Guiding optimization scheduling method and system for orderly charging of electric vehicle based on big data

Abstract

The invention relates to the technical field of electric vehicles and charging scheduling, and discloses an electric vehicle ordered charging guiding optimization scheduling method and system based on big data, and the method comprises the steps: collecting and preprocessing electric vehicle data, traffic data, electric power data and charging facility data; based on the preprocessed user historical data, dividing the electric vehicle users into different user clusters by adopting a K-Means clustering algorithm, and predicting the time-space distribution characteristics of the charging load by using a time sequence prior fitting network TimePFN model; and constructing a charging guide model under the power-traffic constraint, taking the predicted space-time distribution characteristics of the charging load as input, taking the user comprehensive overhead minimization as a target, considering the power constraint, the traffic constraint and the charging facility constraint, and adopting a point-by-point transfer learning Gaussian Bayesian mixture optimization PPTL-MGTO algorithm to solve an optimal charging scheduling scheme. According to the invention, accurate prediction of the charging load can be realized under the condition of scarcity of multi-source heterogeneous data.

Description

Technical Field

[0001] This invention relates to the field of electric vehicle and charging scheduling technology, and in particular to a guided optimization scheduling method and system for orderly charging of electric vehicles based on big data. Background Technology

[0002] With the rapid popularization of new energy vehicles, the number of electric vehicles (EVs) is growing exponentially, and their concentrated charging behavior has put significant pressure on urban power distribution networks and transportation systems. On the one hand, large-scale disorderly charging can easily cause local overload of the power distribution network, voltage fluctuations, and transformer heat loss; on the other hand, the concentrated flow and queuing charging behavior of electric vehicles on urban roads exacerbates traffic congestion, making the coupling problem between the power system and the transportation system increasingly prominent.

[0003] Current research on orderly charging scheduling mainly focuses on load balancing on the power side or economic optimization on the user side, but lacks joint modeling of traffic flow and grid conditions. Furthermore, real-world scenarios suffer from data scarcity or imbalance, such as a lack of historical load records for newly built charging stations and limited behavioral samples of new user groups, leading to poor model generalization. In addition, existing optimization algorithms often rely on fixed parameters or single-source Bayesian inference, making it difficult to adapt to dynamic, multi-constraint problems across regions and time periods.

[0004] To address the aforementioned shortcomings, there is an urgent need for a new method that integrates synthetic data augmentation, time-series predictive modeling, and intelligent optimization decision-making. This method should be able to accurately predict charging loads under conditions of scarce multi-source heterogeneous data and output feasible scheduling schemes under the dual constraints of power and transportation, thereby improving the overall safety and economy of the system. Summary of the Invention

[0005] To address the problems of data scarcity, inaccurate prediction, and complex optimization solutions in existing electric vehicle charging scheduling, this paper proposes a big data-driven method and system for orderly charging guidance and optimization scheduling of electric vehicles.

[0006] In a first aspect, embodiments of the present invention provide a method for optimizing and scheduling orderly charging of electric vehicles based on big data, comprising the following steps:

[0007] Data on electric vehicles, traffic, electricity, and charging facilities are collected through a big data platform, and the data is preprocessed.

[0008] Based on the user historical data in the preprocessed data, the K-Means clustering algorithm is used to divide electric vehicle users into different user clusters. Furthermore, by integrating user cluster characteristics, traffic flow data and power distribution network load data, the TimePFN model of the time series prior fitting network is used to predict the spatiotemporal distribution characteristics of charging load.

[0009] A charging guidance model under power-traffic constraints is constructed, taking the predicted spatiotemporal distribution characteristics of charging load as input, minimizing the comprehensive cost of users as the goal, considering power constraints, traffic constraints and charging facility constraints, and using the point-by-point transfer learning hybrid Gaussian Bayesian optimization PPTL-MGBO algorithm to solve the optimal charging scheduling scheme.

[0010] Further, the data preprocessing includes data cleaning and standardization processing, and synthetic data enhancement; the synthetic data enhancement includes: for the data involving scarce or imbalanced data in the real scene after data cleaning and standardization processing, using the LMC-Synth method to perform synthetic data enhancement to generate multivariate time series data.

[0011] Further, the synthetic data enhancement using the LMC-Synth method includes:

[0012] The number L of latent functions is sampled from the Weibull distribution, ensuring that L is between the preset minimum number n and the total number of variables N;

[0013] The Kernel-Synth method is used to generate L independent latent functions by randomly combining Gaussian process kernels; each latent function corresponds to a class of basic time series features, and the classes of the basic time series features include periodicity and trend;

[0014] The L independent latent functions are linearly combined into a multivariate time series by a linear collaborative regionalization model, ensuring that the variables have real correlation.

[0015] Further, the Gaussian process kernel includes an additive combination kernel function of a linear kernel and a periodic kernel, and its expression is:

[0016]

[0017]

[0018]

[0019] wherein, : a linear kernel function for capturing linear trend characteristics of the time series; : a periodic kernel function for capturing periodic characteristics of the time series; is an exponential function for mapping the kernel function calculation result to a non-negative interval; : an index of the first time point for positioning a specific time in the time series; : an index of the second time point, together with determines the time window for kernel function calculation; is a circular constant for calculating a periodic related trigonometric function term; is the absolute difference between two time points, reflecting the time interval between two time instants; : the period length of the period kernel, i.e., the time period at which the time series repeats, set as 24 hours in the charging load forecasting and corresponding to the daily period; : the width parameter of the period kernel, controlling the smoothness of the periodic feature.

[0020] Further, the formula for linearly combining L independent latent functions into a multivariate time series by the linear co-localization model is:

[0021]

[0022] wherein, : the value of the i-th variable at the time instant t, representing a specific dimension data in the synthesized multivariate time series (MTS), including: the load of charging station A or the traffic flow of road section B; : variable index, taking values in the range of , used to distinguish different types of variables; : time point index, representing a specific time instant, used to locate the time position in the time series; : the weight coefficient of the i-th variable to the j-th latent function, controlling the influence of the latent function on the variable : the value of the j-th latent function at the time instant t, generated by the Kernel-Synth method, representing a class of basic time series features. : the value of the i-th variable at the time instant t, representing a specific dimension data in the synthesized multivariate time series (MTS), including: the load of charging station A or the traffic flow of road section B; : variable index, taking values in the range of , used to distinguish different types of variables; : time point index, representing a specific time instant, used to locate the time position in the time series; : the weight coefficient of the i-th variable to the j-th latent function, controlling the influence of the latent function on the variable : the value of the j-th latent function at the time instant t, generated by the Kernel-Synth method, representing a class of basic time series features.

[0023] Further, wherein the adopting the LMC-Synth method for synthetic data enhancement further comprises: supplementing independent variable scenarios, generating variable-independent synthetic data batches to supplement real data samples to simulate and cover low correlation or independence events in real scenarios, specifically including:

[0024] The number of latent functions L is set to be equal to the total number of variables N of the multivariate time series;

[0025] The weight coefficient matrix is configured as: when the variable index i = the latent function index j, = 0, so that each variable is only influenced by the corresponding one latent function ​​​The variables are independent of each other and have no correlation between them.

[0026] The proportion of the independent variable synthetic data in the total generated data is set to be not less than 25%.

[0027] Further, the user historical data includes charging frequency, single charging amount and charging time period preference; and the different user clusters include commuting type users, commercial type users and family type users.

