propagation process of charging pile demand and improve the accuracy of prediction and analysis.

[0015] Furthermore, in step S43, the dynamic function is parameterized using a hybrid expert system and higher-order differential constraints. The mathematical expression is: In the formula, Indicates the first An expert ODE function, Indicates the first Dynamic weighting function for each expert 1 represents the number of experts. Indicates the first A higher-order differential constraint Indicates the first Dynamic weighting function for a higher-order differential constraint, This indicates the order of the differential constraint. Representing the hidden state in terms of time The first derivative.

[0016] In this scheme, by using a hybrid expert system and high-order differential constraints to parameterize the dynamic function, the advantages of different expert functions can be combined, and the high-order differential constraints can be used to enhance the accurate characterization of the dynamic changes of the hidden state, thereby improving the accuracy and flexibility of the model in modeling the dynamic characteristics of charging pile demand.

[0017] Furthermore, the wave velocity parameters in the dynamic function are adjusted based on the wave velocity adjustment term to obtain adaptive wave velocity parameters. The mathematical expression is: Furthermore, the wave velocity adjustment term is parameterized using a neural network. The mathematical expression for the wave velocity adjustment term is as follows: In the formula, Indicates the fundamental wave velocity. This indicates the wave speed adjustment item. Represents spatial coordinates, Represents the time coordinate. Indicates environmental characteristics, This represents the wave velocity parameter network.

[0018] In this scheme, by introducing a wave velocity adjustment term and combining it with a neural network to parameterize it, the basic wave velocity can be dynamically adjusted according to spatial and temporal coordinates and environmental characteristics to obtain adaptive wave velocity parameters. This allows the model to better adapt to the propagation characteristics of charging pile demand fluctuations under different scenarios and improves prediction accuracy.

[0019] Furthermore, S5 includes the following steps: S51. Construct the prediction loss based on the prediction results, mean squared error, and true value of the wave propagation layer. The mathematical expression is: , This represents the mean square error. The predicted value represents the prediction result of the wave propagation layer. Represents the actual value. Indicates the mean absolute error. express loss, Mean square error The weighting coefficients, Mean absolute error The weighting coefficients, express The weighting coefficients of the loss; S52, Based on velocity field and density field Construct physical consistency loss, physical consistency loss The mathematical expression is: , Denotes the divergence operator, This represents the rate of change of the density field over time. Represents the curl operator, Represents the Laplace operator; S53, based on wave speed and disturbance terms Constructing constraints for the wave equation, wave equation constraints The mathematical expression is: , This represents the second-order partial derivative of the density field with time; S54. Construct the regularization term. The mathematical expression is: , Indicates model parameters, This represents the L2 regularization weighting coefficients used to control the parameters. Represents model parameters The square of the L2 norm, These represent the weight coefficients used to control the smoothing regularization of hidden states. This represents the total number of time steps. Indicates the first Hidden states at each time step Indicates the first Hidden states at each time step These represent the weighting coefficients used to control entropy regularization. The entropy represents the predicted distribution; S55. Construct a multi-task loss function and use a dynamic weight adjustment strategy to adjust the weight coefficients in the multi-task loss function. The mathematical expression is: In the formula, Indicates the loss for physical consistency Adjusted weighting coefficients This indicates the constraint used on the wave equation. Adjusted weighting coefficients Indicates the term used for regularization. The weighting coefficients to be adjusted; The mathematical expression for the dynamic weight adjustment strategy is: In the formula, Indicates the first The weights after adjusting the weight coefficients, Indicates the first The initial weights of each weight coefficient. Indicates the first The adjustment rate of each weight coefficient Indicates the first The current loss value of each weight coefficient. Indicates the first The initial loss value for each weight coefficient.

[0020] In this scheme, prediction accuracy is improved by constructing a prediction loss that combines mean squared error and mean absolute error; physical consistency loss is used to ensure that the prediction conforms to physical laws; a wave equation is constructed to constrain and enhance the simulation of wave characteristics; a regularization term is introduced to prevent overfitting and smooth the hidden state; a multi-task loss function is constructed and a dynamic weight adjustment strategy is adopted, which can flexibly adjust the weights according to different task characteristics and initial loss conditions, thereby comprehensively improving the accuracy of the model in predicting charging pile demand.

[0021] Furthermore, S6 includes the following steps: S61. Construct a time multi-scale representation learning based on hourly, daily, weekly, and monthly scale representation learning, and obtain the time features of each time scale based on the time multi-scale representation learning. S62. Construct a spatial multi-scale representation learning based on point-scale representation learning, road segment-scale representation learning, regional-scale representation learning, and city-scale representation learning, and obtain the spatial features of each spatial scale based on the spatial multi-scale representation learning; S63. A spatiotemporal scale fusion network constructed using an attention mechanism dynamically fuses the temporal features of each time scale and the spatial features of each spatial scale to obtain fused features. The mathematical expression is: In the formula, Represents a spatiotemporal scale fusion network. Represents time characteristics on an hourly scale. Representing the time characteristics on a daily scale, Represents the time characteristics on a weekly scale. Represents the time characteristics on a monthly scale. Representing spatial characteristics at the point scale, Spatial characteristics representing road segment scale. Represents spatial characteristics at the regional scale. Indicates spatial characteristics at the city scale; S64. Construct a cross-scale interaction mechanism based on a message passing network, and exchange information on the fused features through the cross-scale interaction mechanism to obtain interactive enhanced features. The mathematical expression for the information exchange process is: ， In the formula, The message represents the distance from scale i to scale j. This represents the message generation function. Indicates the first Input data at each scale, Indicates the first Features updated at each scale This represents a message aggregation function. Indicates the first Input data at each scale, Representing scale The set of neighbors; S65. Perform consistency verification on the interaction enhancement features through spatiotemporal consistency constraints to obtain consistency verification features and spatiotemporal consistency constraints. The mathematical expression is: ,in, Let i represent the transformation function from scale i to scale j; S66. Predict each scale in the consistency verification features to obtain the prediction result for each scale, and then perform weighted fusion of the prediction results for each scale to obtain the final prediction result. The mathematical expression is: , Indicates the first Prediction results at each scale, Indicates the first Weights for each scale prediction.

[0022] In this scheme, by constructing temporal and spatial multi-scale representation learning, features of different temporal granularities and spatial ranges can be comprehensively captured; a spatiotemporal scale fusion network using an attention mechanism can dynamically fuse multi-scale spatiotemporal features; a cross-scale interaction mechanism based on a message passing network enables effective information exchange to obtain interactive enhanced features; features are verified through spatiotemporal consistency constraints to improve rationality; and weighted fusion of multi-scale prediction results can more accurately uncover the spatiotemporal patterns of charging pile demand and improve prediction accuracy.

[0023] Furthermore, S7 includes the following steps: S71. Obtain the spatiotemporal sequence dataset of the charging pile demand to be tested and standardize it to obtain a standardized dataset to be tested. Divide the standardized dataset to be tested according to the preset dataset division ratio to obtain the corresponding data training set, data validation set and data test set. S72. Input the data training set into the multi-level spatiotemporal representation learning framework for training, and use the Adam optimizer, learning rate scheduling strategy and multi-task loss function for optimization during the training process to obtain the trained multi-level spatiotemporal representation learning framework. S73. Input the data validation set into the trained multi-level spatiotemporal representation learning framework for iterative optimization to obtain the iteratively optimized multi-level spatiotemporal representation learning framework. S74. Input the data test set into the iteratively optimized multi-level spatiotemporal representation learning framework for testing. When the test accuracy of the iteratively optimized multi-level spatiotemporal representation learning framework meets the preset accuracy, output the charging pile demand prediction model.

