New energy vehicle charging network multi-source coupling dynamic optimization system and method

Abstract

This application relates to the field of supporting technology for new energy vehicle infrastructure, and discloses a multi-source coupling dynamic optimization system and method for new energy vehicle charging networks. This invention solves the problem of time misalignment between vehicle mobile data and grid fixed data by using spatiotemporal coupling of multi-source heterogeneous data and manifold completion technology. It realizes the extraction of user latent features and real-time quantification of carbon footprint by using semantic weight kernel principal component analysis and dynamic carbon flow model, breaking through the limitations of static accounting. It adopts hidden Markov model and Weiber distribution to correct the prediction bias of traditional queuing theory for long-tail behavior of private cars. Finally, based on two-level game and model predictive control, it achieves millisecond-level accurate allocation and collaborative optimization of charging power while balancing grid stability, low-carbon goals and user experience.

Description

Technical Field

[0001] This application relates to the field of supporting technologies for new energy vehicle infrastructure, and in particular to a multi-source coupling dynamic optimization system and method for new energy vehicle charging networks. Background Technology

[0002] As the global energy structure transitions towards a low-carbon model, the new energy vehicle industry has entered a critical stage of market expansion. In particular, the number of privately owned new energy charging vehicles has experienced explosive growth, making them a core node in the interaction between urban transportation networks and power energy networks. In the current smart city transportation system, charging infrastructure is no longer just a simple energy replenishment station, but a complex coupled system carrying the interaction of energy flow, information flow, and carbon flow. Large-scale new energy charging vehicles exhibit characteristics such as highly random travel trajectories, discrete spatiotemporal distribution of charging behavior, and significantly differentiated requirements for brand and service experience. This massive and highly heterogeneous charging demand poses severe challenges to the operational efficiency of charging networks, the stability of grid load, and the accuracy of carbon emission accounting.

[0003] However, existing charging network scheduling and management technologies have significant limitations when facing application scenarios dominated by private cars. The main problem is that most existing queuing time prediction and resource scheduling models rely on the traditional M / M / 1 queuing theory model. This model is based on the assumption that service time follows an exponential load distribution. This assumption is more applicable to high-frequency, short-duration ride-hailing or taxi operation scenarios, but for new energy private cars, their charging behavior has significant long-tail characteristics, exhibiting a complex pattern of low frequency, long charging time, charging at the store, and long-term parking. The probability density function of their actual charging time is more in line with the Weibull distribution than the exponential distribution. Continuing to use the traditional model will lead to a huge deviation in the estimation of queuing time, resulting in a situation where there is both traffic congestion during peak hours and idle resources during off-peak hours.

[0004] Furthermore, in terms of carbon emission management, existing technologies generally adopt carbon accounting methods based on static factors, which often separate the charging process from the vehicle's driving trajectory. This ignores the dynamic correlation between the vehicle's instantaneous displacement in the transportation network and the time-varying carbon emission factors on the grid side, making it impossible to accurately quantify the true carbon footprint of each vehicle throughout its entire life cycle. Meanwhile, in terms of scheduling and control architecture, traditional centralized optimization algorithms struggle to balance real-time performance and global optimality when faced with massive, dispersed charging requests due to the exponential explosion of computational dimensions. They also lack differentiated guidance mechanisms based on different user brand preferences and willingness to pay, making it difficult to effectively regulate grid load and carbon emission intensity through market mechanisms. Therefore, there is an urgent need to establish a new charging network management system that can deeply integrate multi-source heterogeneous data, accurately describe user behavior characteristics, and achieve collaborative optimization between charging stations and regions. Summary of the Invention

[0005] This application proposes a multi-source coupling dynamic optimization system and method for new energy vehicle charging networks to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this application adopts the following technical solution: a multi-source coupling dynamic optimization system for new energy vehicle charging networks, comprising: a data acquisition and standardization processing module, a multi-dimensional feature extraction and carbon flow coupling module, a user behavior classification and queuing estimation module, and a multi-agent collaborative scheduling module;

[0007] The data acquisition and standardization processing module is configured to collect driving trajectory data, brand preference data, payment willingness data, and real-time carbon emission factor data of new energy charging vehicles, and preprocess the above data to output a standardized multi-source heterogeneous dataset.

