Long-term classification fast charge load spatial and temporal distribution prediction method and system

Abstract

The invention discloses a long-term classification fast charge load spatial and temporal distribution prediction method and system. The method comprises the following steps: firstly, performing multi-type clustering on charging load data based on a vehicle type, a space type and a date type to generate a load curve of a multi-cluster reference year; then establishing an inventory prediction model to predict the inventory of a target year, correcting the load curve of the reference year according to the inventory of the target year, and obtaining a classification type charging load curve of the target year; performing clustering analysis on the load curve of the target year according to spatial-temporal characteristics, and identifying and extracting a fast charging load curve of the target year with fast charging characteristics; and finally, considering spatial-temporal heterogeneity influence factors of the target year, predicting a correction coefficient of the target year, correcting the fast charging load curve of the target year, and obtaining a fast charging load prediction result of the target year. According to the method provided by the invention, rapid charging load accurate modeling and dynamic correction for different space-time scenes and vehicle types can be realized, and the accuracy, adaptability and application reliability of medium and long term load prediction are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of urban power grid load forecasting, and specifically relates to a method and system for forecasting the spatiotemporal distribution of long-term type-based fast charging loads. Background Art

[0002] Predicting the spatiotemporal distribution of fast-charging loads in urban power grids is a complex process characterized by significant regional heterogeneity and time-varying characteristics. Its essence reflects the evolution of electric vehicle users' charging behavior across different spatial regions and time scales. In recent years, with the rapid growth of electric vehicle ownership and the increasing proportion of fast-charging piles installed, the proportion of fast-charging loads in the overall charging load has continued to rise, becoming a significant factor influencing the operational stability and load fluctuation characteristics of urban power distribution systems. However, systematic research on fast-charging loads remains relatively weak, particularly in the medium- and long-term forecasting realm, where relevant model methods and data support are still incomplete.

[0003] The patent application with publication number CN110751314A uses a harmony algorithm to optimize the random forest parameters of various types of vehicles to reduce algorithm errors. Taking into account the differences in driving habits of individual car owners, the characteristic data of a single electric vehicle is used to train the random forest, and a single-vehicle prediction model for all vehicles is established in parallel. The charging load and its spatiotemporal distribution of the electric vehicle cluster in the area are then obtained through parallel calculation. The characteristics of the driving behavior of individual car owners are fully considered, and the spatiotemporal distribution of the charging load is predicted at the same time, making the prediction results more accurate. Patent application publication number CN115907122A clusters the load variation characteristics of electric vehicle charging stations within a region and, in combination with the Pearson correlation coefficient method, constructs a highly correlated static feature map of charging stations. A CNN-IPSO-BiGRU network prediction model is then established, integrating the static feature map with extended meteorological and holiday information into the network prediction model in a time series format to obtain optimized model parameters. Regional charging station location information and historical load data are then input into the network prediction model, including the optimized model parameters, to produce a predicted time-varying curve of charging station load. This approach exploits the spatiotemporal correlations in data, improving the accuracy of load prediction for all charging stations within a region. Patent application publication number CN108830414A proposes a road traffic model based on road condition information, taking into account network topology and travel time. It utilizes an extreme learning machine to intelligently predict the required travel time for each road section and constructs a model for the proportion of service vehicles to total vehicles. Ultimately, this model predicts the load at charging points, thus overcoming the difficulty of predicting electric vehicle load in commercial charging areas, given the diverse vehicle types and complex spatiotemporal distribution.

[0004] These methods often focus on modeling and forecasting one dimension or a combination of these two: date type (such as weekdays and weekends), vehicle type (such as private cars, taxis, buses), or spatial area division (such as residential areas, commercial areas, and public areas). They lack a research path for systematic integration and dynamic coupling modeling of the multidimensional interactions of these three key factors, making it difficult to fully reflect the spatiotemporal coupling characteristics of fast-charging loads in different area types, time periods, and vehicle types. Therefore, it is urgent to develop a medium- and long-term urban power grid type-specific spatiotemporal distribution forecasting method that integrates spatial structure, temporal behavior characteristics, and vehicle type differences. Summary of the Invention

[0005] In order to address the deficiencies in the prior art, the present invention provides a long-term classification-based fast-charging load spatiotemporal distribution prediction method and system. First, the charging load data is clustered into multiple types based on vehicle type, space type, and date type to generate multiple clusters of base year load curves; then, a vehicle ownership prediction model is established to predict the target year ownership, and the base year load curve is corrected according to the target year ownership to obtain the target year classification-based charging load curve; the target year load curve is clustered according to spatiotemporal characteristics to identify and extract the target year fast-charging load curve with fast-charging characteristics; finally, the target year's spatiotemporal heterogeneity influencing factors are considered to predict the target year's correction coefficient, and the target year fast-charging load curve is corrected to obtain the target year fast-charging load prediction result.

[0006] The first aspect of the present invention provides a method for predicting the spatiotemporal distribution of long-term fast charging loads by type, which adopts the following technical solutions:

[0007] Based on vehicle type, space type, and date type, the charging load data is clustered into multiple types to generate a typical charging load curve for multiple base years.

[0008] Establishing a vehicle ownership prediction model to predict the target year vehicle ownership, and revising the typical charging load curve of each type in the base year according to the target year vehicle ownership to obtain the target year vehicle ownership curve of each type;

[0009] Performing cluster analysis on the target year type-based charging load curves based on spatiotemporal characteristics to identify and extract the target year fast charging load curve;

[0010] The spatiotemporal heterogeneity influencing factors of the target year are obtained, and a correction coefficient of the target year is predicted based on the influencing factors. The fast charging load curve of the target year is corrected to obtain a fast charging load prediction result of the target year.

[0011] Furthermore, the process of constructing the typical charging load curve for the base year includes:

[0012] The original charging load data is divided into multiple typical clusters and reorganized in combination with time nodes to form the charging load data X∈R of a four-dimensional tensor structure. C×D×S×T ; Where C is the vehicle type, D is the date type, S is the space type, T is the time node, T∈{1,2,...,24};

[0013] Let X∈R C×D×S×T Normalized to the interval [0,1];

[0014] Perform one-hot encoding on vehicle type, date type and space type to generate cluster labels as conditional feature vectors C∈R C×D×S ; Where C×D×S is the number of clusters;

[0015] C∈R C×D×S and the normalized X∈R C×D×S×T After splicing, the data are input into a conditional variational autoencoder for feature learning, and a typical charging load curve of the base year by type is generated for each cluster.

[0016] Furthermore, the conditional variational autoencoder adopts an encoder-decoder architecture, including:

[0017] The encoder receives C∈R C×D×S and the normalized X∈R C×D×S×T , from X∈R C×D×S×T Extract the data mean vector μ and logarithmic standard deviation vector σ of each cluster, and obtain the latent variable z through random sampling;

[0018] The decoder is composed of latent variables z and conditional feature vector C∈R C×D×S , reconstruct the sequence of 24-hour charging load values, and output the charging load feature reconstruction data X′∈R C×D×S×T , generate the typical charging load curve of the base year by type.

