Electric vehicle charging load high-precision multi-scale space-time probability prediction method and system

Abstract

This invention discloses a high-precision multi-scale spatiotemporal probabilistic prediction method and system for electric vehicle charging load. The method includes establishing a charging load model considering capacity constraints and Markov processes of user behavior; a spatiotemporal characteristic analysis model of charging load based on self-attention mechanism and gated recurrent neural network; a spatiotemporal prediction model based on frequency domain graph convolutional neural network combined with Fourier transform; a Transformer model based on spatiotemporal scale; and using interval prediction evaluation indicators to quantify the uncertainty of spatiotemporal data of charging load from micro and macro dimensions, establishing an uncertainty quantification loss function to model interval prediction for the fluctuation composition of charging load. Finally, a spatiotemporal interval prediction model of charging load based on sequence-to-sequence convolutional long short-term memory neural network is established to obtain multi-scale spatiotemporal probabilistic interval prediction results for charging load at multiple sites. This invention improves the prediction reliability in complex environments.

Description

Technical Field

[0001] This invention relates to the field of smart grid load forecasting and vehicle-grid interactive coordinated control technology, specifically to a high-precision multi-scale spatiotemporal probability prediction method and system for electric vehicle charging load. Background Technology

[0002] The widespread adoption of electric vehicles (EVs) provides an important pathway for energy transition in the transportation sector and the construction of new power systems. However, EV charging behavior is influenced by multiple heterogeneous conditions, including user travel patterns, road network conditions, and weather factors, resulting in highly random and high-frequency fluctuations in charging load across both time and space. Traditional load forecasting methods mostly remain at the level of deterministic "point forecasting," failing to effectively extract the spatial correlation characteristics between adjacent stations in complex road networks, and also unable to scientifically quantify the uncertainty risks brought about by drastic load fluctuations. Once the forecast is significantly inaccurate, it will pose a great threat to the safe and stable operation and dispatch of the power grid.

[0003] Therefore, to achieve safe and efficient power grid operation in a vehicle-grid interactive environment, it is necessary not only to fully explore the spatiotemporal evolution patterns hidden in multi-source data, but also to perform multi-scale decomposition and uncertainty quantification of the spatiotemporal sequence of charging load. By comprehensively evaluating the overall sharpness of the macro-network and the reliability of micro-sites, and replacing single numerical predictions with probability intervals, reliable risk assessment boundaries and decision margins covering both "optimal and worst-case scenarios" can be provided to power grid dispatchers.

[0004] In view of this, it is necessary to provide a method for high-precision multi-scale spatiotemporal probability prediction of electric vehicle charging load using deep learning, which helps to quantify spatiotemporal uncertainties, overcome the limitations of traditional point prediction, and thus ensure the safe and stable dispatch of the power grid. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a high-precision multi-scale spatiotemporal probability prediction method and system for electric vehicle charging load. The aim is to overcome the limitations of traditional deterministic point prediction in the face of complex environmental interference by performing deep feature fusion and probability interval modeling on multi-source charging load data with strong spatiotemporal heterogeneity and fluctuation uncertainty. This provides a decision boundary that balances macroscopic reliability and microscopic sensitivity, which is of great significance for ensuring the safe and stable scheduling of new power systems and promoting the efficient operation of vehicle-to-grid (V2G) interaction.

[0006] According to a first aspect of the present disclosure, a high-precision multi-scale spatiotemporal probability prediction method for electric vehicle charging load is provided, the method comprising the following steps: Establish a charging load model that considers capacity constraints and Markov processes of user behavior, and use queuing theory to realize the transformation of electric vehicle traffic flow into charging facility load; A spatiotemporal characteristic analysis model for charging load based on self-attention mechanism and gated recurrent neural network is established to extract the spatial characteristics of multi-source fusion data and automatically generate the topology of spatiotemporal map of charging load. A spatiotemporal prediction model based on frequency domain graph convolutional neural network and Fourier transform is established to predict the charging load of multiple sites in spatiotemporal terms, and the maximum correlation entropy criterion is introduced to improve the robustness of the model. Establish a Transformer model based on spatiotemporal scales to measure the contribution of multi-source features to charging load tags; The uncertainty of charging load spatiotemporal data is quantified from micro and macro dimensions using interval prediction evaluation indicators. An uncertainty quantification loss function is established to perform interval prediction modeling for the fluctuation composition of charging load. A spatiotemporal prediction model for charging load based on a sequence-to-sequence convolutional long short-term memory neural network was established. The model was learned from the spatiotemporal characteristics of traffic flow and the decomposed data. The model was iteratively updated by combining the uncertainty quantization loss function to obtain multi-scale spatiotemporal probability interval prediction results of charging load at multiple sites.

[0007] A further technical solution of the present invention is as follows: the charging load model is based on a queuing theory model, a user behavior quantification model, a charging system state transition model, and a charging load analytical solution model, wherein, The queuing theory model is based on the stochastic characteristics of vehicle arrival and charging services, and specifically includes: The number of electric vehicles n arriving at the charging station at each time follows a Poisson distribution, and the charging duration of electric vehicles follows a negative exponential distribution. The model uses the number of charging piles C and the maximum number of vehicles the system can accommodate K as capacity constraint parameters. Vehicles exceeding the capacity K are refused entry into the system. The user behavior quantification model is used to quantify the random behavioral characteristics of users queuing and leaving midway, specifically including: Define system state The probability of a user joining the queue Intensity of users leaving midway The system status The total number of electric vehicles charging and waiting in line at the charging station is used to quantify the degree of congestion at the charging station. The charging system state transition model is based on an improved Markov birth-death process, specifically including: Define system state The birth rate of Lower Markov state transitions The kill rate is defined as the product of the basic arrival rate and the user joining probability. The sum of the completion rate and the mid-way departure rate for charging is used, and the transition only exists between adjacent states; Based on birth rate and extinction rate The construction and charging facilities are in a system state. probability The relevant steady-state equilibrium equations; The analytical solution model for charging load is used to calculate the number of electric vehicles charging and the load on charging facilities. By distinguishing between two operating scenarios—a system without queuing and a system with queuing—the number of charging vehicles in each state is summed using probability weighting to obtain the expected number of electric vehicles currently charging at the charging station. The expected number of electric vehicles charging is multiplied by the rated charging power of a single electric vehicle to obtain the load on the charging facility, thus realizing the conversion of traffic flow into charging load.

[0008] A further technical solution of the present invention is: generating the topology of the spatiotemporal graph of charging load, the specific process of which includes: A gated recurrent neural network is used to extract the time-series characteristics of charging load at a single site, capturing the long-term dependence and dynamic change patterns of the load sequence. The charging load time-series data is input into the gated unit, and the gated recurrent neural network filters and memorizes the time-series information through the reset gate and update gate mechanism. Define the hidden state R of the last time step of the gated recurrent neural network as the representation of the entire time series, and input it into the self-attention mechanism; The weight matrix W is obtained through a self-attention mechanism and used as the adjacency matrix in the graph structure to achieve data-driven spatial association modeling. Based on the temporal features extracted by the gated recurrent neural network and the spatial dependencies between sites captured by the self-attention mechanism, a spatiotemporal map of charging load representing the global evolution law is constructed.