[0028] Further, the TimePFN model predicts the spatio-temporal distribution characteristics of the charging load, including:

[0029] Model input construction: convert the preprocessed multi-source data into a fixed-length time series input window, i.e. an input matrix;

[0030] One-dimensional convolution and residual connection are performed on each variable of the input matrix to extract local time series features, including load rising trend and traffic congestion mutation;

[0031] The local time series features after feature extraction are segmented into overlapping patches and mapped into fixed-dimension embedding vectors;

[0032] The patch embeddings of all variables are mixed in the channel to obtain a total embedding matrix;

[0033] The total embedding matrix is converted into a query (Q), key (K), and value (V) matrix through a linear layer, and the output feature matrix is obtained through multi-head attention calculation;

[0034] The output feature matrix is input into the Transformer encoder for iterative processing to capture cross-variable dependencies;

[0035] The output results of the Transformer encoder are reorganized by variable and the final charging load spatio-temporal distribution prediction results are generated.

[0036] Further, the formula for segmenting the local time series features after feature extraction into overlapping patches and mapping them into fixed-dimension embedding vectors is:

[0037]

[0038] wherein, OverlapExtract is an overlapping patch extraction function; P is the patch size; S is the step; PE is a two-dimensional sine position encoding; : a 2-layer feed-forward network (Feed-Forward Network) for mapping the flattened patch vector to a fixed-dimension embedding space; Flatten : flatten function, convert the two-dimensional patch matrix into a one-dimensional vector, facilitate the processing of the feedforward network; : the first variable after feature extraction of local time sequence characteristics; : the first variable after overlapping patch extraction, flattening, feedforward network mapping and position coding of the final embedding vector set.

[0039] Further, the Transformer encoder adopts an 8-layer structure, each layer containing a multi-head attention module, a feedforward network and layer normalization, wherein the multi-head attention calculation adopts 8 attention heads.

[0040] Further, wherein, taking the minimization of the user's comprehensive cost as the goal, the objective function F considering the power constraint, the traffic constraint and the charging facility constraint is as follows:

[0041]

[0042] wherein, is the total number of electric vehicles to be scheduled; is the charging unit price of the user , unit: yuan / kWh; is the charging demand of the user , unit: kWh; is the traffic travel time of the user from the current location to the target charging station, unit: minute; is the traffic time weight coefficient; is the waiting time of the user at the target charging station, unit: minute; is the waiting time weight coefficient, which can be dynamically adjusted according to the user's demand preference.

[0043] Further, wherein, the PPTL-MGBO algorithm is used to solve the optimal charging scheduling scheme, including:

[0044] The historical similar scenarios are taken as the source domain, and multiple Gaussian process (GP) models are constructed for pre-training;

[0045] The initial data of the current scheduling scenario is taken as the initial distribution of the target domain, and a mixed Gaussian distribution is constructed together with each source domain model;

[0046] The similarity factor of each optimized charging scheduling scheme in the target domain and each domain is calculated, and the similarity weight is obtained based on the Bayesian rule;

[0047] Based on the similarity weight, the weighted posterior mean and variance of the target domain model are calculated;

[0048] The iteration update of the optimal scheduling scheme is achieved by maximizing the acquisition function with the weighted expected improvement EI as the acquisition function.

[0049] Further, the formula for constructing the mixed Gaussian distribution is:

[0050]

[0051] wherein, : the probability density function of the mixed Gaussian distribution; : the weight coefficient of the zth domain; : the Gaussian distribution probability density function of the zth domain; is the number of source domains, i.e., the total number of historical similar scenarios.

[0052] Further, the formula of the weighted expected improvement EI is:

[0053]

[0054] wherein, : the expected improvement value of the scheme under the current data set ; : the charging scheduling scheme to be evaluated; : the data set of the zth iteration; Φ is the cumulative distribution function of the standard normal distribution, is the probability density function of the standard normal distribution; : the weighted posterior mean of the scheme ; : the weighted posterior standard deviation of the scheme .

[0055] Further, the power constraint condition is that the load of each node does not exceed the rated capacity and the voltage deviation is within the allowed range through power flow calculation of the distribution network;

[0056] The traffic constraint condition is to limit the traffic flow of the road section around the charging station to be less than the congestion threshold, and to quantify the influence of road network congestion on the travel time through the road impedance function;

[0057] The charging facility constraint is to consider the power limit and operating state of the charging pile, and to ensure that the scheduled charging demand does not exceed the real-time available capacity of the charging facility.

[0058] In a second aspect, the embodiments of the present application further provide an electric vehicle orderly charging guidance optimization scheduling system based on big data, which is used to implement the electric vehicle orderly charging guidance optimization scheduling method based on big data as described in the first aspect, and the system comprises:​

[0059] a data acquisition and preprocessing module for collecting multi-source data and performing cleaning, standardization and synthetic data enhancement processing;

[0060] a load prediction module for dividing user clusters and predicting the spatio-temporal distribution characteristics of charging load;

[0061] a scheduling optimization module for constructing a charging guidance model under the constraints of power and traffic and solving an optimal scheduling scheme.

[0062] Compared with the prior art, the present application has the following beneficial effects:

[0063] 1. The present application proposes a synthetic data enhancement method based on the LMC-Synth method. Multivariate time series data is generated by combining a linear collaborative regionalization model (LMC) and a Gaussian process kernel combination technology (Kernel-Synth), which supplements the scarce load and traffic data in the real scene, realizes high-fidelity data expansion, and provides sufficient time series sample support for subsequent modeling.

[0064] 2. The present application proposes a generated independent latent function mechanism based on Kernel-Synth. The mechanism generates independent latent functions with trend, periodicity and randomness characteristics through adaptive combination of kernel functions, thereby preserving the real physical laws and statistical dependencies in multivariate modeling, and significantly improving the authenticity and diversity of data simulation.

[0065] 3. The present application fuses user cluster characteristics, traffic flow data and distribution network load data to construct a TimePFN model. The TimePFN model uses overlapping patch embedding, Transformer multi-head attention mechanism and time prior feature extraction based on multi-source data input to accurately predict the spatio-temporal distribution characteristics of electric vehicle charging load and identify potential peak areas and key time periods.

[0066] 4. The present application adopts a point-by-point transfer learning hybrid Gaussian Bayesian optimization (PPTL-MGBO) algorithm. The algorithm fuses source domain historical experience and target domain real-time data, realizes multi-constraint optimization solution through point-by-point similarity transfer and weighted Gaussian process regression, and outputs a user cost minimized charging scheduling scheme under the premise of considering the safety of the distribution network and the efficiency of the traffic flow.

[0067] 5. The application takes the weighted expected improvement (EI) as the acquisition function, maximizes the acquisition function, and realizes the iteration of the optimal candidate scheduling scheme. Under the optimization framework of the PPTL-MGBO algorithm, the weighted expected improvement (EI) function is defined to balance the "local development" and "global exploration", realize the step-by-step iteration and dynamic optimization of the charging scheduling scheme, and accelerate the convergence process of the global optimal solution.