[0024] In this solution, data quality can be improved by acquiring and standardizing the dataset to be tested. The multi-level spatiotemporal representation learning framework is trained using the Adam optimizer, learning rate scheduling strategy and multi-task loss function, which can accelerate convergence, dynamically adjust the learning rate and comprehensively optimize multiple tasks. The model is iteratively optimized using the validation set to enhance generalization ability. The prediction model is output with the test accuracy as the standard, which can more accurately build the charging pile demand prediction model and improve the reliability of the prediction.

[0025] Furthermore, in step S72, the learning rate scheduling strategy employs a cosine annealing strategy, the mathematical expression of which is: In the formula, Indicates the first The learning rate of the step. This represents the minimum learning rate. This represents the maximum learning rate. This represents the total number of steps.

[0026] In this scheme, by adopting a cosine annealing strategy as the learning rate scheduling strategy, the learning rate can be dynamically changed according to the law of cosine function. In the early stage of training, a large learning rate is maintained to accelerate convergence, and in the later stage, the learning rate is steadily reduced to avoid oscillation, thereby improving the stability and convergence effect of the charging pile demand prediction model training. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating a charging pile demand prediction method based on frequency domain multi-head spatiotemporal attention and wave propagation, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the process for constructing the wave propagation layer in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the process of constructing a multi-level spatiotemporal representation learning framework in an embodiment of the present invention. Detailed Implementation

[0028] The following detailed description illustrates the specific implementation method: like Figure 1 The diagram shown is a flowchart illustrating a charging pile demand prediction method based on frequency domain multi-head spatiotemporal attention and fluctuation propagation according to an embodiment of the present invention, including the following steps: S1. Collect spatiotemporal sequence data of charging pile demand, and preprocess the spatiotemporal sequence data of charging pile demand to obtain spatiotemporal sequence data of actual charging pile demand. S2. The spatiotemporal sequence data of actual charging pile demand is transformed into the frequency domain by fast Fourier transform to obtain the frequency domain representation. A self-calibrated spectral filter is constructed, and the frequency domain representation is enhanced by the self-calibrated spectral filter to obtain the frequency domain enhanced representation. S3. Construct a multi-head spatiotemporal attention mechanism based on frequency domain enhanced representation, and dynamically fuse the outputs of the multi-head attention according to the multi-head spatiotemporal attention mechanism to obtain the multi-head fusion output matrix; S4. Construct a wave propagation layer based on hierarchical high-order differential equations, multi-head fusion output matrix, and stochastic turbulence model; S5. Determine the prediction loss based on the prediction results of the wave propagation layer, and construct a multi-task loss function based on the prediction loss, physical consistency loss, wave equation constraints and regularization terms. S6. Construct a multi-level spatiotemporal representation learning framework through temporal multi-scale representation learning and spatial multi-scale representation learning; S7. Train the multi-level spatiotemporal representation learning framework based on the spatiotemporal sequence dataset of the charging pile demand and the multi-task loss function, and output the multi-level spatiotemporal representation learning framework that meets the preset accuracy as the charging pile demand prediction model.

[0029] In this embodiment, this application collects spatiotemporal sequence data on charging pile demand from multiple sources, including charging facility usage records, traffic flow data, user activity data, and environmental factor data. The spatiotemporal sequence data collection includes collecting charging facility usage records. ,in, Indicates the number of charging stations. Indicates the length of time. express One charging station A real-valued matrix of usage records within a given time period; traffic flow data is also collected. ,in, Indicates the number of road monitoring points. express Each road monitoring point is located at A real-valued matrix of traffic flow over a given time period; collecting user activity data. ,in, Indicates the number of urban functional zones. express Each urban functional area A real matrix of user activity data within a time period; and collection of environmental factor data. ,in, Indicates the number of environmental monitoring stations. express Environmental monitoring stations at A real-value matrix of environmental factor data within a given time period. Data cleaning and outlier handling: The raw data is cleaned, and missing values, outliers, and duplicate records are detected and processed. For missing values, spatiotemporal interpolation is used for imputation; the specific calculation formula is as follows. ,in, Indicates position time The estimated value, This represents the observation at position j at time t. Indicates position Spatial neighbor set, Indicates time Time neighbor set, and These represent spatial weights and temporal weights, respectively. This represents the observation value at location i at time t. Spatial weights are typically determined based on distance, geographical relationship, or other spatial characteristics between locations; closer locations or stronger spatial relationships result in greater weights. Temporal weights are determined based on time intervals and the characteristics of the time series; shorter time intervals or stronger correlations within the time series result in higher weights. For outliers, a detection method based on local anomaly factors is used, combined with domain knowledge for correction. Spatiotemporal feature extraction and enhancement are performed on the preprocessed data. For charging pile usage records, temporal features (such as hourly distribution, daily variation patterns, and weekly variation patterns) and spatial features (such as geographical location, surrounding facilities, and accessibility) are calculated. For traffic flow data, road network topology features and traffic flow change patterns are extracted. For user activity data, activity intensity and distribution features are extracted. For environmental factor data, weather conditions and seasonal variation features are extracted. These extracted features are then used to form an enhanced spatiotemporal feature representation through feature fusion methods. ,in Representation of feature dimensions. Multi-scale spatiotemporal representation construction: Based on the original data and extracted features, a multi-scale spatiotemporal representation is constructed, including hourly scale representation. Daily scale representation Weekly scale representation Lunar scale representation This allows us to capture patterns of change across different time scales. Simultaneously, a point-scale representation is constructed. Road segment scale representation Regional scale representation and city-scale representation ,in, , These represent the number of road segments and the number of areas, respectively. This represents the feature dimension at the point scale. This represents the feature dimension at the road segment scale. Represents the feature dimension at the region scale. This represents the feature dimensions at the city scale. This multi-scale representation provides the model with comprehensive spatiotemporal information. Spatiotemporal relationship modeling is then used to construct a spatial relationship matrix between charging stations. ,in, This represents the spatial relationship strength between node i and node j. Spatial relationship considers Euclidean distance, road network distance, and functional similarity; the calculation formula is as follows: ,in, Represents Euclidean distance. Indicates road network distance, This represents the functional feature vector of node i. This represents the functional feature vector of node j. This represents the similarity of the functional feature vectors of node i and node j. This represents the weighting coefficient used to adjust the exponential decay term based on Euclidean distance. This represents the weighting coefficient used to adjust the exponential decay term based on road network distance. These represent the weighting coefficients used to adjust the similarity of functional feature vectors. This represents a parameter used to adjust the influence of Euclidean distance on the strength of spatial relationships. This represents the parameter used to adjust the influence of road network distance on the strength of spatial relationships; where, Based on domain knowledge, if Euclidean distance is considered to play a dominant role in spatial relationships, a larger value can be assigned, and vice versa. Considering the specific application scenario, if road network distance is more critical in spatial relationships, such as in urban transportation applications, it can be appropriately increased. The value; This is determined by analyzing the correlation between functional feature vectors and spatial relationships. If functional feature similarity has a significant impact on spatial relationships, then... The larger the value of ; The appropriate values ​​are determined by statistical analysis of Euclidean distance data, such as calculating the mean and standard deviation of the Euclidean distance, and then considering the model's sensitivity to distance. Initial settings can be made based on the statistical characteristics of road network distance data, and then adjusted through model optimization. A time relationship matrix should also be constructed. ,in, Indicates time and time The strength of the relationship between them should be considered in terms of periodic patterns and trend characteristics.