[0019] Furthermore, the forecast of the target annual holdings includes:

[0020] Obtain vehicle ownership data for each vehicle type over multiple historical years, as well as the corresponding population, GDP per capita, and number of charging stations. Standardize the population, GDP per capita, and number of charging stations to use as standardized variables.

[0021] A generalized logistic growth model with a dynamic shape factor was used to model and fit the vehicle ownership of each type and the total vehicle ownership, and the target annual vehicle ownership forecast results for each vehicle type were obtained.

[0022] The dynamic shape factor takes the standardized variable as input and optimizes the growth model parameters by using the Levenberg-Marquardt algorithm in the nonlinear least squares method.

[0023] Furthermore, the typical charging load curve of each type in the base year is revised, including:

[0024] According to the linear relationship between the total charging load and the number of vehicles in stock, the growth rate of charging load is replaced by the growth rate of vehicle inventory, which is defined as:

[0025]

[0026] Among them, α i,j,k is the charging load growth rate of vehicle type i under space type j and date type k; β i is the growth rate of electric vehicle ownership of vehicle type i; and are the target annual population and base annual population of vehicle type i, respectively;

[0027] The charging load growth rate is multiplied by the load data at 24 time points in the typical charging load curve of the base year to obtain the load data of the charging load curve of the target year, thereby obtaining the charging load curve of the target year.

[0028] Furthermore, the extraction of the target annual fast charging load curve includes:

[0029] Calculate the morphological similarity between charging load curves of different target year types, and construct a similarity matrix S based on the similarity;

[0030] An undirected weighted graph G is constructed using the similarity matrix S. In the undirected weighted graph G, each node represents a target year type charging load curve, and the edge weight between nodes is the morphological similarity between the two target year type charging load curves;

[0031] Spectral clustering is performed based on the undirected weighted graph G, and spectral gap analysis is used to determine the optimal number of clusters c;

[0032] Statistical characteristics are calculated for each of the c clusters. Load curves with fast charging characteristics are identified based on the statistical characteristics to obtain the target year fast charging load curve.

[0033] Furthermore, the target annual fast charging load curve is corrected, including:

[0034] The XGBoost model is used to train the target year load correction model. The model output is the 24-hour correction coefficient of the target year, and the correction coefficient curve of the target year typical load curve is obtained;

[0035] The input of the target year load correction model consists of two parts: the first part is a 6-dimensional discrete variable generated by one-hot encoding after combining space type × date type; the second part is a 4-dimensional continuous variable composed of the spatiotemporal heterogeneity influencing factors of the target year;

[0036] Multiply the normalized ratio of the correction coefficient at each time point by the load data at the corresponding time point in the target year fast charging load curve to obtain the target year fast charging load forecast result.

[0037] Furthermore, the influencing factors of spatiotemporal heterogeneity in the target year include the proportion of fast charging facilities, charging pile density, road network density and urbanization rate in the target year, and the prediction methods are as follows:

[0038] Collect data on the number of fast-charging piles and total charging piles in the three years before the base year, calculate the average annual growth rate of the number of fast-charging piles, and use linear extrapolation based on the fast-charging percentage in the base year to obtain the percentage of fast-charging facilities in the target year;

[0039] Extract the charging pile density growth data for the three years before the base year, calculate the average annual growth rate of charging pile density, and multiply the average annual growth rate by the charging pile density in the base year to obtain the charging station density in the target year;

[0040] Based on the total length of newly added roads and the regional area in the three years before the base year, the annual increase in road length per unit area is calculated and superimposed on the road network density in the base year to obtain the road network density in the target year;

[0041] The average annual growth rate of the urbanization rate is calculated based on the urbanization rate in the three years before the base year, and the urbanization rate in the target year is calculated based on the said average annual growth rate.

[0042] Further, training the target year load correction model;

[0043] An XGBoost model is trained based on the temporal and spatial heterogeneity influencing factors and load correction coefficients in the base year to capture the mapping relationship between the temporal and spatial heterogeneity influencing factors and the load correction coefficients.

[0044] The series of correction factors for the base year is defined as Among them, λ jt Represents the correction coefficient of space type j at time point t, R base,jt and R CVAE,jt are the original charging load data of the base year, R CVAE,jt Charging load data of typical charging load curves by type in the base year;

[0045] The XGBoost input sample consists of two parts: the first part is a 6-dimensional discrete variable generated by one-hot encoding after combining the spatial type and date type; the second part is a 4-dimensional continuous variable composed of the spatiotemporal heterogeneity influencing factors of the base year;

[0046] During the training process, the multi-output regression model wrapper is used to perform synchronous regression modeling of correction coefficients at multiple time points, and multiple XGBoost regression sub-models are constructed based on the gradient boosting tree.

[0047] The second aspect of the present application provides a long-term type-based fast charging load spatiotemporal distribution prediction system, using the method provided in the second aspect of the present application, the system includes:

[0048] Base year curve generation module: clusters charging load data into multiple types based on vehicle type, space type, and date type, and generates multiple clusters of typical charging load curves for the base year;

[0049] A target year curve generation module; establishing a vehicle ownership prediction model to predict the target year vehicle ownership, and modifying the typical charging load curve of the base year according to the target year vehicle ownership to obtain the target year charging load curve of the target year;

[0050] A fast charging curve generation module is used to perform cluster analysis on the target year type charging load curve according to the spatiotemporal characteristics, and identify and extract the target year fast charging load curve;

[0051] The target year fast charging curve prediction module obtains the temporal and spatial heterogeneity influencing factors of the target year, and predicts the correction coefficient of the target year based on the influencing factors, corrects the target year fast charging load curve, and obtains the target year fast charging load prediction result.

[0052] The beneficial effect of the present invention is that, compared with the prior art,

[0053] 1. This application performs fine-grained clustering of raw charging load data based on a three-dimensional combination of vehicle type, space type, and date type, and introduces a conditional variational autoencoder (CVAE) model to perform feature learning and reconstruction of load curves under multi-dimensional label conditions, generating representative benchmark year typical classification charging load curves. This method not only retains the key structural information of the original load data, but also enhances the model's ability to capture differences in heterogeneous scenarios, thereby improving the expression diversity and generation accuracy of load curves in different traffic, regional, and time scenarios, providing a reliable foundation for subsequent target year load forecasts.

[0054] 2. This application incorporates spatial heterogeneity factors related to urban development into the fast-charging load forecasting process, including the proportion of fast-charging facilities, charging pile density, road network density, and urbanization rate. By combining structured discrete variables of spatial and date types, an XGBoost multi-output regression model is used to establish a mapping relationship between regional development characteristics and fast-charging load morphological deviations. This model can adaptively modify the original forecast results based on the level of infrastructure development in different regions, enhancing the model's responsiveness to charging demand fluctuations under evolving spatiotemporal scenarios.