[0068] In summary, the application realizes systematic innovation in three aspects of data enhancement, time series prediction and multi-constraint optimization, not only improves the accuracy and robustness of electric vehicle charging load prediction, but also significantly optimizes the global coordination performance of the charging guidance strategy, and provides a new technical approach for efficient operation of future intelligent energy-transportation integrated system. BRIEF DESCRIPTION OF DRAWINGS

[0069] Other features, objects, and advantages of the application will become more apparent from the following detailed description of non-limiting embodiments when read in conjunction with the accompanying drawings. The drawings are only for the purpose of illustrating preferred embodiments and are not to be construed as limiting the application. Moreover, like reference numerals designate similar parts throughout the drawings. In the drawings:

[0070] Figure 1 is a flowchart of an electric vehicle orderly charging guidance optimization scheduling method based on big data provided by an embodiment of the application;

[0071] Figure 2 is a system architecture diagram of an electric vehicle orderly charging guidance optimization scheduling method based on big data provided by an embodiment of the application;

[0072] Figure 3 is a flowchart of step S2 user charging demand and load space-time distribution prediction provided by an embodiment of the application;

[0073] Figure 4 is a flowchart of step S3 charging guidance model under power-transportation constraints provided by an embodiment of the application;

[0074] Figure 5 is a module diagram of an electric vehicle orderly charging guidance optimization scheduling system based on big data provided by an embodiment of the application. DETAILED DESCRIPTION

[0075] The application will be further described in detail below in conjunction with the drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the application, but not to limit the application. In addition, it should be noted that, for the purpose of description, only the parts related to the application are shown in the drawings, not all the structures.

[0076] Before discussing the example embodiments in more detail, it should be mentioned that some of the example embodiments are described as processes or methods depicted as flow charts. Although the processes are described in a certain order, many of the operations (or steps) can be performed concurrently, in parallel, or simultaneously. In addition, the order of the operations can be re-arranged. The processes can be terminated when their operations are completed, but can also have additional steps not included in the figure. The processes can correspond to methods, functions, procedures, subroutines, subprograms, etc.

[0077] In the technical scheme of the present application, the acquisition, storage, use, and processing of data comply with relevant provisions of national laws and regulations. It should be noted that in the embodiments of the present application, some existing industry solutions such as software, components, or models may be mentioned, which should be considered as exemplary, and the purpose is only to illustrate the feasibility of the implementation of the technical scheme of the present application, but it does not mean that the applicant has or will necessarily use the solution.

[0078] Figure 1 is a flowchart of a big data-based orderly charging guidance optimization scheduling method for electric vehicles provided by an embodiment of the present application; Figure 2 is a system architecture diagram of a big data-based orderly charging guidance optimization scheduling method for electric vehicles provided by an embodiment of the present application. As shown in Figure 1 and Figure 2 The big data-based orderly charging guidance optimization scheduling method 100 for electric vehicles includes the following steps:

[0079] S1: Collect electric vehicle data, traffic data, power data, and charging facility data through a big data platform, and perform data preprocessing.

[0080] Step S1 is used to realize data collection and preprocessing. Specifically, it includes:

[0081] Step S1.1 Multi-source data collection:

[0082] Collect four types of core data through a big data platform to build a full-dimensional database:

[0083] (1) Electric vehicle data: including real-time vehicle location, battery remaining capacity (SOC), charging demand (such as expected charging capacity, charging cutoff time), historical charging trajectory;

[0084] (2) Traffic data: including city road network topology, real-time traffic flow, road congestion index, charging facility surrounding road travel time;

[0085] (3) Power data: including real-time load of each node of the distribution network, voltage deviation, output of distributed power sources (such as photovoltaic and wind power), transformer load rate;

[0086] (4) Charging facility data: including charging site location, charging pile power level, real-time running state (idle / occupied), queuing waiting time.

[0087] Step S1.2: Preprocessing of collected data:

[0088] S1.2.1: Cleaning operation:

[0089] Delete abnormal values (such as battery SOC exceeding reasonable range, invalid geographic location data), fill in missing values (use adjacent time step mean to fill in short missing data), remove duplicate records;

[0090] S1.2.2: Standardization operation:

[0091] Map data of different dimensions (such as charging power, traffic flow) to the [0, 1] interval, construct a unified format structured database, and facilitate subsequent fusion analysis.

[0092] S1.2.3: Synthetic data enhancement based on LMC-Synth method:

[0093] For the problem of data scarcity in some fields in real scenarios (such as historical load data of new charging stations, charging behavior data of new user groups), the LMC-Synth (Linear Co-Regionalization Model combined with Gaussian Process Kernel) method is used to generate large-scale synthetic multivariate time series (MTS) data.

[0094] Further, for the data involving scarce or imbalanced data in real scenarios after data cleaning and standardization, the LMC-Synth method is used for synthetic data enhancement to generate multivariate time series data, including: sampling the number of potential functions L from the Weibull distribution, ensuring that L is between the preset minimum number n and the total number of variables N; using the Kernel-Synth method to generate L independent potential functions by randomly combining Gaussian process kernels; wherein each potential function corresponds to a class of basic time series features, and the class of basic time series features includes periodicity and trend; linearly combining the L independent potential functions into a multivariate time series through a linear co-regionalization model to ensure that the variables have real correlation. Specifically, the following sub-steps are included:

[0095] Step S1.2.3.1: Determine the number of potential functions

[0096] Sample the number of potential functions from the Weibull distribution , ensuring that the preset minimum number is between between, to avoid the distortion of synthetic data caused by extreme values, the formula is as follows:

[0097]

[0098] : the final sampling number of the sampling potential function, used for subsequent combination to generate multivariate time series; max : the maximum function, to ensure that the number of potential functions is not less than the preset minimum number , to avoid incomplete coverage of time series patterns due to too few potential functions; : the minimum function, to ensure that the number of potential functions does not exceed the total number of variables of the data to be synthesized , to avoid resource waste and data redundancy; Weibull : Weibull probability distribution function, used to randomly sample the initial number of potential functions, and its output is a random integer conforming to the distribution; : the shape parameter of Weibull distribution, to control the steepness of the distribution curve, the larger the value, the more concentrated the distribution, which is estimated according to the characteristics of historical data, and set to 1.2 in the experiment; : the scale parameter of Weibull distribution, used to control the sampling range, set to ( total number of variables), to ensure that the sampling range matches the variable scale; : the total number of variables of the multivariate time series (MTS) to be synthesized, such as containing 8 variables including charging load, traffic flow, and distributed power output, ; : the preset minimum number of potential functions, to ensure that at least 3 types of core time series patterns (such as trend, periodicity, and randomness) can be captured, set to 3 in the experiment.

[0099] Step S1.2.3.2: generating independent potential functions based on Kernel-Synth

[0100] Using the Kernel-Synth method, generate independent potential functions , each potential function corresponds to a type of basic time series feature (such as periodicity, trend). Generate a periodic sequence with linear trend, suitable for simulating the daily periodic change of charging load (such as morning peak and evening peak), the kernel function formula is:

[0101]

[0102] , where : the additive combination kernel function of linear kernel and periodic kernel, output the two kernel functions at time and : the sum of the values at the two time points, used to generate periodic time series with linear trend (e.g. daily cycle of charging load); : linear kernel function, used to capture the linear trend of time series (e.g. the slow upward / downward trend of charging load over time), whose formula is ; : the index of the first time point (e.g. "5th minute", "10th minute"), used to locate the specific time in the time series; : the index of the second time point, together with determines the time window for kernel calculation; : periodic kernel function, used to capture the periodicity of time series (e.g. daily peak, weekly peak of charging load), whose formula is ; : exponential function, used to map the kernel calculation results to the non-negative interval, consistent with the semi-positive requirement of Gaussian process kernel; : pi (approximately 3.1416), used to calculate the trigonometric term related to periodicity; : the absolute difference between and , reflecting the time interval between the two time points; : the period length of the periodic kernel, i.e. the time period of the repeated occurrence of time series, set to 24 hours (corresponding to daily cycle) in charging load prediction; : the width parameter of the periodic kernel, controlling the smoothness of the periodicity, the smaller the value the more acute the periodicity, set to 0.8 in the experiment.