[0030] Specifically, in S2, the frequency domain representation The mathematical expression is: In the formula, Represents the Fast Fourier Transform. This represents the spatiotemporal sequence data of charging pile demand. This indicates a self-calibrating spectral filter. This represents the parameters of the self-calibrating spectrum filter; the mathematical expression for the self-calibrating spectrum filter is: In the formula, This represents the transfer function between frequency component i and frequency component j. These represent learnable filtering parameters used to control the bandwidth or sharpness of the filter. These represent learnable filtering parameters used to adjust the gain or amplitude response of the filter. These represent learnable filtering parameters used to adjust the phase response or symmetry of the filter. Indicates the target frequency component. Represents the frequency components of the input signal. Represents the target frequency component and the frequency components of the input signal The frequency distance between them.

[0031] In this embodiment, the spatiotemporal sequence data of actual charging pile demand is obtained through Fast Fourier Transform. Applying Fast Fourier Transform (FFT) in the time dimension transforms the spatiotemporal sequence data of actual charging pile demand into a frequency domain representation. ,in This indicates the number of frequencies. The self-calibration spectrum filter is jointly optimized using low-pass, band-pass, and high-pass filters, resulting in three basic filter types: low-pass filter... bandpass filter and high-pass filter Low-pass filters emphasize low-frequency components, and their calculation formula is as follows: , These represent the parameters of a low-pass filter; a band-pass filter emphasizes the intermediate frequency component, and its calculation formula is... ,in, These represent the parameters of the bandpass filter. The center frequency is used; high-pass filters emphasize high-frequency components, and the calculation formula is as follows: , These parameters represent the strength of the high-pass filter. These three filters are jointly optimized using weighted coefficients to form a composite filter. ,in This represents the learnable weight coefficients of the low-pass filter. This represents the learnable weight coefficients of the bandpass filter. This represents the learnable weight coefficients of the high-pass filter. The determination of each learnable weight coefficient can be achieved during model training by using the weight coefficients as one of the model's parameters and defining a loss function that measures the difference between the model's output and the true value. By calculating the gradient of the loss function with respect to the weight coefficients, gradient descent (such as stochastic gradient descent, Adam, or other optimization algorithms) is used to iteratively update the weight coefficients, gradually decreasing the value of the loss function. This process finds an optimal set of weight coefficients, enabling the model to better fit the data. A frequency domain attention mechanism is introduced to enhance the model's focus on key frequency components. The formula for calculating the frequency domain attention weights is as follows: ,in Represents frequency The corresponding frequency domain representation, , , , For the parameters of the attention mechanism, Let be the hidden layer dimension. The frequency domain attention-enhanced representation is computed as follows: This mechanism allows the model to automatically adjust its focus on different frequency components based on the input data.

[0032] Specifically, S3 includes the following steps: S31. Divide the frequency domain enhancement representation into H heads, determine the preset frequency range that is adapted to each head, and divide each head into the corresponding frequency domain components according to the frequency range adapted to each head. S32. Construct adaptive query, key, and value matrices in the frequency domain components corresponding to each head, and perform spatiotemporal domain transformation on the adaptive query, key, and value matrices according to the inverse Fourier transform to obtain the spatiotemporal domain query matrix, spatiotemporal domain key matrix, and spatiotemporal domain value matrix. S33. Construct a topology-sensitive distance matrix based on the location, distance, and topological relationship between each charging station. The mathematical expression is: In the formula, Represents the topology-sensitive distance matrix The element in the i-th row and j-th column, This represents the Euclidean distance between node i and node j. This represents the directional angle between node i and node j. Represents the distance parameter. Indicates the angle parameter. Indicates topology-sensitive terms, , Represents the normalized diffusion Laplace matrix. These represent the topological parameters used to control the diffusion rate. This represents the topological parameters used to adjust the power of the diffusion process. This represents the diffusion distance between node i and node j. This indicates that the direction depends on the modulation function. , Indicates the direction-dependent strength. Indicates the reference direction angle. S34. Construct a topology-aware attention mask based on the urban road network structure. The mathematical expression of the topology-aware attention mask is: , This represents the element in the i-th row and j-th column of the topology-aware attention mask. Indicates the initial mask. , This represents the shortest path length between node i and node j. This represents the distance threshold for the h-th head. Indicates an indicator function, Indicates the flow impact coefficient. This represents the traffic flow between node i and node j; S35. Query the spatiotemporal domain matrix based on the h-th head. The spatiotemporal key matrix of the h-th head The topological sensitivity distance matrix of the h-th head Topological attention mask of the h-th head A multi-head spatiotemporal attention mechanism is constructed, and its mathematical expression is as follows: , This represents the attention weight of the h-th head. Represents the transpose of the spatiotemporal domain key matrix. Indicates the scaling factor. Indicates the activation function; S36. Calculate the attention output of each head based on the multi-head spatiotemporal attention mechanism to obtain the multi-head attention output. Then, perform concatenation and linear merging operations on the multi-head attention outputs to generate a multi-head fusion output matrix. The mathematical expression for the attention output of the h-th head is: , The mathematical expression for the attention output of the h-th head, the connection operation, and the linear merging is: In the formula, This represents the multi-head fusion output matrix. Indicates a connection operation. This indicates the output for all H headers. Perform a connection operation. Represents the dynamic fusion matrix. , Represents the parameterized dynamic function. Indicates contextual information, This represents the output transformation matrix of the h-th head. This represents the set of output transformation matrices for all H heads.

[0033] In this embodiment, for the h-th head, the query matrix is ​​calculated as follows: The key matrix is ​​calculated as follows: The value matrix is ​​calculated as follows ,in , , Let h be the parameter matrix of the h-th head. , For the query key, an adaptive mask is used. and These represent the dimensions of the key and value, respectively. For spatiotemporal domain transformation, an inverse Fourier transform is performed on the frequency domain query, key, and value matrices to transform them back to the spatiotemporal domain. The spatiotemporal domain query matrix is ​​calculated as follows: The spatiotemporal key matrix is ​​calculated as follows: The spatiotemporal domain value matrix is ​​calculated as follows: ,in This represents the inverse fast Fourier transform. This indicates the time length of the h-th header.

[0034] like Figure 2 As shown, S4 includes the following steps: S41. Construct hierarchical higher-order differential equations; S42. By analyzing the reflection and attenuation behavior of urban boundaries using boundary normal vectors and attenuation coefficients, the urban boundary condition processing mechanism is obtained. The mathematical expression of the urban boundary condition processing mechanism is: In the formula, This represents the boundary condition value at spatial location x and current time t. The reflection matrix represents the spatial location x and the current time t. This represents the fluctuation value at spatial location x and current time t. Represents the computational domain The boundary; S43. Based on the urban boundary condition processing mechanism and the multi-head fusion output matrix, an external condition vector is constructed. Then, based on the hidden state, the external condition vector, and the dynamic function, a hierarchical high-order differential equation state update equation is constructed. The mathematical expression of the state update equation is: In the formula, This indicates a hidden state. Represents hidden state rate of change over time Represents the external condition vector. Represents the parameterized dynamic function; S44. Based on the Gaussian process and diffusion tensor, a stochastic turbulence model is obtained by modeling the stochastic characteristics of charging pile demand changes. The mathematical expression of the stochastic turbulence model is: In the formula, This represents the turbulence term at spatial location x and current time t. Denotes the divergence operator, Represents the diffusion tensor. This represents the gradient of the fluctuation value at spatial location x and current time t. Indicates the diffusion term. Represents random noise; S45. Based on the stochastic turbulence model, the state update equation of the layered high-order differential equation is updated to obtain the wave propagation layer. The mathematical expression for the update process of the wave propagation layer is: In the formula, This indicates the current state of the wave propagation layer. This indicates the state of the wave propagation layer at the next moment. Indicates the time step.