[0008] The multidimensional feature extraction and carbon flow coupling module is configured to receive standardized multi-source heterogeneous datasets, extract core behavioral feature vectors using an improved kernel principal component analysis algorithm, and generate full life cycle carbon footprint data by combining the core behavioral feature vectors and real-time carbon emission factor data of the power grid through a dynamic carbon flow coupling model calibrated by a genetic algorithm. The core behavioral feature vectors are then transmitted to the user behavior classification and queuing estimation module, and the full life cycle carbon footprint data is transmitted to the multi-agent collaborative scheduling module.

[0009] The user behavior classification and queuing estimation module is configured to receive core behavior feature vectors, run a hidden Markov model to calculate user type weight coefficients, call an improved queuing theory model based on user type weight coefficients, use the Weibull distribution function to fit the probability density of charging time to replace the exponential distribution assumption, and calculate and output the predicted value of queuing waiting time.

[0010] The multi-agent collaborative scheduling module is configured to receive full lifecycle carbon footprint data and queuing time predictions, solve for the Nash equilibrium point based on a two-layer game architecture, and generate the optimal charging power allocation strategy.

[0011] Furthermore, the data acquisition and standardization module performs preprocessing on the data to output a standardized multi-source heterogeneous dataset. The specific operations are as follows:

[0012] A1 executes the heterogeneous data dynamic spatiotemporal anchor point synchronization mapping program, calls the timestamp sequence in the driving trajectory data of new energy charging vehicles as the reference clock, establishes a dynamic spatiotemporal anchor point mapping field, obtains real-time carbon emission factor data of the power grid, performs time decay operation based on exponential function on the real-time carbon emission factor data of the power grid, generates carbon emission factors with time weight, and then performs tensor product expansion operation to map the carbon emission factors with time weight to the vector space of driving trajectory data, generating a preliminary synchronization spatiotemporal matrix;

[0013] A2 executes a manifold completion program based on topological graph Laplacian regularization to obtain the preliminary synchronized spatiotemporal matrix and pre-stored road network topology data. It constructs a regularization objective function containing the graph Laplacian matrix, which consists of a data fidelity term and a manifold constraint term. Iterative minimization is performed on the regularization objective function to calculate the Frobenius norm between the preliminary synchronized spatiotemporal matrix and the ideal complete data matrix to be solved. At the same time, it calculates the trace of the transpose of the ideal complete data matrix, the graph Laplacian matrix, and the ideal complete data matrix. The ideal complete data matrix is ​​updated using the gradient descent algorithm until convergence, and the topological complete reconstruction matrix is ​​output.

[0014] Furthermore, the data acquisition and standardization module's preprocessing of data to output a standardized multi-source heterogeneous dataset includes the following specific operations:

[0015] A3 executes a semantically driven hybrid dimensional nonlinear modulation program to obtain a topologically complete reconstruction matrix, brand preference data, and willingness-to-pay data. It constructs a brand feature mapping matrix, transforms discrete brand preference data into a continuous semantic potential function, performs dot product weighting, modulates the amplitude of the willingness-to-pay data using the semantic potential function, and generates a semantically weighted payment vector. Subsequently, it performs hyperbolic tangent nonlinear mapping on the topologically complete reconstruction matrix and the semantically weighted payment vector to compress the physical values ​​to a preset normalization interval. It also performs robust scaling based on the median and interquartile range statistics within a local sliding window to generate a semantically hybrid normalized tensor.

[0016] A4 executes a kernel-space-oriented feature tensor stacking and pre-mapping procedure, calculates the covariance matrix of the semantic mixture normalized tensor, determines the bandwidth parameter of the Gaussian kernel function based on the distribution characteristics of the covariance matrix, performs tensor concatenation operation on the semantic mixture normalized tensors of different vehicles along the sample dimension to construct a global feature matrix, performs eigenvalue decomposition operation on the kernel matrix corresponding to the global feature matrix to generate an eigenvalue diagonal matrix, performs inverse square root operation on the eigenvalue diagonal matrix to generate a whitening transformation matrix, multiplies the global feature matrix with the whitening transformation matrix, and outputs a normalized multi-source heterogeneous dataset.