[0103] Step S1.2.3.3: Generate Multivariate Sequence with Correlation

[0104] Through the Linear Co-kriging Model (LMC), independent latent functions are linearly combined into multivariate time series, ensuring that variables have real correlation (e.g. charging load is positively correlated with traffic flow, charging load is negatively correlated with distributed power output), the formula is as follows:

[0105]

[0106] : the value of the th variable at time , representing a specific dimension data in the synthesized multivariate time series (MTS), such as "charging station A's load (kW)", "road section B's traffic flow (vehicles / hour)", etc.; : variable index, taking values in the range of ( is the total number of variables), used to distinguish different types of variables; : time point index, representing a specific time point (e.g., "1st hour", "2nd minute", etc.), used to locate the time position in the time series; : summation symbol, indicating the accumulation of all terms from to , i.e., summarizing the contribution of all potential functions to the current variable; : total number of potential functions, representing the number of basic time series patterns used for combination to generate multivariate sequences; : weight coefficient of the th variable to the th potential function, obtained by sampling from a symmetric Dirichlet distribution with parameters [1, 1,..., 1] to simulate possible correlation patterns among variables and control the influence of the potential function on the variable ; : value of the th potential function at time , generated by the Kernel-Synth method in step S1.2.3.2, representing a class of basic time series patterns (e.g., periodicity, trend).

[0107] Step S1.2.3.4: Supplement independent variable scenarios

[0108] Further, in some embodiments of the present application, the LMC-Synth method for synthetic data augmentation further includes: supplementing independent variable scenarios, generating variable-independent synthetic data batches to supplement real data samples to simulate and cover low correlation or independence events in real scenarios, specifically including: setting the number of potential functions L equal to the total number of variables N of the multivariate time series; the weight coefficient matrix is configured as: when variable index i = potential function index j, = 0 when variable index i ≠ potential function index j, so that each variable is independently generated by only one corresponding potential function ; the proportion of variable-independent synthetic data in the total generated data is set to no less than 25%, covering low correlation scenarios (such as suburban charging station load and downtown traffic flow), and finally generating more than 100,000 pieces of synthetic MTS data to supplement real data samples.

[0109] S2: Based on the user historical data in the preprocessed data, the K-Means clustering algorithm is used to divide electric vehicle users into different user clusters, and the user cluster features, traffic flow data and power distribution network load data are fused to use the TimePFN model to predict the spatio-temporal distribution characteristics of the charging load.

[0110] Figure 3 is a flowchart of the step S2 of the embodiment of the present application for predicting user charging demand and load spatio-temporal distribution. The step S2 is configured to realize the prediction of user charging demand and load spatio-temporal distribution. As shown in the figure, the step S2 specifically includes the following steps: Figure 3

[0111] Step S2.1: User cluster division

[0112] Based on the preprocessed user historical data (such as charging frequency, single charging amount, charging time period preference), the feature vectors including daily average charging frequency, average single charging amount, and typical charging time period are extracted, and the K-Means clustering algorithm is used to divide the electric vehicle users into different cluster categories. The number K of clusters can be determined by the elbow method or according to business requirements, for example, divided into “commuter users”, “commercial users”, and “family users” three categories, and the charging behavior characteristics of different clusters are determined.

[0113] Step S2.2: Spatio-temporal load prediction based on TimePFN model

[0114] The TimePFN model is used to predict the spatio-temporal distribution characteristics of charging load by fusing user cluster characteristics, traffic flow data, and power distribution network load data.

[0115] Further, the TimePFN model is used to predict the spatio-temporal distribution characteristics of charging load, including: constructing model input: converting the preprocessed multi-source data into a fixed-length time series input window, i.e., an input matrix; performing one-dimensional convolution and residual connection on each variable of the input matrix to extract local time series features, including load rising trend and traffic congestion mutation; dividing the local time series features after feature extraction into overlapping patches and mapping them into fixed-dimension embedding vectors; mixing the patch embeddings of all variables in the channel to obtain a total embedding matrix; converting the total embedding matrix into a query (Q), key (K), and value (V) matrix through a linear layer, and calculating through multi-head attention to obtain an output feature matrix; inputting the output feature matrix into a Transformer encoder for iterative processing to capture cross-variable dependencies; reorganizing the output results of the Transformer encoder by variable and generating the final charging load spatio-temporal distribution prediction results. More specifically, the following sub-steps are included:

[0116] Step S2.2.1: Constructing model input

[0117] Converting the preprocessed multi-source data into a fixed-length time series input window , the formula is as follows:

[0118] ​

[0119] is the input of the model; (input time step, corresponding to 4 hours, one data point every 5 minutes); is the total number of variables (e.g., 1 user cluster label, 3 traffic flow data, 2 power grid load data, 2 charging facility status data, a total of ; is the user cluster feature matrix (e.g., "1" represents a commuter user, "2" represents a commercial user); is the traffic flow matrix (e.g., road congestion index, travel time, traffic volume); is the power grid data matrix (e.g., node load, voltage deviation); is the charging facility data matrix (e.g., charging pile idle rate, queue length). Step S2.2.2: Time series feature extraction One-dimensional (1D) convolution and residual connection are performed on each variable of the input matrix

[0120] to extract local time series features (e.g., load rising trend, traffic congestion mutation), and the formula is as follows:

[0121]

[0122]

[0123] : the output feature matrix of the th variable after 1D convolution and residual connection, with a dimension of , containing original variable information and multiple sets of locally extracted time series features; : variable index, with a value range of (the total number of input variables), used to distinguish different variables (e.g., traffic flow, power grid load) in the input matrix ; Concat : concatenation function, combining 1D convolution results into a matrix along the variable dimension (i.e., feature dimension), ensuring that multiple sets of local time series features are retained; is the result set of 1D convolution operations, where is the convolution kernel index , each corresponding to a local feature extraction method; Conv1D : the th 1D convolution operation, used to extract the variable ​​​​local temporal features (e.g. load rising trend, traffic flow mutation) of the time series; : total number of convolution kernels, corresponding to 9 different local temporal features; : input matrix : the th variable in the dimensional input matrix, representing the input time step (corresponding to 4 hours, with one data point every 5 minutes); : residual connection operator, expanding the original variable along the feature dimension to avoid information loss caused by convolution operation.

[0124] Step S2.2.3: Overlapping patch embedding

[0125] The local temporal features extracted in step 2.2.2 are divided into overlapping patches and mapped into fixed-dimension embedding vectors, as follows:

[0126] : the final embedding vector set of the th variable after overlapping patch extraction, flattening, feed-forward network mapping, and position encoding, with each embedding vector having a dimension of , and a total of 12 embedding vectors generated for a single variable; : variable index, consistent with in step S2.2.2, corresponding to the th variable in the input matrix; : 2-layer feed-forward network (FFN) for mapping the flattened patch vector to a fixed-dimension embedding space, enhancing feature expression capability; activation function: ReLU, used to introduce non-linear transformation, improving the model's fitting ability for complex features; Flatten : flattening function, converting the two-dimensional patch matrix to a one-dimensional vector for easy processing by the feed-forward network; OverlapExtract : overlapping patch extraction function, dividing the variable extracted from the temporal features into multiple overlapping local patches to capture local temporal information, with a single patch dimension of , and a total of patches extracted for each variable; input : feature matrix of the th variable, with a dimension of ; : patch size, i.e. each patch contains 16 consecutive time steps of data, balancing local information granularity and computational efficiency; : Step size, i.e. the window moves 8 time steps each time the patch is extracted, ensuring that there is an overlap of 8 time steps between patches to avoid losing temporal continuity; PE: 2D (2D) sinusoidal position encoding, used to inject temporal position and channel position information for embedding vectors, making up for the defect that the Transformer model is not sensitive to position, the formula is:

[0127]

[0128] : Temporal index of the patch, the value range is (12 patches for a single variable), representing the position of the patch in the time series; : Channel index, the value range is (128 dimensions of embedding vectors), representing the channel position of the embedding vector; = 128: The dimension of the embedding vector is consistent with the output dimension of the feedforward network; : Scale factor of position encoding, used to adjust the encoding frequency of different channel positions, to ensure that different channels capture position information of different scales.