[0035] Specifically, in S43, the dynamic function is parameterized using a hybrid expert system and higher-order differential constraints. The mathematical expression is: In the formula, Indicates the first An expert ODE function, Indicates the first Dynamic weighting function for each expert 1 represents the number of experts. Indicates the first A higher-order differential constraint Indicates the first Dynamic weighting function for a higher-order differential constraint, This indicates the order of the differential constraint. Representing the hidden state in terms of time The first derivative.

[0036] Specifically, the wave velocity parameters in the dynamic function are adjusted based on the wave velocity adjustment term to obtain adaptive wave velocity parameters. The mathematical expression is: Furthermore, the wave velocity adjustment term is parameterized using a neural network. The mathematical expression for the wave velocity adjustment term is as follows: In the formula, Indicates the fundamental wave velocity. This indicates the wave speed adjustment item. Represents spatial coordinates, Represents the time coordinate. Indicates environmental characteristics, This represents the wave velocity parameter network.

[0037] In this embodiment, the reflection matrix The calculation formula is: ,in, Represents the identity matrix. Represents the boundary normal vector. The transpose matrix of the boundary normal vector. Represents the attenuation coefficient. Diffusion tensor The calculation formula is: , Represents the fundamental diffusion tensor. This represents the diffusion adjustment term. The random noise term is generated using a Gaussian process. ,in, Represents the covariance function. This indicates that the expression follows a pattern with a mean of 0 and a covariance function of . The Gaussian process. The formula for calculating higher-order differential constraints is: , Represents the Laplace operator. Indicates the source term.

[0038] Specifically, S5 includes the following steps: S51. Construct the prediction loss based on the prediction results, mean squared error, and true value of the wave propagation layer. The mathematical expression is: , This represents the mean square error. The predicted value represents the prediction result of the wave propagation layer. Represents the actual value. Indicates the mean absolute error. express loss, Mean square error The weighting coefficients, Mean absolute error The weighting coefficients, express The weighting coefficients of the loss; S52, Based on velocity field and density field Construct physical consistency loss, physical consistency loss The mathematical expression is: , Denotes the divergence operator, This represents the rate of change of the density field over time. Represents the curl operator, Represents the Laplace operator; S53, based on wave speed and disturbance terms Constructing constraints for the wave equation, wave equation constraints The mathematical expression is: , This represents the second-order partial derivative of the density field with time; S54. Construct the regularization term. The mathematical expression is: , Indicates model parameters, This represents the L2 regularization weighting coefficients used to control the parameters. Represents model parameters The square of the L2 norm, These represent the weight coefficients used to control the smoothing regularization of hidden states. This represents the total number of time steps. Indicates the first Hidden states at each time step Indicates the first Hidden states at each time step These represent the weighting coefficients used to control entropy regularization. The entropy represents the predicted distribution; S55. Construct a multi-task loss function and use a dynamic weight adjustment strategy to adjust the weight coefficients in the multi-task loss function. The mathematical expression is: In the formula, Indicates the loss for physical consistency Adjusted weighting coefficients This indicates the constraint used on the wave equation. Adjusted weighting coefficients Indicates the term used for regularization. The weighting coefficients to be adjusted; The mathematical expression for the dynamic weight adjustment strategy is: In the formula, Indicates the first The weights after adjusting the weight coefficients, Indicates the first The initial weights of each weight coefficient. Indicates the first The adjustment rate of each weight coefficient Indicates the first The current loss value of each weight coefficient. Indicates the first The initial loss value for each weight coefficient.

[0039] In this embodiment, The calculation formula is: , The formula for calculating is: , The formula for calculating is: . ， , Neural networks representing velocity fields A neural network representing a density field.

[0040] like Figure 3 As shown, S6 includes the following steps: S61. Construct a time multi-scale representation learning based on hourly, daily, weekly, and monthly scale representation learning, and obtain the time features of each time scale based on the time multi-scale representation learning. S62. Construct a spatial multi-scale representation learning based on point-scale representation learning, road segment-scale representation learning, regional-scale representation learning, and city-scale representation learning, and obtain the spatial features of each spatial scale based on the spatial multi-scale representation learning; S63. A spatiotemporal scale fusion network constructed using an attention mechanism dynamically fuses the temporal features of each time scale and the spatial features of each spatial scale to obtain fused features. The mathematical expression is: In the formula, Represents a spatiotemporal scale fusion network. Represents time characteristics on an hourly scale. Representing the time characteristics on a daily scale, Represents the time characteristics on a weekly scale. Represents the time characteristics on a monthly scale. Representing spatial characteristics at the point scale, Spatial characteristics representing road segment scale. Represents spatial characteristics at the regional scale. Indicates spatial characteristics at the city scale; S64. Construct a cross-scale interaction mechanism based on a message passing network, and exchange information on the fused features through the cross-scale interaction mechanism to obtain interactive enhanced features. The mathematical expression for the information exchange process is: ， In the formula, The message represents the distance from scale i to scale j. This represents the message generation function. Indicates the first Input data at each scale, Indicates the first Features updated at each scale This represents a message aggregation function. Indicates the first Input data at each scale, Representing scale The set of neighbors; S65. Perform consistency verification on the interaction enhancement features through spatiotemporal consistency constraints to obtain consistency verification features and spatiotemporal consistency constraints. The mathematical expression is: ,in, Let i represent the transformation function from scale i to scale j; S66. Predict each scale in the consistency verification features to obtain the prediction result for each scale, and then perform weighted fusion of the prediction results for each scale to obtain the final prediction result. The mathematical expression is: , Indicates the first Prediction results at each scale, Indicates the first Weights for each scale prediction.

[0041] In this embodiment, during the dynamic fusion process, the attention weights for each scale representation are first calculated. , , , , and The parameters of the attention network are then represented by the following formula: , Let represent the linear transformation matrix. The [matrix] is... Weights of each scale prediction Adaptive weighting based on uncertainty is adopted, the first Weights of each scale prediction The calculation formula is: , Indicates the first Uncertainty estimation for forecasts at various scales Indicates the first Uncertainty estimation for predictions at various scales.

[0042] Specifically, S7 includes the following steps: S71. Obtain the spatiotemporal sequence dataset of the charging pile demand to be tested and standardize it to obtain a standardized dataset to be tested. Divide the standardized dataset to be tested according to the preset dataset division ratio to obtain the corresponding data training set, data validation set and data test set. S72. Input the data training set into the multi-level spatiotemporal representation learning framework for training, and use the Adam optimizer, learning rate scheduling strategy and multi-task loss function for optimization during the training process to obtain the trained multi-level spatiotemporal representation learning framework. S73. Input the data validation set into the trained multi-level spatiotemporal representation learning framework for iterative optimization to obtain the iteratively optimized multi-level spatiotemporal representation learning framework. S74. Input the data test set into the iteratively optimized multi-level spatiotemporal representation learning framework for testing. When the test accuracy of the iteratively optimized multi-level spatiotemporal representation learning framework meets the preset accuracy, output the charging pile demand prediction model.

[0043] Specifically, in S72, the learning rate scheduling strategy adopts cosine annealing, and the mathematical expression of the cosine annealing strategy is: In the formula, Indicates the first The learning rate of the step. This represents the minimum learning rate. This represents the maximum learning rate. This represents the total number of steps.

[0044] In this embodiment, the calculation formula for standardization is: , This represents the data after standardization. Represents the original data. This represents the mean of the data. The standard deviation of the data is represented. The determination of the preset dataset partition ratio needs to take into account factors such as the amount of data, model complexity, and task requirements. When there is a large amount of data, the proportion of the training set can be larger; for complex models, in order to ensure generalization ability, the proportion of the validation set and the test set should be appropriately increased. In this embodiment, the data training set, data validation set, and data test set are 70%, 15%, and 15% of the standardized test dataset, respectively.