[0009] A further technical solution of this invention is as follows: A spatiotemporal prediction model based on a frequency domain graph convolutional neural network and Fourier transform is established to perform spatiotemporal prediction of charging load at multiple sites. The maximum correlation entropy criterion is introduced to improve the robustness of the model. Specifically, this includes: Based on the spatiotemporal diagram of charging load A spatiotemporal prediction model based on frequency domain graph convolutional neural networks and Fourier transform is established, wherein... The node feature matrix represents the input of multi-source data, consisting of time-series representations of the charging load at each site. This is the site association adjacency matrix learned by the self-attention mechanism; A spatiotemporal feature extraction module based on frequency domain graph convolution and graph Fourier transform is established, including: For adjacency matrix Perform normalization processing to construct the normalized graph Laplacian matrix. ; normalized graph Laplacian matrix Perform eigenvalue decomposition to extract spatial frequency eigenvalues ​​of the graph structure. ; Based on the transpose of the eigenvector matrix For multiple inputs Perform a graph Fourier transform to obtain This maps the data from the node spatial domain to the frequency domain, making the univariate time series of each node linearly independent; right Perform Discrete Fourier Transform to transform the time series of each single variable from the time domain to the frequency domain, and then study the spectral structure and variation law of the signal. In the joint frequency domain space, the frequency domain signal after discrete Fourier transform is extracted in the frequency domain by one-dimensional convolution, and the information flow is strictly controlled according to the time position by the gated linear unit, which accelerates parallel operation and deeply extracts the time-series feature pattern. Perform inverse discrete Fourier transform on the robust features extracted by one-dimensional convolution and gated linear unit filtering to transform the univariate time series back into a two-dimensional structure; In the spatial frequency domain, a graph convolution operator with learnable weights is used to filter the spectral matrix to obtain the graph-filtered feature matrix. It then performs a graphical inverse Fourier transform, ultimately transforming the signal from the spectral domain back to the time domain. ; Before performing frequency domain graph convolution feature extraction, a robust loss function is constructed by introducing the maximum correlation entropy learning criterion. The output of the spatiotemporal feature extraction module is fed into the fully connected layer to perform nonlinear mapping of multidimensional features, and output high-precision multi-scale spatiotemporal prediction results.

[0010] A further technical solution of this invention is: establishing a Transformer model based on spatiotemporal scales to measure the contribution of multi-source features to charging load tags, specifically including the following steps: A Transformer model based on spatiotemporal scales is established to represent location information, weather indicators, and historical loads using vector representations. The original multi-source heterogeneous data is divided and mapped according to the network branch structure. Specifically, this includes: using the input embedding layer to map the enhanced features and label sequences to a fixed-dimensional feature space as the branch input of the encoder; using the output embedding layer to map the shifted label sequence to a fixed-dimensional feature space as the branch input of the decoder; and using a variant of position encoding to label the spatiotemporal position information of the data after the embedding layer processing, which is used to record the position information of the sequence information. The encoder and decoder of the Transformer model are established based on the multi-head attention mechanism. Multi-dimensional self-attention calculation and cross-module interactive extraction of spatiotemporal features are performed. A feedforward neural network module is introduced, which uses the internal Dense layer with time distribution to process a large number of features in parallel and carry out effective feature crossing to further reconstruct the vector representation output by the attention module. Based on the vector representation output by the Transformer model, by measuring the contribution of multi-source features to the charging load prediction label, a feature vector suitable for multi-scale spatiotemporal prediction scenarios is output as the feature input of the downstream interval prediction model.

[0011] A further technical solution of the present invention is as follows: Using interval prediction evaluation indicators, the uncertainty of spatiotemporal data of charging load is quantified from both micro and macro dimensions; an uncertainty quantification loss function is established to perform interval prediction modeling for the fluctuation composition of charging load; specifically, the following steps are included: Traffic flow interval prediction and evaluation indicators are used to quantify the spatiotemporal interval of charging load, including sharpness indicators and reliability indicators. The sharpness indicator is used to penalize predictions that exceed the interval, quantifying the uncertainty of all stations from a macro perspective; the reliability indicator is used to micro-adjust the reliability of each station. Based on the performance of weighted prediction, a loss function for quantifying the uncertainty of spatiotemporal prediction of charging load is constructed to balance the weighted sum of spatiotemporal weighted interval score and spatiotemporal weighted average coverage error.

[0012] A further technical solution of the present invention is: the charging load spatiotemporal interval prediction model includes an encoder and a decoder, wherein, The encoder receives the spatiotemporal feature information of traffic flow, uses ConvLSTM as the base model, and performs information learning on the periodic components and fluctuation components after spatiotemporal decomposition, as well as the important features selected by feature selection, and finally outputs a high-dimensional hidden state vector to the decoder. The decoder constructs a periodic prediction component and a fluctuation prediction component for the high-dimensional latent state vector, respectively derives the accurate prediction value of the periodic component and the interval prediction value of the fluctuation component, and performs weighted fusion. The periodic component prediction branch applies an accuracy loss function to constrain the point prediction accuracy, while the fluctuation component prediction branch uses a charging load spatiotemporal prediction uncertainty quantification loss function to guide and constrain the interval prediction. The overall loss is calculated based on the gradient descent algorithm, and the parameters of the spatiotemporal interval prediction model for charging load are iteratively updated to output the multi-scale spatiotemporal probability interval prediction results of charging load at multiple sites.

[0013] According to a second aspect of the present disclosure, a high-precision multi-scale spatiotemporal probability prediction system for electric vehicle charging load is provided, comprising: The charging load model building module is used to establish a charging load model that considers capacity constraints and Markov processes of user behavior, and uses queuing theory to realize the transformation of electric vehicle traffic flow into charging facility load. The module for constructing a spatiotemporal characteristic analysis model of charging load is used to establish a spatiotemporal characteristic analysis model of charging load based on self-attention mechanism and gated recurrent neural network, extract the spatial characteristics of multi-source fused data, and automatically generate the topology of the spatiotemporal map of charging load. The spatiotemporal prediction model construction module is used to establish a spatiotemporal prediction model based on frequency domain graph convolutional neural network and Fourier transform, to perform spatiotemporal prediction of charging load at multiple sites, and introduces the maximum correlation entropy criterion to improve the robustness of the model. The Transformer model building module is used to build a Transformer model based on spatiotemporal scales to measure the contribution of multi-source features to the charging load label. The interval prediction modeling module is used to quantify the uncertainty of charging load spatiotemporal data from micro and macro dimensions using interval prediction evaluation indicators, and to establish an uncertainty quantification loss function to perform interval prediction modeling for the fluctuation composition of charging load. The module for constructing a spatiotemporal prediction model for charging load is used to establish a spatiotemporal prediction model for charging load based on a sequence-to-sequence convolutional long short-term memory neural network. It learns the model from the spatiotemporal characteristics of traffic flow and the decomposed data, and iterates and updates the model using a loss function to obtain multi-scale spatiotemporal probability interval prediction results for charging load at multiple sites.

[0014] According to a third aspect of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described high-precision multi-scale spatiotemporal probability prediction method for electric vehicle charging load.

[0015] According to a fourth aspect of the present disclosure, a non-transitory computer-readable storage medium is provided, wherein computer instructions are stored on the storage medium, and when executed by a processor, the instructions implement the steps of the above-described high-precision multi-scale spatiotemporal probability prediction method for electric vehicle charging load.

[0016] This disclosure provides a high-precision multi-scale spatiotemporal probability prediction method, system, and electronic device for electric vehicle charging load, the advantages of which include: On the one hand, this invention fully considers the impact of random user behavior and facility capacity constraints on the accuracy of the spatiotemporal distribution of charging load, establishes an electric vehicle charging load model based on Markov processes, realizes the scientific transformation from disordered traffic flow to specific charging facility load, and provides a solid physical foundation for the efficient feature extraction and prediction of subsequent deep learning networks.

[0017] On the other hand, this invention proposes a deep learning method based on frequency domain graph convolution and Transformer multi-feature fusion for high-precision analytical modeling of spatiotemporal charging load. This method rationally utilizes graph convolutional neural networks to extract spatiotemporal features from data from different charging stations and innovatively combines Fourier transform to transform time-domain signals to the frequency domain, effectively analyzing and learning the complex feature patterns of the underlying traffic flow charging load data. Simultaneously, a Transformer-based deep learning architecture is introduced into the network for multi-feature fusion learning, effectively measuring the contribution of different features to the prediction label and accurately capturing the correlation, trend, and periodicity within the time series. This method aims to improve prediction reliability in complex environments. Under the condition of deep fusion of multi-source features, it plans and constructs a sequence-to-sequence convolutional long short-term memory neural network for prediction and inference. Using this deep learning prediction method based on multi-source data fusion, extremely complex spatiotemporal evolution information can be transformed into high-precision multi-spatiotemporal scale uncertainty prediction intervals, thereby effectively promoting the development of electric vehicle charging load prediction technology and providing optimal decision-making basis for vehicle-grid interaction and grid safety scheduling.