[0129] Step S2.2.4: Channel mixing and Transformer encoding

[0130] Merge the patch embeddings of all variables obtained in step S2.2.3, input the Transformer encoder to capture cross-variable dependencies (such as the correlation of "charging load-traffic flow" and "voltage deviation-charging pile power"). Specifically, the following sub-steps are included:

[0131] Step S2.2.4.1: Channel mixing

[0132] Merge the patch embeddings of all variables obtained in step S2.2.3 to obtain the total embedding matrix Patch :

[0133]

[0134] Patch : Total embedding matrix after merging patch embeddings of all variables, containing all patch tokens of multiple variables, used for subsequent cross-variable dependency capture by the Transformer encoder; : Patch embedding set of the 1st to the variable, each contains 12 embedding vectors with a dimension of (12 patches are extracted for a single variable, and 12 is calculated in step S2.2.3); The matrix concatenation operator embeds and merges patches for all variables along the "number of patch tokens" dimension (i.e., the first dimension), ensuring that patch tokens for different variables can be processed uniformly. : Input the total number of variables, which is set to 8 in the experiment (such as user cluster labels, traffic flow, power distribution load, charging facility status, etc.). Total number of patch tokens, as tested. This indicates that the total embedding matrix contains 9 patch tokens; The data type and dimensions of the total embedding matrix: the first dimension is the number of patch tokens, and the second dimension is the embedding vector dimension (128 and the feedforward network output dimension in step S2.2.3). Consistent).

[0135] Step S2.2.4.2: Multi-head attention calculation

[0136] The total embedding matrix obtained by channel mixing is transformed into a query (Q), key (K), and value (V) matrix through a linear layer:

[0137]

[0138] MultiHead The output matrix of the multi-head attention module, with dimensions of It integrates information from multiple attention heads to capture cross-variable dependencies from different perspectives; The query matrix is ​​composed of the total embedding matrix. Generated through a linear layer, with dimensions of This is used to "query" information about other tokens; : Key matrix, and Homogeneous generation, dimension is Used with Calculate similarity to determine the weight of information interaction; Value matrix, and Homogeneous generation, dimension is , used according to and Similarity conveys information; Concat : Concatenation function, along the embedding vector dimension (the second dimension) The outputs of each attention head are merged and restored to the original embedding dimension. : The output set of a scaled dot product attention head For attention head index ( Each output dimension is ;ScaledDotProduct : No. A scaled dot product attention head is used to calculate the information interaction result of a single attention head; the formula is shown below. : Total number of attention heads, capturing different types of intervariate dependencies (such as "charging load - traffic flow" and "voltage deviation - charging pile power") through multiple sets of independent attention heads. Output projection matrix, dimension 1 It is used to map the concatenated multi-attention head information into a unified-dimensional output, integrating the features of different attention heads; Softmax The Softmax activation function calculates the probability distribution along the "number of key tokens" dimension (the second dimension), converting similarity into attention weights and ensuring that the sum of the weights is 1. : Query and key matrix multiplication, dimension is Each element represents the first The query token and the first The similarity between individual key tokens reflects the degree of association between them; Scaling factor For a single attention head dimension This is used to alleviate the problem of "excessive dimensionality leading to excessively large similarity values ​​and the disappearance of the Softmax gradient".

[0139] Step S2.2.4.3: Encoder Layer

[0140] The output feature matrix of the multi-head attention computation is iteratively processed by a Transformer encoder. An 8-layer Transformer encoder is used, each layer containing a multi-head attention module, two FFNs (512 hidden layers), and layer normalization, as shown in the formula:

[0141]

[0142] LayerNorm

[0143] : No. The output matrix of the layer Transformer encoder has dimensions of This includes the transvariate features processed by this layer; : Number of encoder layers, with a value range of 100%. , representing the current Transformer encoder layer being processed; LayerNorm Layer normalization operation: Normalizes the input matrix along the "embedding dimension" (the second dimension), setting the mean to 0 and the standard deviation to 1, which stabilizes the training process and accelerates convergence; : the output of the first layer encoder The result of multi-head attention calculation is a matrix with a dimension of ; : the output matrix of the first layer encoder (initial input: total embedding matrix) : residual connection operator. FFN : a 2-layer feedforward network for nonlinear transformation of the features output by the multi-head attention layer, enhancing the model's expression ability.

[0144] Steps S2.2.4.1 to S2.2.4.3 form a complete link of data integration preliminary dependency capture deepening of features. Channel mixing solves the problem of scattered multivariate patches, achieving data-level unification; multi-head attention solves the problem of unquantifiable cross-variable correlation, achieving preliminary modeling of dependency relationships; the encoder layer solves the problem of insufficient expression of shallow dependencies, deepening the learning of complex dependency features through multiple iterations, and finally providing high-quality cross-variable fused features for charge load prediction.

[0145] Step S2.2.5: Load prediction output

[0146] The encoder output obtained in step S2.2.4.3 is reorganized by variable and the final prediction result is generated, as follows:

[0147]

[0148] is reorganized by variable as ; is a 2-layer shared feedforward network (hidden layer dimension 1024, output dimension , corresponding to future 4-hour load prediction); The "charge load variable" column in is the final charge load spatiotemporal distribution prediction result, based on which potential load peak areas (such as commercial district charging stations) and time periods (such as 18:00-20:00) are identified.

[0149] In the embodiment of the present application, the TimePFN model is trained by minimizing the mean square error (MSE) between the predicted output and the real historical load data , and its loss function is . The model training uses the Adam optimizer, the learning rate is set to 0.001, and the batch size is set to 32.

[0150] S3: Construct a charging guidance model under power-traffic constraints. Take the predicted spatiotemporal distribution characteristics of charging load as input and minimize the user's comprehensive cost as the objective. Consider power constraints, traffic constraints and charging facility constraints. Use the point-by-point transfer learning hybrid Gaussian Bayesian optimization PPTL-MGBO algorithm to solve the optimal charging scheduling scheme.

[0151] Figure 4 This is a schematic diagram of the charging guidance model under power-traffic constraints in step S3 of an embodiment of the present invention. Figure 3 As shown, step S3 is used to construct the charging guidance model under the power-traffic constraints.

[0152] Based on the spatiotemporal distribution prediction results of charging load output in step S2.2.5 (which clarifies key information such as peak / valley values ​​and cross-variable correlation characteristics of charging load in different time periods and regions), this step S3 will further realize the practical transformation from load prediction to charging behavior optimization. By integrating information such as load distribution patterns and potential peak risks from the prediction data, and focusing on quantifying the comprehensive user expenditure and meeting the dual-dimensional constraints of electricity and transportation, a spatiotemporally coordinated charging guidance model is constructed. Ultimately, an optimal scheduling scheme that balances distribution network safety, traffic efficiency, and user needs is generated, as detailed below:

[0153] Step S3.1: Define the objective function of the charging guidance model.