[0045] Based on the above content, conduct experimental analysis: The experimental environment was configured as follows: The hardware platform adopted a distributed computing environment, with the master node configured as NVIDIA A100 GPU (40GB) × 4, Intel Xeon Platinum 8380 processor (40 cores), and 768GB of memory; the slave nodes were configured as NVIDIA V100 GPU (32GB) × 8, Intel Xeon Gold 6348 processor, and 512GB of memory; the network bandwidth was 100Gbps. The software environment used an optimized version of PyTorch 2.1.0, CUDA 11.8, and Python 3.9, and the Horovod distributed training framework was used to accelerate model training. The experimental dataset covered charging pile usage data in three typical cities (representing first-tier cities, emerging cities, and tourist cities, respectively), containing 1,247 charging stations, 2,583,675 charging records, spanning 4 years, with a granularity of 10 minutes. External data includes meteorological data (temperature, humidity, precipitation, wind speed), traffic flow data, urban activity information (holidays, large-scale events), and power grid load data, used to enhance the model's adaptability to complex environments. Evaluation metrics include Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), Prediction Interval Coverage (PICP), Peak Load Prediction Accuracy (PLA), and Emergency Response Index (ERI).

[0046] Comparison Scheme 1: Improved Spatiotemporal Graph Neural Network (ST-GNN+): This method optimizes the traditional spatiotemporal graph neural network by using a dynamic graph structure to capture the changing spatial relationships between charging stations and combining gated recurrent units (GRUs) to process time-series features. It adds an adaptive neighbor sampling mechanism and time-varying graph convolution operations, enabling it to handle large-scale graph structure data and consider dynamic changes in node attributes. The model uses a three-layer graph convolutional network with 128 hidden layers, 4 attention heads, a learning rate of 0.0008, and a batch size of 128. This method represents an advanced application of graph neural networks in spatiotemporal prediction, demonstrating strong capabilities in modeling relationships between nodes. Comparison Scheme 2: Hybrid Physical Information Neural Network (PINN-T): This hybrid model integrates a physical information neural network with a Transformer architecture. By introducing physical knowledge constraints into the deep learning framework, it enhances the model's physical consistency. The PINN-T model includes a Transformer-based feature extraction module and a regularization layer based on physical equation constraints, utilizing partial differential equations to constrain the network's prediction results. The model is configured with an 8-layer Transformer encoder, a hidden dimension of 256, a feedforward network dimension of 1024, and 8 attention heads. Density propagation equations and continuity equations are introduced as physical constraints. The learning rate is set to 0.0005, weight decay to 0.0001, and physical constraint weights to 0.3. This method represents a cutting-edge research direction combining physical information fusion and deep learning, and possesses strong physical interpretability.

[0047] First, the spatiotemporal sequence data of charging pile demand is preprocessed, with a unified time granularity of 10 minutes, and missing value handling, outlier detection, and feature engineering are performed. Feature engineering includes temporal feature extraction (time, day of the week, month, season, holidays), spatial feature extraction (site geographic coordinates, surrounding POI distribution, road network topology), and external factor features (weather, traffic, urban activities, etc.). To address the differences in feature distribution across different cities, feature normalization and adaptive sampling strategies are employed to ensure data quality. Model training uses a hierarchical training strategy: first, the basic components of the model are pre-trained on single-city data; then, overall fine-tuning is performed on mixed multi-city data; finally, adaptive optimization is performed for specific scenarios in each city. The model training parameters are set as follows: the frequency domain conversion module uses adaptive parameter initialization, and the self-calibrating spectral filter... The initial values ​​are [0.05, 0.1, 0.2]. The range is [0.3, 0.5, 0.7]. The values ​​are [0.6, 0.8, 1.0]; the number of heads H in the multi-head spatiotemporal attention mechanism is set to 12, and the topology parameters are... The range is [0.2, 0.3, 0.4]. The values ​​are [1.2, 1.5, 1.8]; the wave propagation layer uses a stack of 3 nonlinear wave equations, the number of mixed experts is 8, and the dynamic weights adopt a soft attention mechanism; the weights of the multi-task loss function are set to [1.2, 1.5, 1.8]. , , An adaptive weight adjustment mechanism was employed. The optimizer used was AdamW with an initial learning rate of 0.0008, employing a cosine annealing strategy with a minimum learning rate of 1e-6, 300 training epochs, and a batch size of 96. Gradient accumulation was used to handle large-batch training. Model training employed mixed precision training and gradient checkpointing to reduce memory consumption, using an early stopping strategy with a tolerance of 25 epochs. Model evaluation used a sliding window testing strategy, with prediction windows ranging from 1 hour to 1 week, analyzing performance changes under different prediction durations. Five test scenarios were designed to assess the model's performance in different application scenarios: regular weekday scenarios (CR), weekend leisure scenarios (WR), peak holiday scenarios (HP), extreme weather scenarios (EW), and large-scale event scenarios (LE). To comprehensively evaluate model performance, multiple ablation experiments were conducted, including removing the frequency domain transformation module, replacing the multi-head spatiotemporal attention mechanism, removing the fluctuation propagation layer, and removing the physical consistency loss, analyzing the contribution of each component to performance. Simultaneously, hyperparameter sensitivity analysis was performed to evaluate the model's stability to changes in key parameters. The experiment employed a multi-dimensional testing method to comprehensively evaluate model performance. Prediction accuracy was evaluated using three standard metrics: MAE (Magnitude of Absolute Error), RMSE (Real-Time Error), which is more sensitive to large errors and reflects model stability, and MAPE (Magnitude of Predictive Error), which assesses relative error and is more suitable for comparing charging stations of different sizes. The calculation formulas are as follows: ; ; ;in, This is the actual value. For predicted values, The sample size is used for multi-temporal and spatial scale performance evaluation. This evaluation is conducted at four temporal granularities (10 minutes, 1 hour, day, week) and three spatial scales (site level, regional level, city level) to analyze the model's adaptability at different scales. Peak load prediction accuracy (PLA) is used to evaluate the model's accuracy in predicting peak periods. The formula is: PLA = (TP + TN) / (TP + TN + FP + FN), where TP, TN, FP, and FN represent the number of samples that correctly predict peak loads, correctly predict non-peak loads, incorrectly predict peak loads, and incorrectly predict non-peak loads, respectively. Uncertainty quantification is assessed using Prediction Interval Coverage (PICP) and Average Prediction Interval Width (MPIW). PICP calculates the proportion of the 95% confidence interval that contains the actual value, and MPIW evaluates the average width of the prediction interval. The Emergency Response Index (ERI) assesses the model's sensitivity and response speed to emergencies. The formula is: ERI = (1 - MAPE_event) × Rt, where MAPE_event is the MAPE during the emergency, and Rt is the proportion of the model's response time to the emergency. Sparse data processing capability is evaluated through prediction performance on low-usage charging stations and newly built charging stations. Computational efficiency is evaluated including training time, inference latency (single sample / batch), throughput (samples / second), memory usage, number of parameters, and FLOPS. Interpretability is evaluated through physical consistency analysis, feature importance analysis, and attention weight visualization. Scenario adaptability is evaluated by calculating performance metrics in five different scenarios and testing cross-city generalization capabilities against multi-city data.