[0055] 3. This application adopts a generalized logistic model based on dynamic shape factors to fit and predict the growth trend of the ownership of different types of electric vehicles (such as private cars, taxis, and buses), and combines multivariate inputs such as historical population, GDP, and the number of charging piles to achieve high-precision modeling of the target annual ownership, and adjust the total level of charging load accordingly. At the same time, at the morphological level, by constructing a sequence of correction coefficients, the local shape deviation of the generated curve is further fine-grained to achieve dual corrections in quantity and structure. The solution proposed in this application enhances the adaptability and interpretability of fast charging load forecasting in cross-year, multi-regional, and multi-type scenarios, and can provide a scientific basis for medium- and long-term infrastructure construction and distribution system load assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A flowchart for predicting the spatiotemporal distribution of long-term fast charging load by type;

[0057] Figure 2 Typical charging load curve for the base year under vehicle type-space-time characteristics;

[0058] Figure 3 Corrected parameter predictions for representative curves for different space-date combinations;

[0059] Figure 4 The spatiotemporal distribution prediction curve of electric vehicle fast charging load by type in the target year. DETAILED DESCRIPTION

[0060] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. The embodiments described in this application are only part of the embodiments of the present invention, not all of them. Based on the spirit of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0061] Example 1

[0062] This embodiment provides a specific implementation method for predicting the spatiotemporal distribution of long-term fast charging load by type. Figure 1 , the method comprises the following steps:

[0063] Step 1: Based on vehicle type, spatial function, and date type, the charging load data is clustered into multiple types to generate a typical charging load curve for multiple clusters of base years. Specifically:

[0064] 1.1: Considering the differences in charging methods, charging time and power requirements of electric vehicles, vehicle types are divided into electric private cars, electric taxis and electric buses;

[0065] Based on the urban functional zoning and charging demand distribution, the study area is divided into commercial areas, public areas and residential areas, corresponding to the charging behavior characteristics of different types of users;

[0066] Considering the temporal characteristics of charging demand, it is further subdivided into working days and rest days to characterize the impact of different day types on charging load.

[0067] Based on this classification, the raw charging load data was divided into 18 typical clusters (3×3×2) across multiple types, including "electric private cars - commercial areas - weekdays" and "electric taxis - public areas - weekends." This provided high-precision data support for the subsequent typical charging load curve fitting model.

[0068] Supplementary explanation: Based on the actual situation of a certain district, since there are regional differences in the construction and distribution of bus charging stations, only areas where bus charging stations are actually built are considered when classifying the spatial functional areas of bus charging stations.

[0069] 1.2: According to the above classification method of vehicle type, spatial function and time attribute, the 18 types of multi-type typical cluster load data are reorganized according to the four-dimensional labels of vehicle type-date type-space type-time node to form a four-dimensional tensor structure: X∈R C×D×S×T .

[0070] Among them, C represents vehicle type, D represents date type, S represents space type, and T represents time node, where T∈{1,2,...,24} represents time node, totaling four dimensions. The first dimension (C) has three index positions corresponding to three vehicle types: electric private cars, electric buses, and electric taxis; the second dimension (D) has two index positions corresponding to two date types: weekdays and weekends; the third dimension (S) has three index positions corresponding to three spatial functional area types: commercial areas, public areas, and residential areas; and the fourth dimension (T) has 24 index positions corresponding to 1 day and 24 hours.

[0071] In the time series, this embodiment establishes a mapping relationship between the date labels in the original data and the date types through the ISO 8601 date coding standard, and divides Monday to Friday into working days, and Saturday and Sunday into rest days. The spatial dimension adopts a two-layer division method: first, according to the urban main road planning, the charging piles are aggregated into spatial units; then, combined with the POI basic information data downloaded from OpenStreetMap, kernel density calculation is performed to divide the spatial units into three functional areas: commercial areas (including 18 types of POIs such as business offices and shopping centers), residential areas (including 9 types of POIs such as residential communities and community services), and public areas (including 12 types of POIs such as transportation hubs, culture, education, and medical care). Finally, the three types of vehicle type labels and 24 time node labels of the data are combined to reorganize the original data into a four-dimensional tensor structure with a shape of [3,2,3,24]: X∈R C×D×S×T .

[0072] Normalize the reorganized four-dimensional tensor charging load data to the interval [0,1];

[0073] At the same time, the vehicle type, space type and date type are uniquely encoded to generate cluster labels (18 types of cluster labels such as "electric private car-commercial area-weekday") as the conditional feature vector C∈R 18 .

[0074] The normalized four-dimensional tensor and conditional feature vector are concatenated along the feature dimension and input into the Conditional Variational Autoencoder (CVAE).

[0075] 1.3: Feature learning and load curve generation based on CVAE.

[0076] In the charging load curve generation task, the original four-dimensional tensor structure (vehicle type × date type × space type × time node) needs to be converted into a two-dimensional input format suitable for neural network processing. The specific conversion process is as follows:

[0077] First, each combination of (vehicle type, date type, and space type) is treated as an independent sample, each containing 24 hours of load time series data. Simultaneously, a corresponding 18-dimensional one-hot encoded conditional vector is generated for each sample, clearly identifying the cluster category to which it belongs (e.g., "Private Car_Commercial District_Weekday"). The final input data is concatenated along the feature dimension with the normalized 24-hour load data and the 18-dimensional conditional vector, forming a 42-dimensional feature vector. This transformation preserves the key information of the original 4D tensor while meeting the neural network's requirement for 2D input.

[0078] The conditional variational autoencoder (CVAE) uses an encoder-decoder architecture with the following structure:

[0079] The encoder consists of two fully connected layers: the first layer maps the 42-dimensional input (24-dimensional payload data + 18-dimensional conditional vector) to a 64-dimensional feature space, and the second layer further compresses it to a 32-dimensional feature space. Both hidden layers use the ReLU activation function to introduce nonlinear modeling capabilities. The encoder output is divided into two branches: one branch generates the mean vector μ of the latent space through a linear transformation, and the other branch generates the logarithmic variance vector logvar. Random sampling is performed from the latent space by reparameterizing z = μσ⊙ε, where ε follows a standard normal distribution, i.e., ε ~ N(0,1).

[0080] The decoder concatenates the sampled latent variables and the conditional vector as input. First, a fully connected layer maps the (latent dimension + conditional dimension) to a 32-dimensional space, which is then gradually expanded to 64 dimensions, ultimately outputting a 24-dimensional reconstructed payload. All layers except the last use the ReLU activation function. The output layer uses a Sigmoid activation function to ensure that the generated values ​​are in the range [0, 1], matching the normalized input data.

[0081] The processing process of the conditional variational autoencoder (CVAE) is as follows: the encoder receives 18 multi-type typical clusters X and conditional feature vectors C processed by the clustering model, and extracts the data mean vector μ and logarithmic standard deviation vector σ of each cluster;

[0082] The latent variable z is obtained by random sampling;

[0083] The decoder receives the latent variable z and the conditional feature vector C and reconstructs the sequence of 24-hour charging load values;

[0084] The output is the charging load feature reconstruction data X′, which realizes the feature restoration of the charging behavior under a specific cluster. Its dimension is strictly consistent with the input time series data, X′∈R C×D×S×T .