[0154] With the core objective of minimizing overall user costs, and taking into account charging electricity costs, travel time, and charging waiting time, the objective function is... The formula is as follows:

[0155]

[0156] in, The total number of electric vehicles to be dispatched; For users The unit price of charging (yuan / kWh); For users Charging demand (kWh); The travel time (in minutes) for user m from the current location to the target charging station; This refers to the traffic time weighting coefficient. For users Waiting time (in minutes) at the target charging station; This is the waiting time weighting coefficient (the weighting coefficient can be dynamically adjusted according to user needs and preferences). The results are calculated based on real-time traffic data and road network topology using path planning algorithms (such as Dijkstra's algorithm). According to the real-time queuing state of the target charging site and the charging pile power estimation.

[0157] Step S3.2: Set power-traffic two-dimensional constraints

[0158] Incorporate power-traffic two-dimensional constraints to ensure the feasibility of the scheduling scheme:

[0159] 1. Power constraint: Determine the upper limit of each node load through power distribution network flow calculation, to ensure that the node load of the power distribution network after charging scheduling does not exceed the rated capacity, and the voltage deviation is within the range, the formula is as follows:

[0160]

[0161] is the real-time load (kW) of the power distribution network node ; is the upper limit of the rated load (kW) of node ; is the voltage (kV) of node ; and are the upper and lower limits of the voltage (kV) of node , respectively.

[0162] 2. Traffic constraint: Quantify the impact of road congestion through the road impedance function , to limit the traffic flow of the road section around the charging site to not exceed the congestion threshold, the formula is as follows:

[0163]

[0164]

[0165] is the real-time traffic flow (vehicles / hour) of road section ; is the congestion threshold (vehicles / hour) of road section ; is the road impedance function; is the free travel time of road section ; is the congestion coefficient.

[0166] 3. Charging facility constraint: Consider the power limit and operating state of the charging pile, to ensure that the scheduled charging demand does not exceed the real-time available capacity of the charging facility, the formula is as follows:

[0167]

[0168] is the user set scheduled to the th charging site; for the user the charging power (kW); for the i-th charging station the real-time available power (kW).

[0169] Step S3.3: Model solving based on PPTL-MGBO algorithm

[0170] The PPTL-MGBO algorithm is used to solve the above multi-constrained optimization problem.

[0171] Further, the PPTL-MGBO algorithm is used to solve the optimal charging scheduling scheme, including: the historical similar scenarios are taken as source domains, a plurality of Gaussian process (GP) models are constructed for pre-training; the initial data of the current scheduling scenario is taken as the target domain initial distribution, and a mixed Gaussian distribution is constructed together with each source domain model; the similarity factors of each optimized charging scheduling scheme in the target domain and each domain are calculated, and the similarity weights are obtained based on the Bayesian rule; the weighted posterior mean and variance of the target domain model are calculated based on the similarity weights; the weighted expectation improvement EI is taken as the acquisition function, and the optimal scheduling scheme is iteratively updated by maximizing the acquisition function. The specific steps are as follows:

[0172] Step S3.3.1: Source domain model pre-training

[0173] The historical similar scenarios (such as past charging scheduling in the same area, charging data in similar cities) are taken as the “source domain”, a plurality of Gaussian process (GP) models are constructed, and the mapping relationship of “electric vehicle charging demand-power constraint-traffic efficiency” in the source domain is learned respectively to obtain the posterior mean and variance of each source domain model (N is the number of source domains).

[0174] Step S3.3.2: Mixed Gaussian distribution construction

[0175] The initial data of the current scheduling scenario (such as a small amount of real-time charging demand, distribution network load data) is taken as the “target domain initial distribution” (denoted as ), and a mixed Gaussian distribution is constructed together with each source domain model, and the formula is as follows:

[0176]

[0177] : the probability density function of the mixed Gaussian distribution, used to describe the probability distribution of the optimized charging scheduling scheme (feature vector ), which fuses the target domain initial data and the prior knowledge of the multi-source domain; : the feature vector of the optimized charging scheduling scheme, with a dimension of​ For feature dimensions, such as user location, charging power, target site load, etc., representing a specific scheduling decision (such as the combination of "user site 1, user site 2"); : summation symbol, representing the weighted summation of Gaussian distributions of all domains from to , integrating multi-domain information; : domain index, representing the initial distribution of the target domain, representing the first to the source domain; : the number of source domains, i.e. the total number of historical similar scenarios (such as past scheduling data in the same area, charging data in similar cities, set to 3 in the experiment); : the weight coefficient of the th domain, controlling the contribution of the domain in the mixed distribution, satisfying and (such as the target domain initial distribution weight , and each source domain weight ); : the Gaussian (normal) distribution probability density function of the th domain, used to model the distribution of the scheduling scheme feature vector in the domain; : the mean vector of the th domain, with a dimension of , representing the average level of the scheduling scheme features in the domain (such as the average charging power, average travel time of source domain 1); : the covariance matrix of the th domain, with a dimension of , describing the correlation between each dimension of the feature vector in the domain (such as the correlation between charging power and power grid load).

[0178] Step S3.3.3: Point-by-point similarity calculation

[0179] For the charging scheduling scheme to be optimized in the target domain (such as user-charging site matching combination), the similarity factor of each scheme to each domain is calculated, and the similarity weight is obtained based on Bayes rule, as follows:

[0180]

[0181] : the th scheduling scheme to be optimized and the ​The similarity factor between the target domain and the initial / source domain is used to determine the similarity between the proposed solution and the target domain. A higher value indicates a more similar solution to the target domain. The more similar the distributions; : The index of the scheduling scheme to be optimized, with a value range of . (Total number of candidate solutions in R); Domain Index For the initial target domain, For source domain ; : No. The weight coefficients of each domain (and in the Gaussian mixture distribution) Consistent); The normalization constant of the Gaussian distribution ensures that the integral of the probability density function is 1; Feature dimension (and scheduling scheme feature vector) (The dimensions are consistent) Covariance matrix The determinant (scalar) of the domain reflects the "volume" of the characteristic distribution of the domain; Exponential functions are used to map quadratic results to non-negative similarity values. The negative half of the Mahalanobis distance, used as a measurement scheme. With domain mean Normalized distance (eliminating the influence of dimensions); The residual vector is the difference between the feature vector and the domain mean vector of the scheme, with dimension 1. ; The transpose operator converts the residual vector from a column vector to a row vector. Covariance matrix The inverse matrix is ​​used to normalize the residuals and reflect the correlation between features; : No. The scheduling scheme to be optimized is in the first... The similarity weights across each domain satisfy the following conditions: and Used for weighted fusion of multi-domain information; : No. The sum of the similarity factors of each scheme with all domains is used as the normalized denominator to ensure that the total weights are 1.