[0048] Table 1 Experimental Results

[0049] As shown in Table 1, experimental results in complex multi-city scenarios further validate the superiority and scalability of the proposed method. In terms of overall prediction accuracy, the proposed method improves MAE, RMSE, and MAPE by 68.2%, 69.8%, and 68.8% respectively compared to the state-of-the-art PINN-T model, indicating that the combination of frequency-domain multi-head spatiotemporal attention and fluctuation propagation can effectively handle more complex charging demand prediction problems in multi-city environments. Particularly in highly variable scenarios such as extreme weather and large-scale events, the MAPE of the proposed method is only 3.82% and 3.95%, respectively, a reduction of 74.0% and 75.7% compared to the PINN-T model, demonstrating the powerful modeling capability of the fluctuation propagation layer for sudden events and abnormal patterns. The sudden event response index reaches 0.94, an improvement of 30.6% compared to the PINN-T model, indicating that the proposed method can quickly identify and respond to sudden changes in demand.

[0050] Multi-temporal and spatial scale analysis shows that the method of this invention performs excellently across different time and spatial scales. As the prediction time scale increases from 10 minutes to the weekly level, the MAPE gradually decreases (3.42%→1.23%), a significantly larger decrease than the comparative scheme, verifying the advantages of frequency domain transformation and multi-head attention structure in capturing long-term dependencies. In terms of spatial scale, the MAE is 1.42kW, 1.21kW, and 0.98kW from the site level to the city level, respectively, indicating that the model can effectively capture the changing patterns at different spatial scales. This is attributed to the accurate modeling of urban spatial structure by the topology-sensitive distance matrix and the wave propagation layer. The peak load prediction accuracy reaches 97.83%, a 14.3% improvement over the PINN-T model, demonstrating the significant advantage of the method of this invention in high load prediction. This has important practical value for load balancing and peak-hour resource scheduling in charging pile operation. Regarding uncertainty quantification, the method of this invention achieves a prediction interval coverage of up to 99.45%, while the average prediction interval width is only 4.32kW, a reduction of 66.1% compared to the PINN-T model. This signifies a significant reduction in prediction uncertainty, providing more reliable support for operational decisions. In terms of sparse data processing, the method of this invention achieves MAEs of only 0.58kW and 0.76kW for sparse sites and newly built sites, respectively, representing reductions of 92.1% and 91.8% compared to the PINN-T model. This demonstrates the excellent generalization ability of the physical constraint model based on the wave equation under data scarcity. This characteristic is significant for solving the cold start problem and predicting new sites, effectively reducing the initial operational risks of newly built charging stations. In terms of computational efficiency, while maintaining high prediction accuracy, the method of this invention achieves a single-sample inference latency of only 5.8ms and a batch processing throughput of 1,453 samples / second, representing improvements of 69.0% and 199.6% respectively compared to the PINN-T model. Memory usage is only 0.32GB, and the number of parameters is reduced by 55.0%. These performance advantages enable the model to run efficiently in resource-constrained environments, supporting large-scale online services and edge computing deployments.

[0051] Cross-city generalization tests show that the MAPE of the proposed method on unseen city data is 5.28%, a 69.8% reduction compared to the PINN-T model. This indicates that the model successfully captures the general physical laws of charging demand and can be effectively applied to new city scenarios, significantly reducing the data requirements and adaptation costs for model migration. Ablation experiments further confirm the importance of each core module: removing the frequency domain transformation module increases MAPE by 43.7%; replacing the multi-head spatiotemporal attention increases MAPE by 37.2%; removing the fluctuation propagation layer increases MAPE by 53.4%; and removing the physical consistency loss increases MAPE by 32.8%. This shows that the fluctuation propagation layer contributes the most to the model performance, consistent with theoretical analysis, verifying the scientific validity and effectiveness of modeling charging demand as a physical fluctuation process.

[0052] This embodiment details the extended application of the charging pile demand forecasting method based on frequency domain multi-head spatiotemporal attention and fluctuation propagation in a smart city integrated management system, particularly its collaborative working mechanism with smart grids and traffic flow forecasting systems. This embodiment demonstrates the superior performance and scalability of the invention in complex multi-source data fusion scenarios. The forecasting system in this embodiment expands the dimensions of the basic input data, integrating not only historical charging pile demand data but also grid load data, traffic flow data, meteorological data, and urban activity information. The system first performs unified preprocessing and alignment of the multi-source heterogeneous data, including time synchronization, spatial registration, and scale normalization. For missing data, the fluctuation propagation layer of this invention is used for physical consistency completion, significantly improving the completion accuracy. Even with a data missing rate of 47%, the completion accuracy remains above 93.6%. In the frequency domain conversion stage, to adapt to the different frequency characteristics of the multi-source data, this embodiment performs multi-domain adaptive extensions to the self-calibrated spectral filter. For the periodic characteristics of grid load data, the filtering capability of low-frequency components is enhanced; for the high-frequency changes in traffic flow data, the filtering effect of the high-frequency filter is strengthened. Specifically, the expression for the multi-domain self-calibration spectrum filter is optimized as follows: ,in For the number of data fields, For domain weight parameters, , , Domain-specific filtering parameters are used. This design enables the system to adaptively process the unique frequency characteristics of different data sources and extract the most valuable information components. This embodiment introduces a cross-domain attention computation framework in the multi-head spatiotemporal attention mechanism, capturing complex correlations between different data types by constructing intra-domain and inter-domain attention matrices. The cross-domain attention computation expression is: ,in and For data field indexing, For cross-domain attention masking, This is an extended topology-sensitive distance matrix. This matrix not only considers spatial distance but also incorporates power grid topology connectivity, transportation network accessibility, and urban functional similarity. The construction formula is as follows: ,in This is the domain correlation matrix, representing the correlation between nodes in different domains.

[0053] In the design of the wave propagation layer, this embodiment constructs a multi-physics coupled wave equation system, jointly modeling power load propagation, traffic flow, and changes in charging demand as interacting physical fields. Specifically, it employs a coupled set of partial differential equations: ,in Indicates the first The state field of each domain For a specific wave velocity in the domain, The inter-domain coupling coefficient, This represents a random disturbance term. By solving this set of coupled equations, the system can more accurately simulate the interaction between charging demand, grid load, and traffic flow in an urban environment, capturing the mutually reinforcing and inhibiting relationships.

[0054] To handle urban boundary conditions, this embodiment introduces an adaptive boundary processing mechanism. Different boundary conditions are designed for urban arterial roads, urban boundaries, and special functional areas: absorbing boundary conditions are used for open boundary areas to eliminate spurious reflections; reflective boundary conditions are used for closed areas to simulate demand rebound effects; and hybrid boundary conditions are used for special functional areas, combining the influence of power grid transmission corridors and major traffic arteries. The mathematical expression of the boundary conditions is as follows: ,in , The location correlation coefficient, For external driving items, This is the boundary normal vector.

[0055] This embodiment also integrates a more complex turbulence propagation model, employing the RANS (Reynolds-averaged Navier-Stokes) equation framework and introducing the Reynolds stress tensor to capture the stochastic fluctuations in charging demand under uncertain conditions. The turbulence model expression is: ,in For the velocity field, The turbulent diffusion coefficient is... For the source term. The dynamic calculation formula for the turbulent diffusion coefficient is: ,in For turbulent kinetic energy, For turbulent dissipation rate, For turbulent Reynolds number, To predict uncertainty, These are model constants. This is an adjustment function. This design enables the system to adapt to fluctuations in charging demand under different traffic conditions, and it performs particularly well during traffic congestion, extreme weather, and special events.