[0018] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0020] Figure 1 This is a flowchart of the high-precision multi-scale spatiotemporal probability prediction method for electric vehicle charging load in an embodiment of the present invention; Figure 2This is a schematic diagram of the electric vehicle charging load model considering capacity constraints and user Markov behavior in this invention. Figure 3 This is a schematic diagram of the multi-scale spatiotemporal prediction model of charging load based on a robust frequency domain graph convolutional neural network in this invention; Figure 4 This is a schematic diagram of the Transformer model based on spatiotemporal scale in this invention; Figure 5 This is a schematic diagram of the spatiotemporal probability prediction model for charging load based on a sequence-to-sequence convolutional long short-term memory neural network in this invention; Figure 6 This is a structural diagram of the high-precision multi-scale spatiotemporal probability prediction system for electric vehicle charging load in an embodiment of the present invention; Figure 7 This is a schematic diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the present invention are shown in the drawings, not the entire structure.

[0022] Before discussing the exemplary embodiments in more detail, it should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processes, many of these steps can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the steps can be rearranged. The process can be terminated when its operation is complete, but may also have additional steps not included in the figures. The process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0023] This invention, by considering facility capacity constraints and introducing Markov processes to characterize user stochastic behavior, transforms disordered electric vehicle traffic flow into a scientific physical model of the actual load on charging facilities. It utilizes graph convolutional neural networks and Fourier transforms to extract spatiotemporal features of different charging stations and learn underlying data feature patterns. A Transformer-based deep learning network is applied to describe the multidimensional hidden information in the spatiotemporal data of charging load, and to accurately measure the contribution of different features to the prediction label, as well as the correlation, trend, and periodicity within the sequence. Aiming for the highest prediction accuracy under complex environments and the most accurate quantification of multi-spatiotemporal uncertainty intervals, a multi-scale spatiotemporal probability prediction model for electric vehicle charging load is established. A sequence-to-sequence convolutional long short-term memory neural network is used to perform high-precision prediction and deduction of the temporal evolution patterns, spatial heterogeneity characteristics, and uncertainty intervals of multi-scale charging loads under complex environments. This invention fully considers the spatiotemporal heterogeneity and stochastic behavior of charging loads, effectively achieving high-precision interval prediction under multi-source data fusion conditions, which is of great significance for promoting the development of electric vehicle charging load prediction technology and the safe control of new power systems.

[0024] One embodiment is a high-precision multi-scale spatiotemporal probability prediction method for electric vehicle charging load, such as... Figure 1 As shown, the method includes the following steps: S1. Establish a charging load model that considers capacity constraints and Markov processes of user behavior, and use queuing theory to realize the transformation of electric vehicle traffic flow into charging facility load. S2. Establish a spatiotemporal characteristic analysis model of charging load based on self-attention mechanism and gated recurrent neural network, extract the spatial characteristics of multi-source fusion data, and automatically generate the topology of charging load spatiotemporal map; S3. Establish a spatiotemporal prediction model based on frequency domain graph convolutional neural network and Fourier transform to predict the charging load of multiple sites in spatiotemporal terms, and introduce the maximum correlation entropy criterion to improve the robustness of the model. S4. Establish a Transformer model based on spatiotemporal scale to measure the contribution of multi-source features to charging load tags and characterize the correlation, trend and periodicity between spatiotemporal sequences. S5. Utilize interval prediction evaluation indicators to quantify the uncertainty of spatiotemporal data of charging load from both micro and macro dimensions, and establish an uncertainty quantification loss function to perform interval prediction modeling for the fluctuation composition of charging load. S6. Establish a spatiotemporal prediction model for charging load based on a sequence-to-sequence convolutional long short-term memory neural network. Learn the model from the spatiotemporal characteristics of traffic flow and the decomposed data. Iterate and update the model using the uncertainty quantization loss function to obtain multi-scale spatiotemporal probability interval prediction results for charging load at multiple sites.

[0025] In step S1, the charging load model is based on a queuing theory model, a user behavior quantification model, a charging system state transition model, and a charging load analytical solution model. The queuing theory model is based on the stochastic characteristics of vehicle arrival and charging services, and specifically includes: The number of electric vehicles n arriving at the charging station at each time follows a Poisson distribution, and the charging duration of electric vehicles follows a negative exponential distribution. The model uses the number of charging piles C and the maximum number of vehicles the system can accommodate K as capacity constraint parameters. Vehicles exceeding the capacity K are refused entry into the system. The user behavior quantification model is used to quantify the random behavioral characteristics of users queuing and leaving midway, specifically including: Define system state The probability of a user joining the queue Intensity of users leaving midway The system status The total number of electric vehicles charging and waiting in line at the charging station is used to quantify the degree of congestion at the charging station. The charging system state transition model is based on an improved Markov birth-death process, specifically including: Define system state The birth rate of Lower Markov state transitions The kill rate is defined as the product of the basic arrival rate and the user joining probability. The sum of the completion rate and the mid-way departure rate for charging is used, and the transition only exists between adjacent states; Based on birth rate and extinction rate The construction and charging facilities are in a system state. probability The relevant steady-state equilibrium equations; The analytical solution model for charging load is used to calculate the number of electric vehicles charging and the load on charging facilities. By distinguishing between two operating scenarios—a system without queuing and a system with queuing—the number of charging vehicles in each state is summed using probability weighting to obtain the expected number of electric vehicles currently charging at the charging station. The expected number of electric vehicles charging is multiplied by the rated charging power of a single electric vehicle to obtain the load on the charging facility, thus realizing the conversion of traffic flow into charging load.

[0026] Specifically, such as Figure 2 As shown, step S1 specifically includes: S1. Establish a charging load model considering capacity constraints and Markov processes of user behavior, and utilize queuing theory to realize the transformation of electric vehicle traffic flow into charging facility load. S11. Establish a charging load model that considers capacity constraints and Markov processes of user behavior. The model consists of a queuing theory model, a charging system state transition model, and an analytical solution model for the charging load. S12. Establish an M / M / C / N queuing theory model for electric vehicle charging that considers capacity constraints to characterize the stochastic characteristics of vehicle arrival and charging services; S121. The number of electric vehicles (n) arriving at the charging station at each time point follows a Poisson distribution, and the charging duration of the electric vehicles follows a negative exponential distribution:

[0027]

[0028] in, For the average arrival rate of electric vehicles, For single-pile charging service rate, Let the set of possible values ​​for the number of electric vehicles be , The charging time represents an exponential distribution. S122. The model uses the number of charging piles C and the maximum number of vehicles the system can accommodate K as capacity constraint parameters. Vehicles exceeding the capacity K are directly rejected from entering the system. S13. Quantify user random behavior characteristics and construct a quantitative model of user queuing selection and mid-journey departure behavior; S131. System Status The total number of electric vehicles charging and waiting in line at the charging station is used to quantify the congestion level of the charging station and define the system state. The probability of a user joining the queue The expression representing the impact of queue length on users' willingness to join the queue is as follows:

[0029] in, User queuing sensitivity coefficient; S132. Define system state The intensity of users leaving midway The piecewise expression representing the impact of queuing time on user abandonment behavior is as follows:

[0030] in, Sensitivity factor for users leaving midway; S14. Establish a state transition model for the charging system based on an improved Markov birth-death process; S141. Due to the introduction of capacity constraints and user behavior factors, the conclusions of traditional queuing theory cannot be directly applied to the M / M / C / K load model; therefore, based on system state... An improved Markov birth-death process is constructed to effectively solve the M / M / C / K load model; S142. Define system state The birth rate of Lower Markov state transitions The kill rate is the product of the basic reach rate and the probability of user participation. The sum of the completion rate and the mid-course departure rate for charging, with transitions occurring only between adjacent states, is expressed as follows:

[0031]

[0032] S143. Charging facilities are in operation. probability The steady-state equilibrium equation should be satisfied:

[0033] in Indicates from state w Transition to state v The transition rate is denoted by S, where S is the state space of the system.