[0085] The conditional variational autoencoder (CVAE) generates 18 three-dimensional typical charging load curves of the base year by processing the reconstructed data X′. Each curve is a 24-point sequence, such as Figure 2 Each curve corresponds to a different charging load mode, covering the charging load values ​​at 24 time nodes, and accurately reflects the load fluctuation characteristics in different time periods.

[0086] This embodiment applies the deep generative model to the field of spatiotemporal prediction of charging load, uses a conditional variational autoencoder to capture the annual charging load behavior characteristics, and uses a decoder to reconstruct and generate typical daily data of the base year charging load as the data basis for prediction correction.

[0087] During model training, each sample is precisely assigned to a specific cluster category. An 18-dimensional one-hot encoded conditional vector clearly identifies the vehicle type, spatial function, and date combination to which the sample belongs, such as "Taxi_Public Area_Holiday." This design enables the model to learn the load distribution characteristics under different combinations. During data loading, each training batch contains multiple samples, each containing a complete 24-hour load curve and its corresponding conditional vector. It should be noted that due to the unique regional distribution of bus charging stations, invalid spatial types were filtered out during data preprocessing to ensure that each sample represents a valid charging behavior combination.

[0088] During training, the encoder learns to map load curves under specific conditions into a low-dimensional latent space, capturing both common features and differential patterns across cluster types. The decoder then learns to reconstruct load curves that meet specific conditions from the latent space. The loss function consists of two parts: the reconstruction loss uses mean squared error to measure the difference between the generated curve and the true curve; and the KL divergence constrains the latent space distribution to be close to a standard normal distribution, ensuring continuity and interpretability. By jointly optimizing these two objectives, the model learns a deep feature representation of charging loads and is able to generate new load curves that conform to specific vehicle types, spatial functions, and date types, providing data support for charging infrastructure planning.

[0089] Step 2: Based on the number of vehicles in use, perform numerical correction on the typical charging load curve by type to obtain the charging load curve by type for the target year. Specifically:

[0090] The magnitude of EV charging load is directly affected by the number of EVs in use, which is heterogeneous by model and varies in temporal and spatial distribution. Therefore, this step considers the growth of each model's number of EVs separately.

[0091] 2.1: Through the generalized logistic model, a dynamic shape factor function is constructed, integrating multi-dimensional variables such as regional population base, per capita GDP, and regional car-to-charging ratio, to establish a type-based electric vehicle ownership prediction model to obtain the ownership of electric private cars, electric taxis, and electric buses in the target year. Specifically:

[0092] The number of electric vehicles in use is predicted based on the logistic growth model. The steps are as follows:

[0093] Based on public statistical documents, historical data on electric private cars, electric taxis, electric buses, total vehicle ownership, and the corresponding years' population, per capita GDP, and total number of charging piles were collected as model inputs. Z-score preprocessing was performed on the population, per capita GDP, and total number of charging piles to eliminate data dimensional differences.

[0094] A generalized logistic model with dynamic shape factors was introduced, and curve fit was used to fit the parameters of a custom function. Separate models were created for electric private cars, electric taxis, electric buses, and the total number of electric vehicles on the road. The LM algorithm, a nonlinear least squares method, was used to optimize the parameters in each of the four models.

[0095] The coefficient of determination (R 2 ) and root mean square error (RMSE) are used as error evaluation indicators of model fitting, and the prediction errors of the four models of electric private cars, electric taxis, electric buses, and total vehicle types are calculated respectively.

[0096] The original data statistics are used to obtain the number of electric private cars, electric taxis and electric buses in the base year, which is recorded as

[0097] From a micro-behavioral mechanism analysis, individual users' charging decisions are influenced by multiple factors, including daily mileage, starting charging time, charging duration, starting SOC, and travel chain spatiotemporal constraints. There is significant variability and randomness in the choices made regarding charging time, location, and duration. However, at a macro-statistical level, based on the law of large numbers, the spatiotemporal charging patterns of vehicle groupings exhibit significant regularity—the distribution of charging demand periods for each vehicle type within a specific functional area can be viewed as a stable probability density function. This characteristic ensures that, despite random perturbations in individual charging behavior during vehicle ownership growth, the total charging demand in the vehicle type dimension maintains an approximately linear proportional relationship with vehicle ownership growth. Therefore, it can be considered that the total charging load is approximately linearly related to vehicle ownership.

[0098] That is, the load growth calibration model for each vehicle type is defined as follows:

[0099]

[0100] Where, α i,j,k is the charging load growth rate of vehicle type i under space type j and date type k; β i is the growth rate of electric vehicle ownership with vehicle type i; i, j, k represent the vehicle type, spatial area, and day type numbers, respectively, where i = 1, 2, 3, representing private cars, taxis, and buses, respectively; j = 1, 2, 3, representing commercial areas, public areas, and residential areas, respectively; k = 1, 2, representing weekdays and weekends, respectively.

[0101] 2.2: Calculate the charging load growth rate of different vehicle types through the load growth calibration model;

[0102] Corresponding to the charging load growth rate, the charging load growth rate is multiplied by the load values ​​at 24 time points in the typical charging load curve of the base year obtained in step 1 to obtain the load value of the charging load curve of the target year. The process is expressed as:

[0103]

[0104] in, and P ijk (t) are the load values ​​of the charging load curves of vehicle type i in spatial region j and day type k at time t, respectively, in the target year and the base year.

[0105] Ultimately, this process yielded a typical load curve for electric private vehicles, commercial areas, and weekdays in the target year. This curve reflects the changing charging load trends in commercial areas during weekdays, given the expected growth in electric private vehicle ownership in the target year. By applying corrections based on the corresponding growth rates of each type of electric vehicle, we were able to accurately derive typical load curves for different vehicle types in specific areas and time periods in the target year. By quantifying the spatiotemporal evolution of ownership growth, this method optimized the load forecasting model and enhanced the accuracy and applicability of the target year load curves.

[0106] It should be noted that although this step employs a linear approximation between charging load and vehicle population growth in model design, this linear growth only applies to the base-year load curve for a specific vehicle model under given regional and temporal conditions, providing a numerical amplification and not altering the original spatiotemporal distribution of the charging load. In other words, the vehicle population growth ratio serves only as a scaling factor for the curve amplitude. However, the charging behavior of different vehicle models in different spatial regions and on different daytime types exhibits significant heterogeneity, resulting in distinct load distribution patterns and peak periods across a 24-hour period. Therefore, the assumption of linear growth in total population cannot be used to disregard temporal and spatial dimensionality. Therefore, during the numerical correction process in this step, the vehicle population ratio must be calibrated for each representative load curve based on the three-dimensional classification structure of "vehicle model, spatial functional area, and daytime type." This ensures that the model reflects the overall growth trend while preserving the original temporal and spatial variability of charging behavior, thereby enhancing the precision and real-world adaptability of the target year forecast results.