[0017] Furthermore, the specific operation of the multidimensional feature extraction and carbon flow coupling module in extracting core behavioral feature vectors using the improved kernel principal component analysis algorithm is as follows:

[0018] B1 executes a kernel matrix evolution construction procedure based on semantic weights. The multidimensional feature extraction and carbon flow coupling module receives the standardized multi-source heterogeneous dataset output by the data fusion and feature extraction module. It separates the brand preference component and the willingness to pay level component from the standardized multi-source heterogeneous dataset. It performs vector synthesis operation on the brand preference component and the willingness to pay level component to generate semantic weight factors. It calculates the square of the Euclidean distance between any two sample points in the standardized multi-source heterogeneous dataset in the regenerating kernel Hilbert space. It divides the square of the Euclidean distance by twice the square of the preset kernel bandwidth parameter and takes the negative number. It performs an exponential operation on the calculation result to generate the basic Gaussian kernel function value. At the same time, it performs an inner product operation on the semantic weight factors of the two sample points. It performs a hyperbolic tangent nonlinear mapping operation on the result of the inner product operation and adds one to generate the semantic gain coefficient. It performs a multiplication operation on the basic Gaussian kernel function value and the semantic gain coefficient to generate the matrix elements of the semantic weighted kernel evolution matrix.

[0019] B2, then executes the nonlinear projection extraction procedure for the core behavioral feature vector, performs centering processing and eigenvalue decomposition operation on the semantic weighted kernel evolution matrix, selects several feature vectors whose cumulative contribution rate meets the preset threshold in descending order of eigenvalues, projects the standardized multi-source heterogeneous dataset onto the low-dimensional feature subspace spanned by several feature vectors, generates the projected feature matrix, and decomposes the projected feature matrix along the feature dimension into charging demand density component, brand preference component, and low carbon potential value component, and combines the charging demand density component, brand preference component, and low carbon potential value component to output the core behavioral feature vector.

[0020] Furthermore, the multidimensional feature extraction and carbon flow coupling module, combining core behavioral feature vectors and real-time carbon emission factor data from the power grid, generates full life-cycle carbon footprint data through a dynamic carbon flow coupling model calibrated by a genetic algorithm, specifically as follows:

[0021] B3 executes a three-dimensional coupled dynamic carbon flow differential modeling program to obtain the charging demand density component and low-carbon potential value component in the core behavioral feature vector. At the same time, it obtains real-time carbon emission factor data of the power grid and establishes a dynamic carbon flow coupled differential equation describing the change of carbon emission rate over time. The dynamic carbon flow coupled differential equation is composed of the linear superposition of energy source term, traffic source term, and carbon sink offset term. The multidimensional feature extraction and carbon flow coupling module generates the energy source term by performing a multiplication operation on the real-time carbon emission factor data of the power grid and the charging demand density component. It generates the traffic source term by performing a continuous multiplication operation on the basic factor of the private car-specific carbon emission coefficient to be calibrated, the road network traffic flow density data after nonlinear power function transformation, and the vehicle mileage data. It generates the carbon sink offset term by performing a multiplication operation on the preset carbon sink coefficient and the low-carbon potential value component.

[0022] B4. Execute a closed-loop parameter calibration and integration procedure based on a genetic algorithm to construct an optimization objective function. The optimization objective function is configured to calculate the cumulative absolute error between the simulated carbon footprint data generated by the dynamic carbon flow coupling differential equation and the historical measured carbon footprint data. Run the genetic algorithm to perform iterative optimization operations of selection, crossover, and mutation on the basic factors of the private car-specific carbon emission coefficient to be calibrated until the value of the optimization objective function converges to the preset minimum value. Output the optimal private car-specific carbon emission coefficient. Substitute the optimal private car-specific carbon emission coefficient into the dynamic carbon flow coupling differential equation and perform definite integral operations on the dynamic carbon flow coupling differential equation in the entire life cycle time domain to output the full life cycle carbon footprint data.