[0034] For charging facilities, the Markov state transitions have only two forms: the arrival and departure of electric vehicles; therefore, the equilibrium equations can be simplified to:

[0035] S15. Establish an analytical solution model for charging load, and calculate the number of electric vehicles charging and the load on charging facilities. ; S151. Distinguishing between queues With queue For both operating scenarios, the expected number of electric vehicles charging at the charging station is obtained by summing the probability-weighted sums of the number of vehicles charging in each state.

[0036] S152. Multiply the expected number of electric vehicles charging by the rated charging power of a single electric vehicle to obtain the load on the charging facility. This enables the conversion of traffic flow into charging load:

[0037] in, Charging power for a single electric vehicle.

[0038] Step S2 generates the topology of the charging load spatiotemporal graph, and the specific process includes: A gated recurrent neural network is used to extract the time-series characteristics of charging load at a single site, capturing the long-term dependence and dynamic change patterns of the load sequence. The charging load time-series data is input into the gated unit, and the gated recurrent neural network filters and memorizes the time-series information through the reset gate and update gate mechanism. Define the hidden state R of the last time step of the gated recurrent neural network as the representation of the entire time series, and input it into the self-attention mechanism; The weight matrix W is obtained through a self-attention mechanism and used as the adjacency matrix in the graph structure to achieve data-driven spatial association modeling. Based on the temporal features extracted by the gated recurrent neural network and the spatial dependencies between sites captured by the self-attention mechanism, a spatiotemporal map of charging load representing the global evolution law is constructed.

[0039] Step S3 establishes a frequency-domain-based graph convolutional neural network combined with Fourier transform spatiotemporal prediction model to predict the charging load at multiple sites. The maximum correlation entropy criterion is introduced to improve the model's robustness. Specifically, this includes: Based on the spatiotemporal diagram of charging load A spatiotemporal prediction model based on frequency domain graph convolutional neural networks and Fourier transform is established, wherein... The node feature matrix represents the input of multi-source data, consisting of time-series representations of the charging load at each site. This is the site association adjacency matrix learned by the self-attention mechanism; A spatiotemporal feature extraction module based on frequency domain graph convolution and graph Fourier transform is established, including: For adjacency matrix Perform normalization processing to construct the normalized graph Laplacian matrix. ; normalized graph Laplacian matrix Perform eigenvalue decomposition to extract spatial frequency eigenvalues ​​of the graph structure. ; Based on the transpose of the eigenvector matrix For multiple inputs Perform a graph Fourier transform to obtain This maps the data from the node spatial domain to the frequency domain, making the univariate time series of each node linearly independent; right Perform Discrete Fourier Transform to transform the time series of each single variable from the time domain to the frequency domain, and then study the spectral structure and variation law of the signal. In the joint frequency domain space, the frequency domain signal after discrete Fourier transform is extracted in the frequency domain by one-dimensional convolution, and the information flow is strictly controlled according to the time position by the gated linear unit, which accelerates parallel operation and deeply extracts the time-series feature pattern. Perform inverse discrete Fourier transform on the robust features extracted by one-dimensional convolution and gated linear unit filtering to transform the univariate time series back into a two-dimensional structure; In the spatial frequency domain, a graph convolution operator with learnable weights is used to filter the spectral matrix to obtain the graph-filtered feature matrix. It then performs a graphical inverse Fourier transform, ultimately transforming the signal from the spectral domain back to the time domain. ; Before performing frequency domain graph convolution feature extraction, a robust loss function is constructed by introducing the maximum correlation entropy learning criterion. The output of the spatiotemporal feature extraction module is fed into the fully connected layer to perform nonlinear mapping of multidimensional features, and output high-precision multi-scale spatiotemporal prediction results.

[0040] Specifically, such as Figure 3 As shown, steps S2 and S3 specifically include: S2. Establish a spatiotemporal characteristic analysis model of charging load based on self-attention mechanism and gated recurrent neural network, extract the spatial characteristics of multi-source fused data, and automatically generate the topology of the graph, wherein, S21. Utilize a gated recurrent neural network (GRU) to extract the time-series characteristics of single-site charging load, capturing the long-term dependence and dynamic variation patterns of the load sequence; and process the charging load time-series data. The input is fed into the GRU gating unit. GRU uses reset and update gates to filter and memorize temporal information, alleviating the gradient vanishing problem of traditional recurrent neural networks. The formula for GRU is as follows:

[0041]

[0042]

[0043]

[0044] in, Indicates input data, To update the gate, control the amount of historical information passed to the current state; To reset the door, control how historical information is combined with new input information; This is the candidate hidden state; Hide the current state; , , The weight matrix is ​​a learnable weight matrix; It is the sigmoid activation function; S22. Define the hidden state R of the last time step of the GRU unit as the representation of the entire time series, and input it into the self-attention mechanism:

[0045]

[0046]

[0047] Where R is the hidden state at the last time step calculated by the GRU unit, which serves as the input to the self-attention mechanism; Represents the query matrix (Query). Represents the key matrix (Key); and Let represent the learnable linear projection weight matrices corresponding to the query space and key space, respectively. Given a real matrix with N rows and T columns; during network optimization, the system does not directly compare the original hidden state R, but instead... and The high-dimensional linear mapping projects the temporal features of multiple stations onto two independent feature subspaces representing 'node interaction demands' and 'node self-attributes' respectively. This asymmetric mapping mechanism gives the model the ability to adaptively mine deep nonlinear coupling relationships between nodes in extremely complex traffic networks, thereby generating a more accurate and robust spatial dynamic adjacency matrix. S23. The weight matrix W obtained through the self-attention mechanism can be directly used as the adjacency matrix in the graph structure to realize data-driven spatial association modeling. S24. Based on the load temporal features extracted by the GRU and the inter-site spatial dependencies captured by the self-attention mechanism, a spatiotemporal graph of charging load representing the global evolution law is constructed. , The node feature matrix of the spatiotemporal graph represents the input of multi-source data.

[0048] S3. Establish a spatiotemporal prediction model based on frequency domain graph convolutional neural network and Fourier transform to predict the charging load of multiple sites in spatiotemporal terms, and introduce the maximum correlation entropy criterion to improve the robustness of the model. S31. Spatiotemporal diagram of charging load based on the output of step S2 A spatiotemporal prediction model based on frequency domain graph convolutional neural networks and joint Fourier transform is established; among which, The node feature matrix is ​​composed of the time-series representation of the charging load of each site; This is the site association adjacency matrix learned by the self-attention mechanism; S32. Establish a spatiotemporal feature extraction module based on frequency domain graph convolution and graph Fourier transform; S321. For the adjacency matrix Perform normalization processing to construct the normalized graph Laplacian matrix. The calculation formula is:

[0049] in, for An identity matrix of order 1. The degree matrix is ​​a diagonal matrix, and the diagonal elements of the degree matrix are adjacency matrices. The sum of the elements in the corresponding row; through Symmetric normalization is performed to eliminate the impact of absolute differences in the correlation between different charging stations on network stability. S322. Normalized graphical Laplacian matrix Perform eigenvalue decomposition to extract spatial frequency eigenvalues ​​of the graph structure. ;

[0050] in, Let be the eigenvector matrix formed by orthogonal bases. It is an eigenvalue diagonal matrix; S323. Based on the transpose of the eigenvector matrix For multiple inputs Performing a graph Fourier transform (GFT) maps the data from the node spatial domain to the frequency domain, making the univariate time series of each node linearly independent;