[0107] Step 3: Based on the target year-type charging load curve after the inventory correction, identify and extract the charging load curve with typical fast charging characteristics.

[0108] Fast charging loads have significant characteristics of high peak values, time concentration, and high volatility. Since fast charging stations provide high-power charging in a short period of time, the load rises and falls rapidly within a specific period, especially during peak traffic hours or holidays when charging demand is concentrated, showing strong load volatility. In addition, fast charging loads are greatly affected by external factors such as traffic flow and climatic conditions, which may cause sudden increases or fluctuations in charging demand. Based on this characteristic, a refined load model based on multi-scale restricted aligned path length spectrum clustering is constructed to extract the fast charging load curve from the target year-type charging load curve after the inventory is corrected. The specific process is as follows:

[0109] 3.1: Use LDTW (Length-constrained Dynamic Time Warping) algorithm to calculate the similarity between different load curves. Given two load curves, use LDTW to measure the morphological similarity on the time axis, calculate the LDTW distance, and construct the similarity matrix S. For two load curves X1 = {x 11 ,x 12 ,…,x 1p},X2={x 21 ,x 22 ,…,x 2q}, its LDTW distance LDTW(X1,X2) is defined as:

[0110]

[0111] Where Γ is the set of all legal alignment paths, ω(a, b) is the weight function that controls the penalty weight of the alignment point, and ||x 1a -x 2b || represents the Euclidean distance between the load data at the a-th time point in X1 and the load data at the b-th time point in X2, a∈[1,p], b∈[1,q].

[0112] Finally, a symmetric similarity matrix S∈R is obtained based on the LDTW distance calculation M×N , where each element S mn Represents the similarity between load curve m and load curve n. The higher the similarity value, the more similar the curves are. mn =exp(-γ·LDTW(X m ,X n )),γ is the similarity scaling factor, and M is the number of load curves.

[0113] 3.2: Using the similarity matrix S, construct an undirected weighted graph G, where each node represents a load curve and the weights of the edges between nodes are determined by the values ​​in the similarity matrix. Use the k-nearest neighbor method to determine the connectivity in the similarity graph, ensuring that each node is connected only to nodes with high similarity, preserving the strong structural relationship between fast charging loads.

[0114] 3.3: Calculate the degree matrix D of the graph, where D mm Represents the degree of node m, that is, the sum of the similarities between node m and other nodes. Construct the normalized Laplace matrix based on S and D It is used to capture the connection density between different charging load curves, suppress the influence of isolated edges on the spectral structure, and highlight the consistency structure within the load curve cluster.

[0115] 3.4: Normalized Laplace matrix L s Perform eigenvalue decomposition to obtain the eigenvectors corresponding to the first g smallest eigenvalues. These first g eigenvectors are then concatenated to form the characteristic matrix U. Each eigenvector represents the node distribution in a low-dimensional space. This spectral space mapping preserves the morphological similarity between curves. That is, in spectral space, charging curves with similar power structures and different start and end times but consistent morphology are mapped to adjacent locations.

[0116] 3.5: In the spectral space, perform K-means clustering on the feature matrix U to obtain the cluster label of each load curve. Through clustering, load curves with similar spatiotemporal characteristics are assigned to the same cluster.

[0117] 3.6: By analyzing the spectral gap, determine the appropriate number of clusters c to ensure fine classification of the fast charging load curve. Assume that the eigenvalue of the Laplace matrix is ​​0 = λ1≤λ2≤…≤λ M , the number of clusters c should be selected to satisfy That is, the point with the largest difference between adjacent eigenvalues ​​is selected as the optimal number of clusters to ensure high consistency within the class and large difference between classes.

[0118] For the load curves in the c clusters, the statistical characteristics of each cluster are calculated, such as peak power, duration, fluctuation coefficient, etc. In particular, for each load curve, when its peak power exceeds 50kW and the duration is less than 10 minutes, and it appears 3 times or more in a day, these load curves can be judged as having typical fast charging load characteristics. Furthermore, through a comprehensive evaluation of these characteristics, curve clusters that meet typical fast charging characteristics such as "short-term high power", "high-frequency spikes" and "severe fluctuations" can be effectively identified.

[0119] Finally, by extracting the load curves in these clusters, the target year-type fast charging load curve can be obtained.

[0120] 3.7: Use indicators such as silhouette coefficient and adjusted Rand index to evaluate the quality of clustering and ensure the reliability of clustering results.

[0121] Step 4: Consider the spatiotemporal heterogeneity caused by urban development and modify the fast charging load curve by type.

[0122] 4.1: To accurately characterize the factors affecting charging load in different areas of the city under different date types, a 10-dimensional input feature vector containing spatial, temporal and infrastructure indicators is constructed. This vector consists of two parts:

[0123] (1) 6-dimensional discrete variables: spatial type (commercial area / residential area / public area) × date type (weekday / holiday); for each combined sample, one-hot encoding is used to generate a 6-dimensional sparse vector;

[0124] (2) 4-dimensional continuous variable: the proportion of fast charging facilities in the base year P base , charging pile density D base , road network density L base and regional urbanization rate U base ;

[0125] The 6-dimensional discrete variables and the 4-dimensional continuous variables are concatenated to form a 10-dimensional input feature to capture the interaction effects in different spatiotemporal scenarios and establish a multi-dimensional data feature vector input, which is recorded as:

[0126] X j,k =[X region-day ,P base ,D base ,L base ,R base ];

[0127] Where, X j,k is the characteristic vector of each set of data, where j and k represent the numbers of spatial type and date type respectively; X region-day It is the one-hot encoding of the combined features of the spatial type and date type of the data;

[0128] 4.2: Model training; using XGBoost modeling, regression input is X j,k , the output is the correction coefficient sequence corresponding to 24 hours.

[0129] The output of the XGBoost model is set as a 24-point curve correction coefficient, which represents the quantitative index of the morphological deviation between the original base year true load value and the CVAE generated curve in step 1 at 24 time points.

[0130] The correction coefficient is defined as the ratio of the 24-point actual load value of the original data to the 24-point charging load value in the typical charging load curve of the base year generated in step 1:

[0131]

[0132] Where λ jt is the correction factor; R base,jt is the original base year real 24-hour load data of different types of spatial areas; R CVAE,jt The 24-hour load data for a typical day curve in the base year generated by CVAE.

[0133] To achieve multi-dimensional output, the output layer uses the MultiOutputRegressor framework to synchronously generate correction factor prediction values ​​at 24 time points and construct a multi-dimensional regression mapping relationship.

[0134] Using the base year's influencing factor eigenvector X j,k and load correction factor λ jt To train the XGBoost model, we can obtain the mapping relationship between the influencing factors and the load correction coefficients.

[0135] In the load forecasting model, the XGBoost training process begins with the data preparation stage. By creating a "space-date" combined feature (space type_date type), the charging pile density, road network density, proportion of fast charging facilities and urbanization rate in the base year are used as input features, and the 24-hour load correction coefficient is used as the output value.