[0023] Furthermore, the specific operation of the user behavior classification and queuing estimation module in calculating the user type weight coefficients using a hidden Markov model is as follows:

[0024] C1 executes the user hidden state fuzzy decoding program. The user behavior classification and queuing estimation module obtains the core behavior feature vector output by the multi-dimensional feature extraction and carbon flow coupling module. In the hidden Markov model, a hidden state space containing home commuting state, long-distance travel state, and local transportation state is constructed. The core behavior feature vector is mapped to the hidden state space. The home commuting state is configured to represent the physical state of the vehicle charging during the morning and evening peak hours on weekdays and the remaining mileage data is greater than 50 kilometers. The long-distance travel state is configured to represent the physical state of the vehicle charging more than 80% of the total battery capacity in a single charge and the geographical coordinates being located around the highway network. The local transportation state is configured to represent the physical state of the vehicle having less than 30 kilometers of remaining mileage data and requesting slow charging service.

[0025] C2, the user behavior classification and queuing estimation module performs Gaussian mixture model fitting operation, calculates the observation probability density value of the core behavior feature vector in each state in the hidden state space, and calls the forward and backward iterative algorithm to calculate the posterior probability value of the current vehicle observation sequence in the home commuting state, long-distance travel state, and local transportation state respectively. The user behavior classification and queuing estimation module performs normalization operation on the posterior probability value, generates three normalized probability components corresponding to the above three states respectively, and combines the three normalized probability components to output the user type weight coefficient.

[0026] Furthermore, the user behavior classification and queuing estimation module calls an improved queuing theory model based on user type weight coefficients, and uses the Weibull distribution function to fit the probability density of charging time to replace the exponential distribution assumption. The specific operation of calculating and outputting the predicted queuing waiting time is as follows:

[0027] C3 executes the dynamic nonlinear synthesis program of distribution parameters. The user behavior classification and queuing estimation module obtains the user type weight coefficients. At the same time, it calls the pre-stored baseline scale parameters corresponding to the three user types respectively. It uses the user type weight coefficients as the exponent to perform a weighted geometric mean operation on the baseline scale parameters to generate the current vehicle-specific synthetic scale parameters. The value of the synthetic shape parameter of the Weiber distribution function is fixed at 1.8.

[0028] C4 executes an improved queuing time calculation program based on the second-order central moment. The user behavior classification and queuing estimation module calls the gamma function to calculate the second-order central moment of the Weiber distribution function. The specific calculation logic is to calculate the gamma function value of the independent variable as one plus twice the reciprocal of the composite shape parameter, subtract the square of the gamma function value of the independent variable as one plus twice the reciprocal of the composite shape parameter, and multiply the result of the subtraction operation by the square of the composite scale parameter to generate the variance of the charging service duration.

[0029] C5, the user behavior classification and queuing estimation module, obtains the current average vehicle arrival rate data and system service intensity data. It performs calculations using the extended Polachek-Shinchin formula, multiplying the square of the system service intensity data by the square of the product of the average vehicle arrival rate data and the composite scale parameter by the variance coefficient, and then dividing by twice the average vehicle arrival rate data multiplied by one minus the difference in system service intensity data to obtain the physical average waiting time under steady-state conditions. A queuing patience correction factor is then introduced to perform a multiplicative correction operation on the physical average waiting time, and the predicted queuing waiting time value is output.