[0051] S324. (Regarding) Perform Discrete Fourier Transform (DFT) to transform each univariate time series from the time domain to the frequency domain, and then study the spectral structure and variation law of the signal; S325. In the joint frequency domain space, the frequency domain signal processed by DFT is input into the feature capture mechanism composed of one-dimensional convolution (1D-Conv) and gated linear unit (GLU); the local steady-state mode is extracted in the frequency domain by using 1D-Conv, and the information flow rate is strictly controlled according to the time position by using the characteristic of GLU to accelerate parallel operation and deeply extract the temporal feature mode. S326. Perform Inverse Discrete Fourier Transform (IDFT) on the robust features extracted by filtering to transform the univariate time series back into a two-dimensional structure; S327. In the spatial frequency domain, a graph convolution operator with learnable weights is used to filter the spectral matrix to obtain the graph-filtered feature matrix. And perform the inverse graphical Fourier transform (IGFT) to finally transform the signal from the spectral domain back to the time domain:

[0052] in, To incorporate output features that integrate complex spatiotemporal dependencies; S33. Before performing frequency domain graph convolution feature extraction, the maximum correlation entropy learning criterion (MCC) is introduced to construct a robust loss function; S331. Existing time series regression models typically use least squares error (LSE) as the loss function, which measures the sum of squares of the differences between the predicted and actual data. The calculation formula is as follows:

[0053] Since the LSE criterion is only optimal when the prediction error distribution follows a Gaussian distribution, it can lead to inaccurate model parameter estimation when the sample is contaminated by non-Gaussian noise or outliers. S332. To overcome the above limitations, correlation entropy theory is introduced to replace mean squared error as the loss function. The basic framework of correlation entropy and the Gaussian kernel function used are defined as follows:

[0054]

[0055] in, For sample size, For real data, For predictive data, Here, a Gaussian kernel is used as the kernel function. As a local criterion for similarity measurement, The parameters for controlling the kernel size are used to reasonably limit the residual sensitivity range caused by large outliers; S333. Substitute the Gaussian kernel function into the relevant entropy framework, and combine it with network weight parameters. Model prediction output This yields the final maximum correlation entropy loss function used for training optimization. :

[0056] S334. Within the framework of the Maximum Relevance Entropy Criterion (MCC), the loss function It considers that any two measurements have equal probabilities, which is very effective in resisting various noises, especially non-Gaussian distributions and large outliers, and further improves the robustness of the spatiotemporal prediction model; compared with LSE, MCC has better robustness to non-Gaussian errors or outliers. S34. Feed the output of the spatiotemporal feature extraction module into the fully connected layer to perform nonlinear mapping of multidimensional features and output high-precision multi-scale spatiotemporal prediction results.

[0057] In S4, a Transformer model based on spatiotemporal scales is established to measure the contribution of multi-source features to the charging load label. The specific steps include: A Transformer model based on spatiotemporal scales is established to represent location information, weather indicators, and historical loads using vector representations. The original multi-source heterogeneous data is divided and mapped according to the network branch structure. Specifically, this includes: using the input embedding layer to map the enhanced features and label sequences to a fixed-dimensional feature space as the branch input of the encoder; using the output embedding layer to map the shifted label sequence to a fixed-dimensional feature space as the branch input of the decoder; and using a variant of position encoding to label the spatiotemporal position information of the data after the embedding layer processing, which is used to record the position information of the sequence information. The encoder and decoder of the Transformer model are established based on the multi-head attention mechanism. Multi-dimensional self-attention calculation and cross-module interactive extraction of spatiotemporal features are performed. A feedforward neural network module is introduced, which uses the internal Dense layer with time distribution to process a large number of features in parallel and carry out effective feature crossing to further reconstruct the vector representation output by the attention module. Based on the vector representation output by the Transformer model, by measuring the contribution of multi-source features to the charging load prediction label, a feature vector suitable for multi-scale spatiotemporal prediction scenarios is output as the feature input of the downstream interval prediction model.

[0058] Specifically, such as Figure 4 As shown, step S4 specifically includes: S4. Establish a Transformer model based on spatiotemporal scale to measure the contribution of multi-source features to charging load tags and characterize the correlation, trend and periodicity between spatiotemporal sequences; S41. Establish a Transformer model based on spatiotemporal scale to perform vector representation of the load sequence after spatiotemporal decomposition (i.e., periodic components and fluctuating components) and important traffic flow features after feature selection and purification. S42. Divide and map the original multi-source heterogeneous data according to the network branch structure; S421. Use an input embedding layer to map the enhanced features and label sequences to a fixed-dimensional feature space, which serves as the branch input for the encoder. S422. The output embedding layer is used to map the shifted label sequence to a fixed-dimensional feature space, which serves as the branch input of the decoder. S423. Using a variant of position encoding, spatiotemporal position information is labeled onto the data processed by the embedding layer to record the position information of the sequence information:

[0059]

[0060] in, Indicates the position of the sequence data. Dimension index representing the feature This represents the total dimension of the feature vector; S43. Establish the Encoder and Decoder modules of the Transformer model, and perform multidimensional self-attention calculation and cross-module interactive extraction of spatiotemporal features; S431. Construct a multi-head attention mechanism in the encoder to map the augmented features with location labels to the label sequence as a query matrix. Key matrix Sum matrix By integrating multiple self-attention structures to embed and represent the input data in different dimensions, the representation of sequence information is considered from different perspectives. S432. Construct a masked multi-head attention mechanism in the Decoder. Take the shifted label sequence with position labels as input, and use a mask to mask the information to be predicted to prevent the leakage of future label information and complete the vector representation of the historical sequence. S433. Enable data interaction between the Encoder and Decoder by building a cross-module multi-head attention mechanism within the Decoder; this mechanism receives the output of the Encoder module as a query matrix. Bond matrix And receive the output of the masked multi-head attention module as a value matrix. ; S434. In the above multi-head attention mechanism, the scale dot-product attention mechanism is used to focus on the more important parts of the sequence information, and different weights are assigned to features of different dimensions. The attention calculation formula is as follows:

[0061] in, Let be the dimension of the key matrix. This is a scaling factor used to prevent the function from entering the gradient saturation region; S435. Introduce a feedforward neural network module, which uses internal Dense layers with time distribution to process a large number of features in parallel and perform effective feature crossing, and further reconstructs the vector representation output by the attention module. S436. An Add&Norm structure is introduced at the output of each network sublayer (including multi-head self-attention module, masked multi-head attention module and feedforward neural network module) of the Encoder and Decoder; the Add&Norm structure includes a skip connection module and layer normalization, which alleviates the gradient vanishing problem in deep network training by using a residual network and realizes layer normalization of feature distribution. S44. The fully connected layer (Linear) of the Decoder outputs the vector representation reconstructed by the Transformer model. By measuring the contribution of multi-source features to the charging load prediction label, the correlation, trend and periodicity between sequences and within sequences are deeply characterized. Finally, a feature vector suitable for multi-scale spatiotemporal prediction scenarios is output as the feature input of the downstream interval prediction model.

[0062] Step S5 utilizes interval prediction evaluation indicators to quantify the uncertainty of spatiotemporal data of charging load from both micro and macro perspectives, and establishes an uncertainty quantification loss function to model interval prediction for the fluctuation composition of charging load. Specifically, this includes the following steps: Traffic flow interval prediction and evaluation indicators are used to quantify the spatiotemporal interval of charging load, including sharpness indicators and reliability indicators. The sharpness indicator is used to penalize predictions that exceed the interval, quantifying the uncertainty of all stations from a macro perspective; the reliability indicator is used to micro-adjust the reliability of each station. Based on the performance of weighted prediction, a loss function for quantifying the uncertainty of spatiotemporal prediction of charging load is constructed to balance the weighted sum of spatiotemporal weighted interval score and spatiotemporal weighted average coverage error.