[0056] In terms of model training and optimization, this embodiment adopts a multi-level cascaded learning strategy. First, each data domain is pre-trained independently, and then overall optimization is achieved through cross-domain joint fine-tuning. The multi-task loss function is extended as follows: ,in, For the first Prediction loss for each domain, For cross-domain consistency loss, For the first Domain-specific physical constraints of a domain For multi-domain physical consistency loss, Indicates the first The weighting coefficients of the prediction loss for each domain. The weighting coefficients representing the cross-domain consistency loss. Indicates the first The weighting coefficients of domain-specific physical constraints in each domain. The weighting coefficients represent the multi-domain physical consistency loss. This represents another weight coefficient for the regularization term, where, , , , , By constructing a probabilistic model of the objective function (such as the model's performance metrics on the validation set), the next most promising parameter combination is selected for evaluation based on existing evaluation results. The Bayesian method is used to continuously update the understanding of the objective function, thereby searching for the optimal combination of weight coefficients more efficiently.

[0057] This embodiment further constructs a multi-level spatiotemporal representation learning framework to achieve seamless integration of different time and spatial scales. In the time dimension, it constructs multi-scale time representations including minute, hour, day, and week levels; in the spatial dimension, it constructs multi-scale spatial representations at the charging station level, street level, region level, and city level. The multi-scale representations are fused using an adaptive aggregation function. ,in For time scale sets, For spatial scale sets, For feature representation at a specific spatiotemporal scale, For scale-specific transformation functions, For adaptive fusion weights.

[0058] In practical deployment, this embodiment employs a distributed computing architecture, decomposing the prediction task for large-scale urban areas into multiple sub-regions for parallel computation, and ensuring information exchange between regions through a boundary condition synchronization mechanism. This design significantly improves the system's computational efficiency, enabling real-time prediction of a network of millions of charging piles, reducing computational latency to 8.3% of the original, while maintaining prediction accuracy. Furthermore, the system supports heterogeneous computing devices, dynamically adjusting computational accuracy and resource consumption according to the deployment environment, achieving lightweight deployment on edge devices.

[0059] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A charging pile demand prediction method based on frequency domain multi-head space-time attention and wave propagation, characterized by: Includes the following steps: S1. Collect spatiotemporal sequence data of charging pile demand, and preprocess the spatiotemporal sequence data of charging pile demand to obtain spatiotemporal sequence data of actual charging pile demand. S2. The spatiotemporal sequence data of actual charging pile demand is transformed into the frequency domain by fast Fourier transform to obtain the frequency domain representation. A self-calibrated spectral filter is constructed, and the frequency domain representation is enhanced by the self-calibrated spectral filter to obtain the frequency domain enhanced representation. S3. Construct a multi-head spatiotemporal attention mechanism based on frequency domain enhanced representation, and dynamically fuse the outputs of the multi-head attention according to the multi-head spatiotemporal attention mechanism to obtain the multi-head fusion output matrix; S4. Construct a wave propagation layer based on hierarchical high-order differential equations, multi-head fusion output matrix, and stochastic turbulence model; S5. Determine the prediction loss based on the prediction results of the wave propagation layer, and construct a multi-task loss function based on the prediction loss, physical consistency loss, wave equation constraints and regularization terms. S6. Construct a multi-level spatiotemporal representation learning framework through temporal multi-scale representation learning and spatial multi-scale representation learning; S7. Train the multi-level spatiotemporal representation learning framework based on the spatiotemporal sequence dataset of the charging pile demand and the multi-task loss function, and output the multi-level spatiotemporal representation learning framework that meets the preset accuracy as the charging pile demand prediction model.

2. The charging pile demand prediction method based on frequency domain multi-head spatiotemporal attention and fluctuation propagation according to claim 1, characterized in that: In the S2, the mathematical expression of the frequency domain representation is: , wherein, denotes a fast Fourier transform, denotes the charging pile demand spatiotemporal sequence data, denotes a self-calibration spectral filter, denotes a self-calibration spectral filter parameter; The mathematical expression of the self-calibrating spectral filter is: wherein denotes the transfer function between the frequency component i and the frequency component j, denotes a learnable filter parameter for controlling the bandwidth or the sharpness of the filter, denotes a learnable filter parameter for adjusting the gain or the amplitude response of the filter, denotes a learnable filter parameter for adjusting the phase response or the symmetry of the filter, denotes the target frequency component, denotes the frequency component of the input signal, denotes the target frequency component and the frequency component of the input signal between the target frequency component and the frequency component of the input signal.

3. The method of claim 1, wherein the method is characterized by: S3 includes the following steps: S31. Divide the frequency domain enhancement representation into H heads, determine the preset frequency range that is adapted to each head, and divide each head into the corresponding frequency domain components according to the frequency range adapted to each head. S32. Construct adaptive query, key, and value matrices in the frequency domain components corresponding to each head, and perform spatiotemporal domain transformation on the adaptive query, key, and value matrices according to the inverse Fourier transform to obtain the spatiotemporal domain query matrix, spatiotemporal domain key matrix, and spatiotemporal domain value matrix. S33, constructing a topologically sensitive distance matrix based on the positions, distances and topological relations among the charging sites, the topologically sensitive distance matrix The mathematical expression of the topologically sensitive distance matrix is: , wherein, denotes the topologically sensitive distance matrix the element in the ith row and the jth column, denotes the Euclidean distance between the node i and the node j, denotes the direction angle between the node i and the node j, denotes the distance parameter, denotes the angle parameter, denotes the topologically sensitive term, , denotes the normalized diffusion Laplacian matrix, denotes the topological parameter for controlling the rate of diffusion, denotes the topological parameter for adjusting the power of the diffusion process, denotes the diffusion distance between the node i and the node j, denotes the direction-dependent modulation function, , denotes the direction-dependent intensity, denotes the reference direction angle, S34, constructing a topological perception attention mask based on the urban road network structure, and a mathematical expression of the topological perception attention mask is: , denotes an element of the i-th row and the j-th column of the topological perception attention mask, denotes an initial mask, , denotes a shortest path length between the node i and the node j, denotes a distance threshold of the h-th head, denotes an indicator function, denotes a traffic influence coefficient, denotes a traffic flow between the node i and the node j; S35, the spatio-temporal domain query matrix of the h-th head , the spatio-temporal domain key matrix of the h-th head , the topologically sensitive distance matrix of the h-th head , the topologically aware attention mask of the h-th head The multi-head spatio-temporal attention mechanism is constructed, and a mathematical expression of the multi-head spatio-temporal attention mechanism is: , , the attention weight of the h-th head, , the transpose matrix of the spatio-temporal domain key matrix, , the scaling factor, , the activation function; S36, calculating the attention output of each head based on the multi-head spatio-temporal attention mechanism, obtaining the output of the multi-head attention, and performing connection operation and linear combination on the output of the multi-head attention to generate a multi-head fusion output matrix, wherein the mathematical expression of the attention output of the hth head is: , denotes the attention output of the hth head, and the mathematical expression of the connection operation and linear combination is: , wherein denotes the multi-head fusion output matrix, denotes the connection operation, denotes the connection operation performed on the outputs of all H heads, denotes the connection operation performed on the outputs of all H heads, denotes the dynamic fusion matrix, , denotes the parameterized dynamic function, denotes the context information, denotes the output transformation matrix of the hth head, denotes the set of output transformation matrices of all H heads.