[0030] Furthermore, the specific operations of the multi-agent cooperative scheduling module in solving for the Nash equilibrium point based on the two-layer game architecture are as follows:

[0031] D1 executes a multi-objective coupled heterogeneous utility functional construction procedure. The multi-agent collaborative scheduling module obtains the full life cycle carbon footprint data output by the multi-dimensional feature extraction and carbon flow coupling module, and simultaneously obtains the queuing wait time prediction value output by the user behavior classification and queuing estimation module. It constructs a generalized heterogeneous utility functional for each site-level agent participating in the game. The generalized heterogeneous utility functional is configured as a linear superposition combination of the charging service logarithmic gain term, the full life cycle carbon emission index penalty term, and the queuing time nonlinear loss term. The multi-agent collaborative scheduling module performs a natural logarithmic operation on the function with the optimal charging power allocation strategy to be solved as the independent variable and multiplies it by a preset gain sensitivity coefficient to generate the charging service logarithmic gain term. It performs a natural exponential operation on the full life cycle carbon footprint data and multiplies it by a preset carbon sensitivity coefficient to generate the full life cycle carbon emission index penalty term. It performs a power operation on the queuing wait time prediction value and multiplies it by a preset time sensitivity coefficient to generate the queuing time nonlinear loss term.

[0032] D2 executes a dynamic interaction program based on a two-layer Stackelberg game architecture. The multi-agent collaborative scheduling module establishes a game topology structure that includes upper-level regional agents and lower-level station-level agents. Regional agents generate shadow electricity price signals based on the total load constraints of the regional power grid. Station-level agents receive shadow electricity price signals and adjust their own power demand with the goal of maximizing the generalized heterogeneous utility functional. The multi-agent collaborative scheduling module constructs a game-coupled variational inequality that describes the coupling relationship between the upper-level regional agents and the lower-level station-level agents. This variational inequality characterizes the zero-sum equilibrium state of the utility gradients of each agent under the premise of satisfying the physical constraints of the power grid.

[0033] Furthermore, the specific operations of the multi-agent cooperative scheduling module in generating the optimal charging power allocation strategy are as follows:

[0034] D3 executes the Nash equilibrium solution program based on distributed momentum gradient. The multi-agent cooperative scheduling module uses a distributed projection gradient algorithm to iteratively solve the game-coupled variational inequality. In each iteration cycle, each station-level agent only communicates with its neighboring nodes in the topology to exchange power state information. The multi-agent cooperative scheduling module calculates the first-order gradient value of the generalized heterogeneous utility functional with respect to the optimal charging power allocation strategy. It introduces a momentum acceleration factor to perform inertial correction on the power increment generated in the previous iteration cycle. It updates the current power strategy by combining the first-order gradient value with the consensus protocol data of the neighboring nodes, and uses the projection operator to map the updated power strategy back to the allowed physical power range. The above iterative steps are repeated until the power strategies of all station-level agents converge to a steady state to obtain the Nash equilibrium point.

[0035] D4 executes the restricted model predictive control physical execution program. The multi-agent cooperative scheduling module sets the Nash equilibrium point as the reference trajectory and constructs a rolling optimization objective function. The rolling optimization objective function is configured to minimize the weighted sum of the square of the tracking error between the actual output power and the reference trajectory and the square of the power change rate. The multi-agent cooperative scheduling module introduces a power change rate constraint to limit the ramp-up rate of the charging pile power. Under the premise of satisfying the power change rate constraint, it solves the control sequence in the future finite prediction time domain, selects the first time step element in the control sequence as the physical control command, and outputs the final optimal charging power allocation strategy.

[0036] A dynamic optimization method for multi-source coupling in new energy vehicle charging networks includes the following steps:

[0037] Step S1: Collect driving trajectory data, brand preference data, payment willingness data and real-time carbon emission factor data of new energy charging vehicles, and preprocess the above data to output a standardized multi-source heterogeneous dataset.

[0038] Step S2: Receive a standardized multi-source heterogeneous dataset, extract core behavioral feature vectors using an improved kernel principal component analysis algorithm, and combine the core behavioral feature vectors with real-time carbon emission factor data from the power grid. Generate full life cycle carbon footprint data using a dynamic carbon flow coupling model calibrated by a genetic algorithm, and transmit the core behavioral feature vectors to step S3 and the full life cycle carbon footprint data to step S4.

[0039] Step S3: Receive the core behavior feature vector, run the hidden Markov model to calculate the user type weight coefficient, call the improved queuing theory model based on the user type weight coefficient, use the Weibull distribution function to fit the probability density of charging time to replace the exponential distribution assumption, and calculate and output the predicted value of queuing waiting time.