[0063] The charging load spatiotemporal interval prediction model in step S6 includes an encoder and a decoder, wherein, The encoder receives the spatiotemporal feature information of traffic flow, uses ConvLSTM as the base model, and performs information learning on the periodic components and fluctuation components after spatiotemporal decomposition, as well as the important features selected by feature selection, and finally outputs a high-dimensional hidden state vector to the decoder. The decoder constructs a periodic prediction component and a fluctuation prediction component for the high-dimensional latent state vector, respectively derives the accurate prediction value of the periodic component and the interval prediction value of the fluctuation component, and performs weighted fusion. The periodic component prediction branch applies an accuracy loss function to constrain the point prediction accuracy, while the fluctuation component prediction branch uses a charging load spatiotemporal prediction uncertainty quantification loss function to guide and constrain the interval prediction. The overall loss is calculated based on the gradient descent algorithm, and the parameters of the spatiotemporal interval prediction model for charging load are iteratively updated to output the multi-scale spatiotemporal probability interval prediction results of charging load at multiple sites.

[0064] Specifically, such as Figure 5 As shown, steps S5 and S6 specifically include: S5. Utilize interval prediction evaluation indicators to quantify the uncertainty of spatiotemporal data of charging load from both micro and macro perspectives, and establish an uncertainty quantification loss function to perform interval prediction modeling for the fluctuation composition of charging load; S51. Based on the fluctuation composition of charging load, establish a quantification loss function for the spatiotemporal prediction uncertainty of charging load to quantify the uncertainty of spatiotemporal data of charging load; use traffic flow interval prediction evaluation index to quantify the spatiotemporal interval of charging load, and calculate the reliability index and sharpness index of prediction respectively. S52. The Sharpness Index (IS) is used to penalize predictions that exceed the range, quantifying the uncertainty of all stations from a macro perspective. Its calculation formula is as follows:

[0065]

[0066]

[0067] in, Represents actual traffic flow observations. and These are the upper and lower bounds of the prediction interval, respectively. and These are the penalties for when the true value is below the lower bound and above the upper bound, respectively. This is an indicator function that returns 1 if the condition within the parentheses is true, and 0 otherwise. This indicates a custom confidence level.

[0068] S53. The reliability index (ACE) is used to fine-tune the reliability of each site, and its calculation formula is as follows:

[0069]

[0070]

[0071] in, This represents the actual load value. For the prediction interval, It is a Boolean variable. For the sample size, The target confidence level; S54. Considering the differences caused by the geographical locations of different charging stations, a loss function is constructed based on weighted prediction performance to improve the priority of interval prediction on reliability and increase the model prediction accuracy. A loss function for quantifying the uncertainty of spatiotemporal prediction of charging load is established, and its formula is as follows:

[0072]

[0073] in, This represents the number of nodes, or observation stations, in the spatiotemporal dimension. Indicates the number of samples. and For the corresponding dimension, the upper and lower bounds of the prediction are... and For the introduced weighting coefficients, To balance the spatiotemporal weighted interval score Spatiotemporal weighted average coverage error The weight parameters, This is a penalty for exceeding the lower bound. This is a penalty for exceeding the upper limit; This indicates the reliability of the predicted interval coverage; for the nth observation station at the m-th sample time, the model will provide a predicted interval [ , The real world will provide a true observation of traffic flow. If the true value < Triggering penalty = - ,on the contrary = 0; if > Triggering penalty = - Conversely, it is 0.

[0074] S55. The weighted loss function guided by the uncertainty quantification evaluation index will be applied in the sequence-to-sequence convolutional long short-term memory neural network, and its expression is:

[0075] in, Weights are used as buffer parameters for further balancing. and .

[0076] S6. Establish a spatiotemporal interval prediction model for charging load based on a sequence-to-sequence convolutional long short-term memory neural network, learn the model from the spatiotemporal characteristics of traffic flow and the decomposed data, and iteratively update the model by combining the uncertainty quantification loss function to obtain multi-scale spatiotemporal probability interval prediction results for charging load at multiple sites.

[0077] S61. Establish a spatiotemporal prediction model for charging load based on a sequence-to-sequence convolutional long short-term memory neural network. The model includes an encoder and a decoder, and learns the feature information of different components at different times using ConvLSTM as the base model. S62. Using the high-dimensional reconstructed feature vector representation output by the encoder in step S4, information learning is performed on the periodic components and fluctuation components after spatiotemporal decomposition, as well as the important features selected by feature selection, using the ConvLSTM basis model.

[0078] S621. The ConvLSTM replaces the matrix multiplication of the traditional LSTM with convolution operations, and jointly receives the spatiotemporal data of the current time step with the hidden state of the previous time step to capture spatial correlation features. The expression formula for its internal information transmission is as follows:

[0079]

[0080]

[0081]

[0082]

[0083]

[0084] in, , , Control Input gate, forget gate, and output gate for each time state; and The memory cell states and hidden states that participate in information transmission at different times; and These represent convolution and Hadamard product operations, respectively. It is an activation function. It is the feature input at the current time. It is the hidden state from the previous moment. These are the input gate convolution weights. These are the convolution weights of candidate memory cells. These are the convolutional weights of the final CNN layer. These correspond to the biases of the input gate, forget gate, and output gate, respectively. It is the bias of the final CNN layer; S622. After learning the spatiotemporal information through multiple layers of ConvLSTM, the hidden state is... The information is fed into the final CNN convolutional layer for final information learning and mapped into a high-dimensional latent state vector, which is then passed to the decoder. S63. Using the decoder to receive the high-dimensional hidden state vector, construct different component predictors, namely, a periodic component prediction component and a fluctuating component prediction component; derive the accuracy prediction value of the periodic component and the interval prediction value of the fluctuating component from different branches of the charging load spatiotemporal interval prediction model, respectively, with the following function mapping formula:

[0085]

[0086] in, This represents the functional mapping relationship of the spatiotemporal prediction model for the charging load; and These represent the accuracy prediction value of the periodic component and the range prediction value of the fluctuation component at future time steps, respectively. This represents the time-series data of periodic components within a historical time window. Key characteristics of periodic components representing historical time windows. Key characteristics of periodic components representing future time windows; These represent the time-series data and key feature inputs of the fluctuation component within historical and future time windows, respectively. S64. During the overall model learning process, differential constraints are applied to the two prediction branches mentioned above; S641. For the prediction branch of the periodic component, apply the accuracy loss function to constrain the point prediction accuracy:

[0087] in, This represents the actual observed value of the periodic component charging load at that moment; This represents the point prediction value of the periodic component at the corresponding time point, derived from the prediction model; the formula uses a variant of the mean squared error (MSE) as the accuracy loss function, aiming to minimize the predicted value through the algorithm. Compared with the true value The squared difference between the two is used to strictly constrain and optimize the absolute point prediction accuracy of the prediction branches of the model periodic components.

[0088] S642. For the prediction branch of the fluctuation component, apply the uncertainty quantification loss function established in step S5, which combines sharpness and reliability indices. Provide guidance and constraints for interval forecasting; S65. The output information of the different predictors above is weighted and fused. The weighting fusion formula is as follows:

[0089] in, This represents the predicted value for future time steps after fusion; and These are the fusion weight parameters for the corresponding periodic components and fluctuation components, respectively; S66. Calculate the overall loss based on the gradient descent algorithm and iteratively update the model parameters to output the final multi-site charging load multi-scale spatiotemporal probability interval prediction results that combine macro-level road network reliability and micro-level site sensitivity.