4. The method of claim 3, wherein the method is characterized by: S4 includes the following steps: S41. Construct hierarchical higher-order differential equations; S42. By analyzing the reflection and attenuation behavior of urban boundaries using boundary normal vectors and attenuation coefficients, the urban boundary condition processing mechanism is obtained. The mathematical expression of the urban boundary condition processing mechanism is: wherein represents a boundary condition value at spatial position x and current time t, represents a reflection matrix at spatial position x and current time t, represents a wave value at spatial position x and current time t, represents a boundary of the calculation domain ; S43, constructing an external condition vector based on the city boundary condition processing mechanism and the multi-head fusion output matrix, and constructing a state update equation of a hierarchical high-order differential equation based on the hidden state, the external condition vector, and a dynamic function, a mathematical expression of the state update equation being: , wherein, represents a hidden state, represents a hidden state a change rate over time, represents an external condition vector, represents a parameterized dynamic function; S44. Based on the Gaussian process and diffusion tensor, a stochastic turbulence model is obtained by modeling the stochastic characteristics of charging pile demand changes. The mathematical expression of the stochastic turbulence model is: In the formula, This represents the turbulence term at spatial location x and current time t. Denotes the divergence operator, Represents the diffusion tensor. This represents the gradient of the fluctuation value at spatial location x and current time t. Indicates the diffusion term. Represents random noise; S45. Based on the stochastic turbulence model, the state update equation of the layered high-order differential equation is updated to obtain the wave propagation layer. The mathematical expression for the update process of the wave propagation layer is: In the formula, This indicates the current state of the wave propagation layer. This indicates the state of the wave propagation layer at the next moment. Indicates the time step.

5. The charging pile demand prediction method based on frequency domain multi-head spatiotemporal attention and fluctuation propagation according to claim 4, characterized in that: In step S43, the dynamic function is parameterized using a hybrid expert system and higher-order differential constraints. The mathematical expression is: In the formula, Indicates the first An expert ODE function, Indicates the first Dynamic weighting function for each expert 1 represents the number of experts. Indicates the first A higher-order differential constraint Indicates the first Dynamic weighting function for a higher-order differential constraint, This indicates the order of the differential constraint. Representing the hidden state in terms of time The first derivative.

6. The charging pile demand prediction method based on frequency domain multi-head spatiotemporal attention and fluctuation propagation according to claim 5, characterized in that: The wave velocity parameters in the dynamic function are adjusted based on the wave velocity adjustment term to obtain adaptive wave velocity parameters. The mathematical expression is: Furthermore, the wave velocity adjustment term is parameterized using a neural network. The mathematical expression for the wave velocity adjustment term is as follows: In the formula, Indicates the fundamental wave velocity. This indicates the wave speed adjustment item. Represents spatial coordinates, Represents the time coordinate. Indicates environmental characteristics, This represents the wave velocity parameter network.

7. The charging pile demand prediction method based on frequency domain multi-head spatiotemporal attention and fluctuation propagation according to claim 1, characterized in that: S5 includes the following steps: S51. Construct the prediction loss based on the prediction results, mean squared error, and true value of the wave propagation layer. The mathematical expression is: , This represents the mean square error. The predicted value represents the prediction result of the wave propagation layer. Represents the actual value. Indicates the mean absolute error. express loss, Mean square error The weighting coefficients, Mean absolute error The weighting coefficients, express The weighting coefficients of the loss; S52, Based on velocity field and density field Construct physical consistency loss, physical consistency loss The mathematical expression is: , Denotes the divergence operator, This represents the rate of change of the density field over time. Represents the curl operator, Represents the Laplace operator; S53, based on wave speed and disturbance terms Constructing constraints for the wave equation, wave equation constraints The mathematical expression is: , This represents the second-order partial derivative of the density field with time; S54. Construct the regularization term. The mathematical expression is: , Indicates model parameters, This represents the L2 regularization weighting coefficients used to control the parameters. Represents model parameters The square of the L2 norm, These represent the weight coefficients used to control the smoothing regularization of hidden states. This represents the total number of time steps. Indicates the first Hidden states at each time step Indicates the first Hidden states at each time step These represent the weighting coefficients used to control entropy regularization. The entropy represents the predicted distribution; S55. Construct a multi-task loss function and use a dynamic weight adjustment strategy to adjust the weight coefficients in the multi-task loss function. The mathematical expression is: In the formula, Indicates the loss for physical consistency Adjusted weighting coefficients This indicates the constraint used on the wave equation. Adjusted weighting coefficients Indicates the term used for regularization. The weighting coefficients to be adjusted; The mathematical expression for the dynamic weight adjustment strategy is: In the formula, Indicates the first The weights after adjusting the weight coefficients, Indicates the first The initial weights of each weight coefficient. Indicates the first The adjustment rate of each weight coefficient Indicates the first The current loss value of each weight coefficient. Indicates the first The initial loss value for each weight coefficient.

8. The charging pile demand prediction method based on frequency domain multi-head spatiotemporal attention and fluctuation propagation according to claim 1, characterized in that: S6 includes the following steps: S61. Construct a time multi-scale representation learning based on hourly, daily, weekly, and monthly scale representation learning, and obtain the time features of each time scale based on the time multi-scale representation learning. S62. Construct a spatial multi-scale representation learning based on point-scale representation learning, road segment-scale representation learning, regional-scale representation learning, and city-scale representation learning, and obtain the spatial features of each spatial scale based on the spatial multi-scale representation learning; S63. A spatiotemporal scale fusion network constructed using an attention mechanism dynamically fuses the temporal features of each time scale and the spatial features of each spatial scale to obtain fused features. The mathematical expression is: In the formula, Represents a spatiotemporal scale fusion network. Represents time characteristics on an hourly scale. Representing the time characteristics on a daily scale, Represents the time characteristics on a weekly scale. Represents the time characteristics on a monthly scale. Representing spatial characteristics at the point scale, Spatial characteristics representing road segment scale. Represents spatial characteristics at the regional scale. Indicates spatial characteristics at the city scale; S64. Construct a cross-scale interaction mechanism based on a message passing network, and exchange information on the fused features through the cross-scale interaction mechanism to obtain interactive enhanced features. The mathematical expression for the information exchange process is: ， In the formula, The message represents the distance from scale i to scale j. This represents the message generation function. Indicates the first Input data at each scale, Indicates the first Features updated at each scale This represents a message aggregation function. Indicates the first Input data at each scale, Representing scale The set of neighbors; S65. Perform consistency verification on the interaction enhancement features through spatiotemporal consistency constraints to obtain consistency verification features and spatiotemporal consistency constraints. The mathematical expression is: ,in, Let i represent the transformation function from scale i to scale j; S66. Predict each scale in the consistency verification features to obtain the prediction result for each scale, and then perform weighted fusion of the prediction results for each scale to obtain the final prediction result. The mathematical expression is: , Indicates the first Prediction results at each scale, Indicates the first Weights for each scale prediction.

9. The charging pile demand prediction method based on frequency domain multi-head spatiotemporal attention and fluctuation propagation according to claim 1, characterized in that: S7 includes the following steps: S71. Obtain the spatiotemporal sequence dataset of the charging pile demand to be tested and standardize it to obtain a standardized dataset to be tested. Divide the standardized dataset to be tested according to the preset dataset division ratio to obtain the corresponding data training set, data validation set and data test set. S72. Input the data training set into the multi-level spatiotemporal representation learning framework for training, and use the Adam optimizer, learning rate scheduling strategy and multi-task loss function for optimization during the training process to obtain the trained multi-level spatiotemporal representation learning framework. S73. Input the data validation set into the trained multi-level spatiotemporal representation learning framework for iterative optimization to obtain the iteratively optimized multi-level spatiotemporal representation learning framework. S74. Input the data test set into the iteratively optimized multi-level spatiotemporal representation learning framework for testing. When the test accuracy of the iteratively optimized multi-level spatiotemporal representation learning framework meets the preset accuracy, output the charging pile demand prediction model.

10. The charging pile demand prediction method based on frequency domain multi-head spatiotemporal attention and fluctuation propagation according to claim 9, characterized in that: In step S72, the learning rate scheduling strategy employs a cosine annealing strategy, the mathematical expression of which is: In the formula, Indicates the first The learning rate of the step. This represents the minimum learning rate. This represents the maximum learning rate. This represents the total number of steps.

Images