[0040] Step S4: Receive full lifecycle carbon footprint data and queuing time prediction values, solve the Nash equilibrium point based on a two-layer game architecture, and generate the optimal charging power allocation strategy.

[0041] The beneficial effects of this invention are as follows:

[0042] This invention solves the problem of time misalignment between vehicle mobile data and grid stationary data by using spatiotemporal coupling of multi-source heterogeneous data and manifold completion technology. It achieves user latent feature extraction and real-time carbon footprint quantification by using semantic weighted principal component analysis and dynamic carbon flow model, breaking through the limitations of static accounting. It also corrects the prediction bias of traditional queuing theory for long-tail behavior of private cars by using hidden Markov model and Weiber distribution. Finally, based on two-level game theory and model predictive control, it achieves millisecond-level precise allocation and collaborative optimization of charging power while balancing grid stability, low-carbon goals and user experience. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort:

[0044] Figure 1 This is a system framework diagram of the present invention;

[0045] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Example 1

[0048] like Figure 1 As shown, this invention discloses a multi-source coupling dynamic optimization system for new energy vehicle charging networks, comprising: a data acquisition and standardization processing module, a multi-dimensional feature extraction and carbon flow coupling module, a user behavior classification and queuing estimation module, and a multi-agent collaborative scheduling module, wherein;

[0049] In a specific embodiment of the present invention, the data acquisition and standardization processing module is configured to acquire driving trajectory data, brand preference data, payment willingness data and real-time carbon emission factor data of new energy charging vehicles, and preprocess the above data to output a standardized multi-source heterogeneous dataset.

[0050] It is important to note that before data collection begins, the system will obtain the user's explicit, voluntary, and fully informed consent to the processing of their personal information through a clear user agreement and privacy policy. This authorization process strictly adheres to the principles of legality, legitimacy, necessity, and good faith stipulated in the Personal Information Protection Law, ensuring that the purpose, method, and scope of information processing are transparent to the user. In particular, for sensitive information such as driving trajectories that can reflect the patterns of personal activities, their collection and use will be based on the user's separate consent and will be ensured to be carried out only to the extent necessary to achieve the purpose of the service.

[0051] The data acquisition and standardization processing module, as a programmable processing unit configured on the server side, is configured to execute the following rigorous data preprocessing logic to ensure that the spatiotemporal reference and physical dimensions of multi-source heterogeneous data are unified before being input into the core algorithm.

[0052] The data acquisition and standardization processing module performs the following preprocessing operations to output a standardized multi-source heterogeneous dataset: It executes a dynamic spatiotemporal anchor point synchronization mapping program for heterogeneous data, uses the timestamp sequence in the driving trajectory data of new energy charging vehicles as a reference clock, establishes a dynamic spatiotemporal anchor point mapping field, acquires real-time carbon emission factor data of the power grid, performs time-effect decay calculation based on an exponential function on the real-time carbon emission factor data of the power grid to generate carbon emission factors with time weights, and then performs tensor product expansion operation to map the carbon emission factors with time weights into the vector space of the driving trajectory data, generating a preliminary synchronization spatiotemporal matrix.

[0053] In practice, the data acquisition and standardization processing module first locks the timestamp of each discrete trajectory point in the driving trajectory data, and uses this as an anchor point to retrieve the most recently updated time slice data in the real-time carbon emission factor data of the power grid.

[0054] The data acquisition and standardization processing module introduces an information time-decrease operator to perform weighted calculations on the retrieved real-time carbon emission factor data of the power grid.

[0055] The information timeliness decay operator characterizes the effectiveness weight of power grid data as it decreases over time. In this embodiment, the numerical range of the information timeliness decay operator is set to 0.1 to 0.3. The basis for setting this numerical range is that the power grid load scheduling cycle is usually five to fifteen minutes. Taking 0.1 to 0.3 can ensure that the weight curve decays to less than 10% within two scheduling cycles, thereby filtering out outdated power grid status information.