[0090] Another embodiment provides a high-precision multi-scale spatiotemporal probability prediction system for electric vehicle charging load, such as... Figure 6 As shown, system 600 includes: The charging load model building module 610 is used to establish a charging load model that considers capacity constraints and Markov processes of user behavior, and uses queuing theory to realize the transformation of electric vehicle traffic flow into charging facility load. The charging load spatiotemporal characteristic analysis model construction module 620 is used to establish a charging load spatiotemporal characteristic analysis model based on self-attention mechanism and gated recurrent neural network, extract the spatial characteristics of multi-source fusion data, and automatically generate the topology of charging load spatiotemporal map. The spatiotemporal prediction model construction module 630 is used to establish a spatiotemporal prediction model based on frequency domain graph convolutional neural network and Fourier transform, to perform spatiotemporal prediction of charging load at multiple sites, and introduces the maximum correlation entropy criterion to improve the robustness of the model. Transformer model building module 640 is used to build a Transformer model based on spatiotemporal scale to measure the contribution of multi-source features to the charging load label. The interval prediction modeling module 650 is used to quantify the uncertainty of charging load spatiotemporal data from micro and macro dimensions using interval prediction evaluation indicators, and to establish an uncertainty quantification loss function to perform interval prediction modeling for the fluctuation composition of charging load. The charging load spatiotemporal interval prediction model construction module 660 is used to establish a charging load spatiotemporal interval prediction model based on a sequence-to-sequence convolutional long short-term memory neural network. It learns the model from the spatiotemporal characteristics of traffic flow and the decomposed data, and iterates and updates the model in combination with the loss function to obtain multi-scale spatiotemporal probability interval prediction results of charging load at multiple sites.

[0091] In addition to the modules described above, the high-precision multi-scale spatiotemporal probability prediction system 600 for electric vehicle charging load may also include other components. However, since these components are not related to the content of this disclosure, their illustrations and descriptions are omitted here.

[0092] The other specific working processes of the electric vehicle charging load high-precision multi-scale spatiotemporal probability prediction method using the above-mentioned electric vehicle charging load high-precision multi-scale spatiotemporal probability prediction system 600 are described in the above-mentioned embodiment of the electric vehicle charging load high-precision multi-scale spatiotemporal probability prediction method, and will not be repeated here.

[0093] Another embodiment illustrating that the system of the present invention can also be achieved by means of... Figure 7 The architecture of the computing device shown is used to implement this. Figure 7 The architecture of the computing device is shown. For example... Figure 7 As shown, the computer system 710 includes a system bus 730, one or more CPUs 740, input / output 720, and memory 750. The memory 750 can store various data or files used for computer processing and / or communication, as well as program instructions executed by the CPU, including a high-precision multi-scale spatiotemporal probability prediction method for electric vehicle charging load. Figure 7 The architecture shown is merely exemplary and should be adjusted according to actual needs when implementing different devices. Figure 7 One or more components are included. The memory 750, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as the program instructions / modules corresponding to the high-precision multi-scale spatiotemporal probability prediction method for electric vehicle charging load described above in this embodiment of the invention. One or more CPUs 740 execute various functional applications and data processing of the system of the present invention by running the software programs, instructions, and modules stored in the memory 750.

[0094] Of course, the server provided in the embodiments of the present invention is not limited to executing the method operations described above, but can also execute related operations in the high-precision multi-scale spatiotemporal probability prediction method for electric vehicle charging load provided in any embodiment of the present invention.

[0095] The memory 750 may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function; the data storage area may store data created based on terminal usage. Furthermore, the memory 750 may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory, or other non-volatile solid-state storage device. In some instances, the memory 750 may further include memory remotely configured relative to one or more CPUs 740, which can be connected to the device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0096] Input / output 720 can be used to receive input digital or character information, and to generate key signal inputs related to user settings and function control of the device. Input / output 720 may also include a display device such as a display screen.

[0097] This invention also provides a non-transitory computer-readable storage medium storing a computer program. When executed by a processor, this computer program implements the high-precision multi-scale spatiotemporal probability prediction method for electric vehicle charging load described in the above embodiments. The computer-readable storage medium of this invention can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. For example, a computer-readable storage medium can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0098] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0099] The program code contained on the storage medium can be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0100] Furthermore, other specific working processes of a non-temporary computer-readable storage medium are described in the above-described embodiment of the high-precision multi-scale spatiotemporal probability prediction method for electric vehicle charging load, and will not be repeated here.

[0101] In this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a step or method that comprises a list of elements includes not only those elements but also other elements not expressly listed or inherent to such a step or method.

[0102] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A high-precision multi-scale spatiotemporal probability prediction method for electric vehicle charging load, characterized in that, The method includes the following steps: Establish a charging load model that considers capacity constraints and Markov processes of user behavior, and use queuing theory to realize the transformation of electric vehicle traffic flow into charging facility load; A spatiotemporal characteristic analysis model for charging load based on self-attention mechanism and gated recurrent neural network is established to extract the spatial characteristics of multi-source fusion data and automatically generate the topology of spatiotemporal map of charging load. A spatiotemporal prediction model based on frequency domain graph convolutional neural network and Fourier transform is established to predict the charging load of multiple sites in spatiotemporal terms, and the maximum correlation entropy criterion is introduced to improve the robustness of the model. Establish a Transformer model based on spatiotemporal scales to measure the contribution of multi-source features to charging load tags; The uncertainty of spatiotemporal data of charging load is quantified from micro and macro dimensions using interval prediction evaluation indicators. An uncertainty quantification loss function is established to perform interval prediction modeling for the fluctuation composition of charging load. A spatiotemporal prediction model for charging load based on a sequence-to-sequence convolutional long short-term memory neural network was established. The model was learned from the spatiotemporal characteristics of traffic flow and the decomposed data. The model was iteratively updated by combining the uncertainty quantization loss function to obtain multi-scale spatiotemporal probability interval prediction results of charging load at multiple sites.

2. The high-precision multi-scale spatiotemporal probability prediction method for electric vehicle charging load according to claim 1, characterized in that, The charging load model is based on a queuing theory model, a user behavior quantification model, a charging system state transition model, and a charging load analytical solution model. The queuing theory model is based on the stochastic characteristics of vehicle arrival and charging services, and specifically includes: The number of electric vehicles n arriving at the charging station at each time follows a Poisson distribution, and the charging duration of electric vehicles follows a negative exponential distribution. The model uses the number of charging piles C and the maximum number of vehicles the system can accommodate K as capacity constraint parameters. Vehicles exceeding the capacity K are refused entry into the system. The user behavior quantification model is used to quantify the random behavioral characteristics of users queuing and leaving midway, specifically including: Define system state The probability of a user joining the queue Intensity of users leaving midway The system status The total number of electric vehicles charging and waiting in line at the charging station is used to quantify the degree of congestion at the charging station. The charging system state transition model is based on an improved Markov birth-death process, specifically including: Define system state The birth rate of Lower Markov state transitions The kill rate is defined as the product of the basic arrival rate and the user joining probability. The sum of the completion rate and the mid-way departure rate for charging is used, and the transition only exists between adjacent states; Based on birth rate and extinction rate The construction and charging facilities are in a system state. probability The relevant steady-state equilibrium equations; The analytical solution model for charging load is used to calculate the number of electric vehicles charging and the load on charging facilities. By distinguishing between two operating scenarios—a system without queuing and a system with queuing—the number of charging vehicles in each state is summed using probability weighting to obtain the expected number of electric vehicles currently charging at the charging station. The expected number of electric vehicles charging is multiplied by the rated charging power of a single electric vehicle to obtain the load on the charging facility, thus realizing the conversion of traffic flow into charging load.

3. The high-precision multi-scale spatiotemporal probability prediction method for electric vehicle charging load according to claim 1, characterized in that, The topology of the charging load spatiotemporal graph is generated, and the specific process includes: A gated recurrent neural network is used to extract the time-series characteristics of charging load at a single site, capturing the long-term dependence and dynamic change patterns of the load sequence. The charging load time-series data is input into the gated unit, and the gated recurrent neural network filters and memorizes the time-series information through the reset gate and update gate mechanism. Define the hidden state R of the last time step of the gated recurrent neural network as the representation of the entire time series, and input it into the self-attention mechanism; The weight matrix W is obtained through a self-attention mechanism and used as the adjacency matrix in the graph structure to achieve data-driven spatial association modeling. Based on the temporal features extracted by the gated recurrent neural network and the spatial dependencies between sites captured by the self-attention mechanism, a spatiotemporal map of charging load representing the global evolution law is constructed.