[0136] During data preprocessing, we used ColumnTransformer to perform one-hot encoding on categorical features and normalized numerical features to ensure comparability of features of different dimensions. The dataset was then split into training and test sets in an 8:2 ratio, maintaining a random seed of 42 to ensure experimental reproducibility.

[0137] The model training phase uses the MultiOutputRegressor wrapper to train an independent XGBoost regressor for each hour of the 24-hour workload output. Each XGBoost model sequentially builds 1,000 decision trees based on the gradient boosting framework, iteratively optimizing the predictions: each new tree learns to correct the residuals of the previous tree.

[0138] During the training process, the mean square error (MSE) is used as the loss function by default. The objective function is minimized through the gradient descent algorithm. At the same time, a regularization term is introduced to control the complexity of the model and prevent overfitting.

[0139] In the core parameter configuration of this embodiment, n_estimators = 1000 determines the number of basic learners, and learning_rate = 0.1 controls the contribution weight of each tree. The two together affect the convergence speed and accuracy of the model. The default value of the max_depth parameter is 6, which limits the maximum depth of a single tree; the default value of min_child_weight is 1, which constrains the minimum sample weight required for leaf nodes; the default values ​​of subsample and colsample_bytree parameters are both 1, which control the sample and feature sampling ratios respectively. random_state = 42 ensures the reproducibility of the experiment, and objective = 'reg:squarederror' explicitly defines the loss function of the regression problem. This strategy of multiple independent output models can ensure that the hourly load forecasts do not interfere with each other.

[0140] 4.3: By predicting the proportion of urban fast charging facilities, charging station density, road network density and urbanization rate in the target year, the characteristic vector X of each influencing factor in the target year is obtained obj,j,k , and input it into the trained XGBoost model to predict the 24-hour correction coefficient of the target year, and obtain the correction coefficient curve of the typical load curve of the target year.

[0141] Furthermore, the forecast method for the target year's fast charging facility ratio, charging pile density, road network density, and urbanization rate is as follows:

[0142] (1) For the proportion of fast charging facilities, the average annual growth rate is calculated by statistically analyzing the number of fast charging piles and total piles in the past three years. The target year fast charging facility proportion is obtained by linear extrapolation based on the current year fast charging proportion. The specific calculation formula is as follows:

[0143] P obj =P base +g p ΔT;

[0144]

[0145] Where: P obj is the target year’s proportion of fast charging facilities; P base is the proportion of fast charging facilities in the base year (i.e. the number of fast charging piles / total number of piles); g p is the average annual growth rate in the past ΔT years; ΔT is the interval between the target year and the base year (unit: year).

[0146] (2) For the charging station density, that is, the number of charging stations per unit area, a multiplication recursion model is used for prediction. By extracting historical charging station density growth data (about 3 years), its average annual growth rate is estimated, and based on this, the current density is multiplied to obtain the charging station density data for the target year. The specific calculation formula is as follows:

[0147] D obj =D base ·(1+g d ) ΔT ;

[0148]

[0149] Where: D base is the density of charging stations per unit area in the base year (stations / square kilometer); D obj is the target annual charging station density; g d It is the historical average annual increase (growth rate).

[0150] (3) As a representative indicator of the level of urban road construction, road network density uses absolute increment linear prediction because its annual increment is relatively stable and its growth trend is relatively slow. The specific method is to combine the total length of newly added roads in the past three years with the regional area, estimate the annual increment of road length per unit area, and superimpose it on the current value to obtain the target year prediction result. The specific calculation formula is as follows:

[0151] L obj =L base +ΔL·ΔT;

[0152]

[0153] Where: L base The road network density per unit area in the base year (km / km 2 );L obj is the target annual road network density; ΔR is the annual road growth per unit area (km / km 2 / year); base is the total road length in the base year (km); A is the area (km 2 ).

[0154] (4) The urbanization rate is predicted based on the growth rate of the city's permanent population or the expansion rate of the built-up area. Considering that my country's urbanization process is gradually entering a stable development stage, a linear extrapolation method can be used to estimate the urbanization level of the region in the target year based on the average growth value of the urbanization rate in the past three years. The specific calculation formula is as follows:

[0155] U obj =U base +g u ΔT;

[0156]

[0157] Where: U base is the urbanization rate in the base year (%); U obj is the target annual urbanization rate; gu is the average annual growth rate of urbanization, U base-r represents the urbanization rate r years before the base year.

[0158] 4.4: Then multiply the normalized ratio of the correction coefficient at each time point by the load value at the corresponding time point in the fast charging load curve to obtain the new curve after shape correction.

[0159] To address the spatiotemporal heterogeneity of charging behavior between the base year and the target year due to urban development and infrastructure changes, this step proposes a modified model for the typical daily spatial behavior of charging load. This model modifies the shape of the typical fast-charging load curves for each type in the target year to achieve more accurate spatiotemporal predictions.

[0160] The framework focuses on predicting regional spatial development differences, adjusting the spatiotemporal morphology of electric vehicle fast-charging load curves by type in target years, establishing a mapping between charging facility supply capacity and load demand response. Using a multi-output XGBoost model, a curve morphology correction model based on the synergistic interaction of multiple factors was constructed across spatial functional zoning and temporal attribute dimensions. This model fully considers the multiple factors that influence the spatiotemporal evolution of charging load, including key factors such as the differentiated layout of urban fast-charging facilities, charging pile density, road network density, and urbanization rate.

[0161] Among them, the proportion of fast-charging facilities in a city reflects the regulatory influence of charging technology preferences on peak and valley load characteristics; charging pile density reflects the impact of charging service accessibility on user behavior patterns; road network density constrains the spatiotemporal distribution of charging demand through spatial accessibility; and urbanization rate reflects the regional scale of electric vehicles. The synergistic effect of these four factors constitutes the spatial driving mechanism for charging load evolution, dynamically correcting the spatiotemporal migration of charging hotspots caused by infrastructure upgrades and urban development.

[0162] To address the objective constraint of missing dynamic monitoring data on charging behavior in historical years, this application proposes a load correction method based on the assumption of spatial homogeneity. By dividing the study area into functional units with homogenized charging characteristics according to spatial functional zoning, the method attributes the deviation of each unit's charging load from the baseline curve to the combined effects of the fast-charging pile ratio, the scale of charging facilities, and the road network density. This method is used to capture the mapping relationship between the spatial development differences in the forecast area and the load demand response.

[0163] This embodiment designs two comparative verification schemes for the prediction method of the spatiotemporal distribution of electric vehicle fast charging load, in order to systematically evaluate the impact of different correction strategies on the model prediction accuracy. Scheme 1 is a spatiotemporal distribution prediction model for fast charging load of electric vehicles of different types, which only considers the numerical correction of the ownership growth ratio; Scheme 2 is a spatiotemporal distribution prediction model for fast charging load of electric vehicles of different types, which considers the numerical correction of the ownership growth ratio and the dynamic correction of the spatial region. In view of the unavailability of the benchmark value of charging load in the target year, this patent selects the 2023-2024 charging facility operation monitoring database of Fengrun District, Tangshan City as the verification benchmark. During the verification process, the prediction results of the two schemes are compared with the measured data in multiple dimensions, and quantitative analysis is carried out on the three key indicators of load peak time deviation, spatial distribution consistency and daily load curve morphological characteristics. The improvement of the three key indicators obtained by simulation of the above two different schemes is shown in Table 1.