[0182] Step S3.3.4: Weighted Optimization and Iterative Update

[0183] Based on similarity weight Calculate the weighted posterior mean of the target domain model. With variance The formula is as follows:

[0184]

[0185] : weighted posterior mean of target domain model, dimension , fusing target domain initial distribution and mean information of source domains as predicted mean of scheme ; : similarity weight of th scheme on target domain initial distribution; : mean vector of target domain initial distribution (dimension ); : sum of weighted mean of 1st to th source domain; : similarity weight of th scheme on th source domain; : mean vector of th source domain (dimension ); : weighted posterior variance of target domain model (scalar), measuring uncertainty of predicted mean , the larger the value, the more unstable the prediction; : similarity weight of th scheme on th domain; : variance of th domain (scalar), taking diagonal element of covariance matrix corresponding to predicted value of objective function); : square of mean vector of th domain (scalar, usually taking square sum of each dimension mean); : weighted mean square term, used to correct mean cross term in variance calculation, ensuring non-negative variance.

[0186] Step S3.3.5: maximize weighted expected improvement (EI)

[0187] Before the first iteration, generate an initial candidate scheme set in the feasible scheduling scheme space by Latin hypercube sampling method, and evaluate the objective function value to form an initial data set .

[0188] Take the weighted expected improvement (EI) as the acquisition function, and maximize the acquisition function :

[0189]

[0190] : under the current data set , scheme The expected improvement value is used to measure the potential improvement of the solution compared to the current best solution. The larger the value, the more worthwhile it is to explore. Charging scheduling schemes to be evaluated; : No. The dataset for the next iteration (containing the features of the evaluated schemes and the objective function values); :plan The weighted posterior mean (the predicted objective function value); Current optimal scheduling scheme The objective function value (e.g., minimum synthesis cost); Predict the amount of improvement; a positive result indicates a solution. It may be better than the current best; : The cumulative distribution function (CDF) of the standard normal distribution, outputting a probability value (range [0,1) representing the probability that "the improvement is positive"; :plan The weighted posterior standard deviation; Standardized improvement (following a standard normal distribution) ); : The probability density function (PDF) of the standard normal distribution, which outputs probability density values ​​to quantify the distributional concentration of the improvement quantity; : Represents the expected return when the improvement amount is positive; : Represents the potential benefits arising from forecast uncertainty, balancing "exploitation (exploiting the known optimal region)" and "exploration (exploring the region of high uncertainty)".

[0191] Specifically, in one embodiment, the acquisition function is maximized by combining Latin hypercube sampling with local gradient optimization. This allows for the selection of the next candidate solution to be evaluated. .

[0192] Each time the algorithm completes an iteration of "source domain fusion - similarity calculation - weighted optimization - EI maximization", it outputs one optimal candidate scheduling scheme for the current iteration. When the convergence condition is met (e.g., the change in the objective function value for 5 consecutive iterations is <0.1%), it outputs the global optimal scheduling scheme set, which is the final matching scheme for all users to be scheduled: covering all electric vehicles participating in the scheduling (e.g., 500 vehicles), and specifying the target charging station, charging time period, and charging power for each vehicle (e.g., "user C has a charging power of 30kW and is matched with charging pile No. 2 at station 2").

[0193] The embodiment of the application provides a kind of based on big data electric vehicle orderly charging guide optimization scheduling method, through LMC-Synth synthesis data enhancement, the problem of new scene data scarcity is solved, using TimePFN (TimePFN) It is realized that the accurate space-time prediction of charging load is realized, and PPTL-MGBO (PPTL-MGBO) algorithm is used in electric power-traffic double constraint to solve optimal scheduling scheme.The method effectively improves the load prediction accuracy and the global coordination of scheduling decision, significantly reduces the user charging cost, guarantees the safe and stable operation of distribution network, and relieves the pressure of urban traffic congestion.

[0194] Figure 5 It is a kind of module schematic view of based on big data electric vehicle orderly charging guide optimization scheduling system 200 provided by the embodiment of the application, the technical solution of the embodiment can be applicable to the case of electric vehicle orderly charging guide optimization scheduling, the system can be realized by software and / or hardware, and is generally integrated on any electronic device with network communication function, the electronic device includes but is not limited to: server, computer, personal digital assistant and the like equipment.As shown in Figure 5 The embodiment of the application provides a kind of based on big data electric vehicle orderly charging guide optimization scheduling system 200, for realizing the foregoing based on big data electric vehicle orderly charging guide optimization scheduling method 100, the system 200 includes:

[0195] Data acquisition and pretreatment module 210 is used to collect multi-source data and carry out cleaning, standardization and synthetic data enhancement processing;

[0196] Load prediction module 220 is used to divide user cluster and predict the space-time distribution characteristics of charging load;

[0197] Scheduling optimization module 230 is used to construct charging guide model under the constraint of electric power-traffic and solve optimal scheduling scheme.

[0198] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described module can refer to the corresponding process in the foregoing method embodiment, which will not be repeated here.

[0199] It should be understood that the steps can be reordered, added or deleted using the various forms of flow shown above.For example, each step described in the application can be executed in parallel, sequentially or in different order, as long as the desired results of the technical solution of the application can be achieved, which is not limited herein.

[0200] The above embodiments are merely exemplary and are not intended to limit the embodiments. Based on the above description, those skilled in the art can further make modifications and variations to the embodiments. The modifications and variations do not depart from the scope of the embodiments.

Claims

1. A method for optimizing and scheduling orderly charging of electric vehicles based on big data, characterized in that, Includes the following steps: Data on electric vehicles, traffic, electricity, and charging facilities are collected through a big data platform, and the data is preprocessed. Based on the user historical data in the preprocessed data, the K-Means clustering algorithm is used to divide electric vehicle users into different user clusters. Furthermore, by integrating user cluster characteristics, traffic flow data and power distribution network load data, the TimePFN model of the time series prior fitting network is used to predict the spatiotemporal distribution characteristics of charging load. A charging guidance model under power-traffic constraints is constructed. The predicted spatiotemporal distribution characteristics of charging load are used as input, and the goal is to minimize the user's comprehensive cost. Considering power constraints, traffic constraints, and charging facility constraints, the point-by-point transfer learning hybrid Gaussian Bayesian optimization (PPTL-MGBO) algorithm is used to solve the optimal charging scheduling scheme.

2. The method for orderly charging guidance and optimization scheduling of electric vehicles based on big data according to claim 1, characterized in that, in, Data preprocessing includes data cleaning and standardization, as well as synthetic data augmentation. Synthetic data augmentation includes using the LMC-Synth method to perform synthetic data augmentation on data that is scarce or imbalanced in real-world scenarios after data cleaning and standardization, thereby generating multivariate time series data.

3. The method for orderly charging guidance and optimization scheduling of electric vehicles based on big data according to claim 2, characterized in that, in, The synthetic data augmentation using the LMC-Synth method includes: The number of latent functions L is sampled from the Weibull distribution, ensuring that L is between the preset minimum number n and the total number of variables N; The Kernel-Synth method is used to generate L independent latent functions by randomly combining Gaussian process kernels; each latent function corresponds to a class of basic time series features, and the categories of the basic time series features include periodicity and trend. By using a linear co-regionalization model, L independent latent functions are linearly combined into a multivariate time series, ensuring that there is a true correlation between the variables.

4. The method for orderly charging guidance and optimization scheduling of electric vehicles based on big data according to claim 3, characterized in that, The Gaussian process kernel comprises an additive combination of linear and periodic kernels, and its expression is as follows: ； ； ； in, Linear kernel function, used to capture the linear trend characteristics of time series; Periodic kernel functions are used to capture the periodic characteristics of time series. It is an exponential function used to map the result of the kernel function calculation to the non-negative interval; The index of the first time point is used to locate a specific moment in the time series; The index of the second time point, and Together, we determine the time window for kernel function computation; Pi is used to calculate trigonometric function terms related to the period. For time points and The absolute difference reflects the time interval between two moments; The period length of the period kernel is the time period during which the time series repeats. In the charging load forecast, it is set to 24 hours and corresponds to a daily cycle. The width parameter of the periodic kernel controls the smoothness of the periodic features.