4. The high-precision multi-scale spatiotemporal probability prediction method for electric vehicle charging load according to claim 1, characterized in that, A spatiotemporal prediction model based on frequency domain graph convolutional neural network and Fourier transform is established to predict the charging load of multiple sites. The maximum correlation entropy criterion is introduced to improve the robustness of the model. Specifically, this includes: Based on the spatiotemporal diagram of charging load A spatiotemporal prediction model based on frequency domain graph convolutional neural networks and Fourier transform is established, wherein... The node feature matrix represents the input of multi-source data, consisting of time-series representations of the charging load at each site. This is the site association adjacency matrix learned by the self-attention mechanism; A spatiotemporal feature extraction module based on frequency domain graph convolution and graph Fourier transform is established, including: For adjacency matrix Perform normalization processing to construct the normalized graph Laplacian matrix. ; normalized graph Laplacian matrix Perform eigenvalue decomposition to extract spatial frequency eigenvalues ​​of the graph structure. ; Based on the transpose of the eigenvector matrix For multiple inputs Perform a graph Fourier transform to obtain This maps the data from the node spatial domain to the frequency domain, making the univariate time series of each node linearly independent; right Perform Discrete Fourier Transform to transform the time series of each single variable from the time domain to the frequency domain, and then study the spectral structure and variation law of the signal. In the joint frequency domain space, the frequency domain signal after discrete Fourier transform is extracted into local steady-state mode by one-dimensional convolution, and the information flow is strictly controlled according to the time position by the gated linear unit, which accelerates parallel operation and deeply extracts time-series feature mode. Perform inverse discrete Fourier transform on the robust features extracted by one-dimensional convolution and gated linear unit filtering to transform the univariate time series back into a two-dimensional structure; In the spatial frequency domain, a graph convolution operator with learnable weights is used to filter the spectral matrix to obtain the graph-filtered feature matrix. It then performs a graphical inverse Fourier transform, ultimately transforming the signal from the spectral domain back to the time domain. ; Before performing frequency domain graph convolution feature extraction, a robust loss function is constructed by introducing the maximum correlation entropy learning criterion. The output of the spatiotemporal feature extraction module is fed into the fully connected layer to perform nonlinear mapping of multidimensional features, and output high-precision multi-scale spatiotemporal prediction results.

5. The high-precision multi-scale spatiotemporal probability prediction method for electric vehicle charging load according to claim 1, characterized in that, A Transformer model based on spatiotemporal scales is established to measure the contribution of multi-source features to charging load labels. This includes the following steps: A Transformer model based on spatiotemporal scales is established to represent location information, weather indicators, and historical loads using vector representations. The original multi-source heterogeneous data is divided and mapped according to the network branch structure. Specifically, this includes: using the input embedding layer to map the enhanced features and label sequences to a fixed-dimensional feature space as the branch input of the encoder; using the output embedding layer to map the shifted label sequence to a fixed-dimensional feature space as the branch input of the decoder; and using a variant of position encoding to label the spatiotemporal position information of the data after the embedding layer processing, which is used to record the position information of the sequence information. The encoder and decoder of the Transformer model are established based on the multi-head attention mechanism. Multi-dimensional self-attention calculation and cross-module interactive extraction of spatiotemporal features are performed. A feedforward neural network module is introduced, which uses the internal Dense layer with time distribution to process a large number of features in parallel and carry out effective feature crossing to further reconstruct the vector representation output by the attention module. Based on the vector representation output by the Transformer model, by measuring the contribution of multi-source features to the charging load prediction label, a feature vector suitable for multi-scale spatiotemporal prediction scenarios is output as the feature input of the downstream interval prediction model.

6. The high-precision multi-scale spatiotemporal probability prediction method for electric vehicle charging load according to claim 1, characterized in that, The uncertainty of charging load spatiotemporal data is quantified from both micro and macro perspectives using interval prediction evaluation indicators. An uncertainty quantification loss function is established to model interval prediction for the fluctuation components of charging load. The specific steps include: Traffic flow interval prediction and evaluation indicators are used to quantify the spatiotemporal interval of charging load, including sharpness indicators and reliability indicators. The sharpness indicator is used to penalize predictions that exceed the interval, quantifying the uncertainty of all stations from a macro perspective; the reliability indicator is used to micro-adjust the reliability of each station. Based on the performance of weighted prediction, a loss function for quantifying the uncertainty of spatiotemporal prediction of charging load is constructed to balance the weighted sum of spatiotemporal weighted interval score and spatiotemporal weighted average coverage error.

7. The high-precision multi-scale spatiotemporal probability prediction method for electric vehicle charging load according to claim 6, characterized in that, The spatiotemporal prediction model for charging load includes an encoder and a decoder, wherein, The encoder receives the spatiotemporal feature information of traffic flow, uses ConvLSTM as the base model, and performs information learning on the periodic components and fluctuation components after spatiotemporal decomposition, as well as the important features selected by feature selection, and finally outputs a high-dimensional hidden state vector to the decoder. The decoder constructs a periodic prediction component and a fluctuation prediction component for the high-dimensional latent state vector, respectively derives the accurate prediction value of the periodic component and the interval prediction value of the fluctuation component, and performs weighted fusion. The periodic component prediction branch applies an accuracy loss function to constrain the point prediction accuracy, while the fluctuation component prediction branch uses a charging load spatiotemporal prediction uncertainty quantification loss function to guide and constrain the interval prediction. The overall loss is calculated based on the gradient descent algorithm, and the parameters of the spatiotemporal interval prediction model for charging load are iteratively updated to output the multi-scale spatiotemporal probability interval prediction results of charging load at multiple sites.

8. A high-precision multi-scale spatiotemporal probability prediction system for electric vehicle charging load, characterized in that, include: The charging load model building module is used to establish a charging load model that considers capacity constraints and Markov processes of user behavior, and uses queuing theory to realize the transformation of electric vehicle traffic flow into charging facility load. The module for constructing a spatiotemporal characteristic analysis model of charging load is used to establish a spatiotemporal characteristic analysis model of charging load based on self-attention mechanism and gated recurrent neural network, extract the spatial characteristics of multi-source fused data, and automatically generate the topology of the spatiotemporal map of charging load. The spatiotemporal prediction model construction module is used to establish a spatiotemporal prediction model based on frequency domain graph convolutional neural network and Fourier transform, to perform spatiotemporal prediction of charging load at multiple sites, and introduces the maximum correlation entropy criterion to improve the robustness of the model. The Transformer model building module is used to build a Transformer model based on spatiotemporal scales to measure the contribution of multi-source features to the charging load label. The interval prediction modeling module is used to quantify the uncertainty of charging load spatiotemporal data from micro and macro dimensions using interval prediction evaluation indicators, and to establish an uncertainty quantification loss function to perform interval prediction modeling for the fluctuation composition of charging load. The module for constructing a spatiotemporal prediction model for charging load is used to establish a spatiotemporal prediction model for charging load based on a sequence-to-sequence convolutional long short-term memory neural network. It learns the model from the spatiotemporal characteristics of traffic flow and the decomposed data, and iterates and updates the model using a loss function to obtain multi-scale spatiotemporal probability interval prediction results for charging load at multiple sites.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the high-precision multi-scale spatiotemporal probability prediction method for electric vehicle charging load as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, When the instruction is executed by the processor, it implements the steps of the high-precision multi-scale spatiotemporal probability prediction method for electric vehicle charging load as described in any one of claims 1 to 7.

Images