[0164] Table 1 Comparative analysis of different fast charging load spatiotemporal distribution prediction methods

[0165]

[0166] The research results show that: in the temporal dimension, the charging demand timing offset caused by the commuting tidal effect is effectively captured, and the average absolute deviation of the peak time is reduced from 28.0 minutes to 18.0 minutes (a decrease of 35.7%). In the spatial dimension, by constructing a spatial response function that includes four dynamic correction factors, including the layout of fast charging facilities, the differentiation of road network density, and the urban urbanization rate, the model's ability to characterize the heterogeneity of the spatial distribution of load is significantly improved, and the spatial correlation coefficient is increased from 0.72 to 0.89, verifying the model's ability to analyze the gradient characteristics of charging demand in urban functional areas. In terms of curve morphological fidelity, morphological similarity analysis based on dynamic time warping shows that Scheme 2 reduces the DTW distance value from 152.3 to 98.5 (a decrease of 35.3%), confirming that the model has a stronger ability to characterize the electric vehicle charging load forecast curve.

[0167] This paper uses historical data on charging loads for various types of electric vehicles in Fengrun District from 2023 to 2024, as well as historical data on regional spatial factors such as charging station density, the proportion of fast-charging stations, and urbanization rate, to predict the spatiotemporal distribution of charging loads by type in Fengrun District in 2025. First, a multi-type clustering CVAE (Clustered Automatic Emission Analyzer) was used to generate the charging load curve for the base year. Specific CVAE parameters are shown in Table 2.

[0168] Table 2 Conditional variational autoencoder parameter settings

[0169]

[0170] The CVAE fitting results are as follows Figure 3As shown in the figure, bus charging is mainly concentrated between 12:00 and 6:00 at night and around 12:00 noon, showing typical morning and evening peak charging characteristics, reflecting the behavior of concentrated charging during the idle period of operation. Among them, the load in public areas is the most significant, indicating that buses mainly charge intermittently at night or in the afternoon at transportation hubs and dedicated public charging stations;

[0171] In commercial and residential areas, bus charging loads are generally low, indicating limited charging demand in areas where operations are not concentrated. Taxi charging loads exhibit a multimodal distribution, with peaks occurring around 9:00 AM, 2:00 PM, and 8:00 PM. This distribution is particularly pronounced on weekdays, reflecting a combination of high-frequency taxi operations and flexible charging. Their load is highly concentrated in public and commercial areas, indicating that operating vehicles are primarily active in the city's core. Private car charging loads are primarily concentrated between 8:00 PM and 12:00 AM, demonstrating a typical "nighttime concentrated charging" pattern, particularly in residential areas. This pattern is consistent across both weekdays and weekends. Private cars also exhibit some daytime charging activity on weekends, particularly in public and commercial areas, indicating that increased travel frequency during holidays leads to more dispersed charging times. Overall, the CVAE model's fitting results for fast-charging loads for various types of vehicles demonstrate not only excellent temporal distribution patterns but also strong adaptability to spatial behavior, effectively reflecting the diversity and regional characteristics of actual charging behavior.

[0172] Based on the CVAE model's prediction results and data on the number of electric vehicles in use, we calculated the growth rate of electric vehicle charging load by type. Based on this growth rate, we normalized it and constructed corresponding numerical correction coefficients to adjust the typical charging load curve for the base year to reflect the dynamic changes in the load structure in the target year as the number of electric vehicles increases. The correction coefficients are detailed in Table 3.

[0173] Table 3 Growth rate of electric vehicle charging load by type

[0174]

[0175] According to the characteristics of fast charging, high power and short charging time, a refined load model based on multi-scale restricted aligned path length (LDTW) spectral clustering is constructed, and the fast charging load part is extracted from the typical load curve of the classification after the ownership ratio is corrected. Then, a multi-output XGBoost model is used to construct a curve morphology correction model based on the synergy of multiple factors in the spatial functional zoning and time attribute dimensions. Through the analysis of various influencing factors, the proportion of urban fast charging facilities, charging station density, road network density and urbanization rate in different functional areas in the base year are represented in the form of feature vectors. The specific values ​​are shown in Table 4, respectively denoted as X 1j,X 2j ,X 3j ,X 4j ,j=1,2,3,establish multi-dimensional data feature vector input.The normalized correction factor curve of the typical load curve of the target year is obtained through the multi-output XGBoost model, such as Figure 2 As shown, the normalized ratio of the correction factor is multiplied by the load value of each time point of the fast charging load curve to obtain the new curve after shape correction, as shown in Figure 4 shown.

[0176] Table 4 Data of various influencing factors in different spatial regions

[0177]

[0178] Example 2

[0179] This embodiment provides a long-term, type-specific, fast-charging load spatiotemporal distribution prediction system, using the specific implementation of the prediction method described in the embodiment. The system includes:

[0180] Base year curve generation module: clusters charging load data into multiple types based on vehicle type, space type, and date type, and generates multiple clusters of typical charging load curves for the base year;

[0181] A target year curve generation module; establishing a vehicle ownership prediction model to predict the target year vehicle ownership, and modifying the typical charging load curve of the base year according to the target year vehicle ownership to obtain the target year charging load curve of the target year;

[0182] A fast charging curve generation module is used to perform cluster analysis on the target year type charging load curve according to the spatiotemporal characteristics, and identify and extract the target year fast charging load curve;

[0183] The target year fast charging curve prediction module obtains the temporal and spatial heterogeneity influencing factors of the target year, and predicts the correction coefficient of the target year based on the influencing factors, corrects the target year fast charging load curve, and obtains the target year fast charging load prediction result.

[0184] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A long-term fast charging load spatiotemporal distribution prediction method based on different types of fast charging loads, characterized by: include: The charging load data is clustered into multiple types based on vehicle type, space type, and date type, and a typical charging load curve of multiple clusters in the base year is generated. Establishing a vehicle ownership prediction model to predict the target year vehicle ownership, and revising the typical charging load curve of each type in the base year according to the target year vehicle ownership to obtain the target year vehicle ownership curve of each type; Performing cluster analysis on the target year type-based charging load curves based on spatiotemporal characteristics to identify and extract the target year fast charging load curve; The spatiotemporal heterogeneity influencing factors of the target year are obtained, and a correction coefficient of the target year is predicted based on the influencing factors. The fast charging load curve of the target year is corrected to obtain a fast charging load prediction result of the target year.