5. The method for orderly charging guidance and optimization scheduling of electric vehicles based on big data according to claim 4, characterized in that, The formula for linearly combining L independent latent functions into a multivariate time series using a linear co-regionalization model is as follows: ； in, : No. The variables at time... The value represents a specific dimension of data in the synthetic multivariate time series (MTS), including: the load of charging station A or the traffic flow of road segment B; : Variable index, value range is , used to distinguish different types of variables; : Time point index, representing a specific moment, used to locate the time position in a time series; : No. The variable for the first The weight coefficients of a latent function control the effect of that latent function on the variables. The degree of influence was obtained by sampling from the Dirichlet distribution; : No. A latent function at time t The value of is generated by the Kernel-Synth method and represents a class of basic time series features.

6. The method for orderly charging guidance and optimization scheduling of electric vehicles based on big data according to claim 5, characterized in that, in, The synthetic data augmentation using the LMC-Synth method also includes: supplementing independent variable scenarios by generating batches of synthetic data with independent variables to supplement real data samples, thereby simulating and covering low-correlation or irrelevant events in real-world scenarios. Specifically, this includes: The number of latent functions L is set to be equal to the total number of variables N in the multivariate time series; Weight coefficient matrix Configured so that when variable index i = latent function index j, When variable index i ≠ latent function index j, =0, making each variable =0, so that each variable Only one corresponding latent function They are generated independently, and there is no correlation between the variables; The proportion of the variable-independent synthetic data in the total generated data is set to be no less than 25%.

7. The method for orderly charging guidance and optimization scheduling of electric vehicles based on big data according to claim 1, characterized in that, The user's historical data includes charging frequency, single charging amount, and charging time preference; the different user clusters include commuter users, business users, and family users.

8. The method for orderly charging guidance and optimization scheduling of electric vehicles based on big data according to claim 1, characterized in that, The method of using the Time-Prior Fitting Network (TimePFN) model to predict the spatiotemporal distribution characteristics of charging load includes: Model input construction: The preprocessed multi-source data is transformed into a fixed-length time-series input window, i.e., the input matrix; One-dimensional convolution and residual connection are performed on each variable of the input matrix to extract local temporal features, including load increase trends and traffic congestion abrupt changes; The extracted local temporal features are segmented into overlapping patches and mapped to fixed-dimensional embedding vectors; The patch embeddings of all variables are channel-mixed and merged to obtain the total embedding matrix; The total embedding matrix is ​​transformed into a query (Q), key (K), and value (V) matrix through a linear layer, and then the output feature matrix is ​​obtained through multi-head attention calculation. The output feature matrix is ​​input into the Transformer encoder for iterative processing to capture cross-variable dependencies; The output of the Transformer encoder is reorganized by variables to generate the final spatiotemporal distribution prediction of the charging load.

9. The method for orderly charging guidance and optimization scheduling of electric vehicles based on big data according to claim 8, characterized in that, The formula for segmenting the extracted local temporal features into overlapping patches and mapping them to fixed-dimensional embedding vectors is as follows: ； Among them, OverlapExtract P is the overlapping patch extraction function; S is the patch size; PE is the step size; A 2-layer feed-forward network is used to map the flattened patch vectors to a fixed-dimensional embedding space; Flatten The flattening function converts a two-dimensional patch matrix into a one-dimensional vector, making it easier for feedforward networks to process. : No. Local temporal features of each variable after feature extraction; : No. The final set of embedded vectors after the variables are extracted by overlapping patches, flattened, mapped by feedforward networks, and encoded at their positions.

10. The method for orderly charging guidance and optimization scheduling of electric vehicles based on big data according to claim 8, characterized in that, The Transformer encoder adopts an 8-layer structure, with each layer containing a multi-head attention module, a feedforward network, and layer normalization. The multi-head attention calculation uses 8 attention heads.

11. The method for orderly charging guidance and optimization scheduling of electric vehicles based on big data according to claim 1, characterized in that, in, With the goal of minimizing overall user costs, the objective function F, considering constraints such as electricity, traffic, and charging infrastructure, is formulated as follows: ； in, The total number of electric vehicles to be dispatched; For users The unit price for charging is yuan / kWh; For users The charging demand is expressed in kWh. For users Travel time from the current location to the target charging station, in minutes; This refers to the traffic time weighting coefficient. For users Waiting time at the target charging station, in minutes; This is a waiting time weighting coefficient, which can be dynamically adjusted according to user needs and preferences.

12. The method for orderly charging guidance and optimization scheduling of electric vehicles based on big data according to claim 1, characterized in that, in, The optimal charging scheduling scheme is solved using the point-by-point transfer learning mixture Gaussian Bayes optimization (PPTL-MGBO) algorithm, including: Using historically similar scenarios as the source domain, multiple Gaussian process (GP) models were constructed and pre-trained. The initial data of the current scheduling scenario is used as the initial distribution of the target domain, and a mixture Gaussian distribution is constructed together with the models of each source domain. Calculate the similarity factor between each charging scheduling scheme to be optimized in the target domain and each domain, and obtain the similarity weight based on Bayesian rules; Based on the similarity weights, the weighted posterior mean and variance of the target domain model are calculated; We use the weighted expectation improvement EI as the acquisition function, and maximize the acquisition function to achieve iterative updates of the optimal scheduling scheme.

13. The method for orderly charging guidance and optimization scheduling of electric vehicles based on big data according to claim 12, characterized in that, in, The formula for constructing a Gaussian mixture distribution is: ； in, The probability density function of a mixture Gaussian distribution; The weight coefficient of the z-th domain; : The Gaussian probability density function of the z-th domain; This represents the number of source domains, i.e., the total number of similar historical scenarios.

14. The method for orderly charging guidance and optimization scheduling of electric vehicles based on big data according to claim 12, characterized in that, The formula for the weighted expected value improvement of EI is: ； in, In the current dataset The following is the plan Expected improvement value; Charging scheduling schemes to be evaluated; : No. The dataset for the next iteration; Φ The cumulative distribution function of the standard normal distribution. is the probability density function of the standard normal distribution; :plan The weighted posterior mean; :plan The weighted posterior standard deviation.

15. The method for orderly charging guidance and optimization scheduling of electric vehicles based on big data according to claim 1, characterized in that, in, The power constraints are: ensure that the load at each node does not exceed the rated capacity and the voltage deviation is within the allowable range through power flow calculations of the distribution network; The traffic constraints are: limiting the traffic flow on the road sections around the charging station to not exceed the congestion threshold, and quantifying the impact of road network congestion on travel time through the road impedance function; The constraints on charging facilities are: considering the power limitations and operating status of charging piles, ensuring that the scheduled charging demand does not exceed the real-time available capacity of the charging facilities.

16. A big data-based electric vehicle orderly charging guidance and optimization scheduling system, used to implement the big data-based electric vehicle orderly charging guidance and optimization scheduling method as described in any one of claims 1-15, characterized in that, The system includes: The data acquisition and preprocessing module is used to collect multi-source data and perform cleaning, standardization, and synthetic data enhancement processing. The load forecasting module is used to divide user clusters and predict the spatiotemporal distribution characteristics of charging load; The scheduling optimization module is used to build a charging guidance model and solve for the optimal scheduling scheme under the dual constraints of power and transportation.

Images