2. A long-term type-based fast charging load spatiotemporal distribution prediction method according to claim 1, characterized in that: The process of constructing the typical charging load curve for the base year includes: The original charging load data is divided into multiple typical clusters and reorganized in combination with time nodes to form a four-dimensional tensor structure of charging load data X∈R C×D×S×T ; Where C is the vehicle type, D is the date type, S is the space type, T is the time node, T∈{1,2,...,24}; Let X∈R C×D×S×T Normalized to the interval [0,1]; Perform one-hot encoding on vehicle type, date type and space type to generate cluster labels as conditional feature vectors C∈R C×D×S ; Where C×D×S is the number of clusters C∈R C×D×S and the normalized X∈R C×D×S×T After splicing, the data are input into a conditional variational autoencoder for feature learning, and a typical charging load curve of the base year by type is generated for each cluster.

3. The method for predicting the spatiotemporal distribution of long-term fast charging load by type according to claim 2 is characterized in that: The conditional variational autoencoder adopts an encoder-decoder architecture, including: The encoder receives C∈R C×D×S and the normalized X∈R C×D×S×T , from X∈R C×D×S×T Extract the data mean vector μ and logarithmic standard deviation vector σ of each cluster, and obtain the latent variable z through random sampling; The decoder is composed of latent variables z and conditional feature vector C∈R C×D×S , reconstruct the sequence of 24-hour charging load values, and output the charging load feature reconstruction data X′∈R C×D×S×T , generate the typical charging load curve of the base year by type.

4. The method for predicting the spatiotemporal distribution of long-term fast charging load by type according to claim 1 is characterized in that: The forecast of the target annual holdings includes: Obtain vehicle ownership data for each type over multiple historical years, as well as the corresponding population, GDP per capita, and number of charging stations. Standardize the population, GDP per capita, and number of charging stations to use as standardized variables. A generalized logistic growth model with a dynamic shape factor was used to model and fit the vehicle ownership of each type and the total vehicle ownership, and the target annual vehicle ownership forecast results for each vehicle type were obtained. The dynamic shape factor takes the standardized variable as input and optimizes the growth model parameters by using the Levenberg-Marquardt algorithm in the nonlinear least squares method.

5. The method for predicting the spatiotemporal distribution of long-term fast charging load by type according to claim 1 is characterized in that: The typical charging load curves by type in the base year are revised, including: According to the linear relationship between the total charging load and the number of vehicles in stock, the growth rate of charging load is replaced by the growth rate of vehicle stock, which is defined as: Among them, α i,j,k is the charging load growth rate of vehicle type i under space type j and date type k; β i is the growth rate of electric vehicle ownership of vehicle type i; and are the target annual population and base annual population of vehicle type i, respectively; The charging load growth rate is multiplied by the load data at 24 time points in the typical charging load curve of the base year to obtain the load data of the charging load curve of the target year, thereby obtaining the charging load curve of the target year.

6. The method for predicting the spatiotemporal distribution of long-term fast charging load by type according to claim 1 is characterized in that: Extraction of the target year fast charging load curve includes: Calculate the morphological similarity between charging load curves of different target year types, and construct a similarity matrix S based on the similarity; An undirected weighted graph G is constructed using the similarity matrix S. In the undirected weighted graph G, each node represents a target year type charging load curve, and the edge weight between nodes is the morphological similarity between the two target year type charging load curves; Spectral clustering is performed based on the undirected weighted graph G, and spectral gap analysis is used to determine the optimal number of clusters c; Statistical characteristics are calculated for each of the c clusters. Load curves with fast charging characteristics are identified based on the statistical characteristics to obtain the target year fast charging load curve.

7. The method for predicting the spatiotemporal distribution of long-term fast charging load by type according to claim 1 is characterized in that: The target annual fast charging load curve is modified, including: The XGBoost model is used to train the target year load correction model. The model output is the 24-hour correction coefficient of the target year, and the correction coefficient curve of the target year typical load curve is obtained; The input of the target year load correction model consists of two parts: the first part is a 6-dimensional discrete variable generated by one-hot encoding after combining space type × date type; the second part is a 4-dimensional continuous variable composed of the spatiotemporal heterogeneity influencing factors of the target year; Multiply the normalized ratio of the correction coefficient at each time point by the load data at the corresponding time point in the target year fast charging load curve to obtain the target year fast charging load forecast result.

8. The method for predicting the spatiotemporal distribution of long-term fast charging load by type according to claim 1 is characterized in that: The factors affecting the spatiotemporal heterogeneity in the target year include the proportion of fast charging facilities, charging pile density, road network density and urbanization rate in the target year, and the prediction methods are as follows: Collect data on the number of fast charging piles and total charging piles in the three years before the base year, calculate the average annual growth rate of the number of fast charging piles, and use linear extrapolation based on the fast charging percentage in the base year to obtain the fast charging facility percentage in the target year; Extract the charging pile density growth data for the three years before the base year, calculate the average annual growth rate of charging pile density, and multiply the average annual growth rate by the charging pile density in the base year to obtain the charging station density in the target year; Based on the total length of newly added roads and the regional area in the three years before the base year, the annual increase in road length per unit area is calculated and superimposed on the road network density in the base year to obtain the road network density in the target year; The average annual growth rate of the urbanization rate is calculated based on the urbanization rate in the three years before the base year, and the urbanization rate in the target year is calculated based on the said average annual growth rate.

9. The method for predicting the spatiotemporal distribution of long-term fast charging load by type according to claim 1, characterized in that: training the target year load correction model; An XGBoost model is trained based on the temporal and spatial heterogeneity influencing factors and load correction coefficients in the base year to capture the mapping relationship between the temporal and spatial heterogeneity influencing factors and the load correction coefficients. The series of correction factors for the base year is defined as Among them, λ jt Represents the correction coefficient of space type j at time point t, R base,jt and R CVAE,jt are the original charging load data of the base year, R CVAE,jt Charging load data of typical charging load curves by type in the base year; The XGBoost input sample consists of two parts: the first part is a 6-dimensional discrete variable generated by one-hot encoding after combining space type × date type; the second part is a 4-dimensional continuous variable composed of the spatiotemporal heterogeneity influencing factors of the base year; During the training process, the multi-output regression model wrapper is used to perform synchronous regression modeling of correction coefficients at multiple time points, and multiple XGBoost regression sub-models are constructed based on the gradient boosting tree.

10. A long-term, type-based, spatiotemporal distribution prediction system for fast-charging loads, utilizing the method according to any one of claims 1 to 9, characterized in that: The system comprises: Base year curve generation module: clusters charging load data into multiple types based on vehicle type, space type, and date type, and generates multiple clusters of typical charging load curves for the base year; A target year curve generation module; establishing a vehicle ownership prediction model to predict the target year vehicle ownership, and modifying the typical charging load curve of the base year according to the target year vehicle ownership to obtain the target year charging load curve of the target year; A fast charging curve generation module is used to perform cluster analysis on the target year type charging load curve according to the spatiotemporal characteristics, and identify and extract the target year fast charging load curve; The target year fast charging curve prediction module obtains the temporal and spatial heterogeneity influencing factors of the target year, and predicts the correction coefficient of the target year based on the influencing factors, corrects the target year fast charging load curve, and obtains the target year fast charging load prediction result.

Images