[0056] The specific logic of the data acquisition and standardization processing module for performing the time-effect decay calculation based on the exponential function is as follows: calculate the negative exponent of the natural constant, where the exponent term is the product of the information time-effect decay operator and the time difference, and then multiply the calculation result by the original real-time carbon emission factor data of the power grid.

[0057] Subsequently, the data acquisition and standardization processing module performs tensor product expansion operation, namely Kronecker product operation, which multiplies the scalar carbon emission factor with time weights and the multidimensional driving trajectory data vector by dimension expansion, so that the real-time carbon emission factor data of the power grid is aligned with the driving trajectory data in terms of dimension, and finally generates a preliminary synchronous spatiotemporal matrix containing spatiotemporal coupling characteristics.

[0058] Following the above steps, the data acquisition and standardization module executes a manifold completion procedure based on topological graph Laplacian regularization to obtain the preliminary synchronized spatiotemporal matrix and pre-stored road network topology data. It then constructs a regularization objective function containing the graph Laplacian matrix, which consists of a data fidelity term and a manifold constraint term. An iterative minimization operation is performed on the regularization objective function, calculating the Frobenius norm between the preliminary synchronized spatiotemporal matrix and the ideal complete data matrix to be solved. Simultaneously, the trace of the transpose of the ideal complete data matrix, the graph Laplacian matrix, and the ideal complete data matrix is ​​calculated. The ideal complete data matrix is ​​then updated using a gradient descent algorithm until convergence, outputting the topologically complete reconstruction matrix.

[0059] In practice, the data acquisition and standardization processing module first loads the road network topology data and converts it into an adjacency matrix, and then calculates the difference between the degree matrix and the adjacency matrix to generate the graph Laplace matrix.

[0060] The graph Laplacian matrix in the algorithm represents the connectivity constraints of the physical road network. When the data acquisition and standardization processing module constructs the regularization objective function, a regularization smoothing factor is introduced to adjust the weight ratio between the data fidelity term and the manifold constraint term.

[0061] In this embodiment, the regularization smoothing factor is set to a value of 0.01 to 0.05. The basis for setting this value is the average connectivity statistics of urban road network nodes. This value range can ensure that while repairing the missing trajectory values, the repaired trajectory points are forced to fall on the legal road connectivity map.

[0062] The data acquisition and standardization module performs the iterative minimization solution operation as follows: it uses the gradient descent method to calculate the partial derivative of the regularization objective function with respect to the ideal complete data matrix, and updates the element values ​​of the ideal complete data matrix along the negative gradient direction.

[0063] When the difference between the regularization objective function values ​​obtained from two consecutive iterations is less than the preset convergence threshold (e.g., 10 to the power of negative 6), the data acquisition and standardization processing module stops iterating and outputs the finally converged topological complete reconstruction matrix.

[0064] The data acquisition and standardization module preprocesses data to output a standardized multi-source heterogeneous dataset. Specific operations include: executing a semantically driven hybrid dimensional nonlinear modulation program to obtain a topologically complete reconstruction matrix, brand preference data, and willingness-to-pay data; constructing a brand feature mapping matrix; converting discrete brand preference data into a continuous semantic potential function; performing dot-multiplication weighting operations; using the semantic potential function to modulate the amplitude of the willingness-to-pay data to generate a semantically weighted payment vector; subsequently performing a hyperbolic tangent nonlinear mapping operation on the topologically complete reconstruction matrix and the semantically weighted payment vector to compress physical values ​​to a preset normalization interval; and performing robust scaling operations based on the median and interquartile range statistics within a local sliding window to generate a semantically hybrid standardized tensor.

[0065] In practice, the data collection and standardization processing module first constructs a brand feature mapping matrix by querying a pre-set brand value database.

[0066] The brand feature mapping matrix contains a set of brand premium weight coefficients, which represent the sensitivity of different brand users to the value of time cost.

[0067] In this embodiment, the brand premium weighting coefficient is set to 1.2 to 1.5 for high-end brands and 0.8 to 1.0 for economy brands. This value is based on the regression analysis results of the historical charging order payment amounts of car owners of different brands.

[0068] The data collection and standardization module uses brand premium weighting coefficients to multiply and weight the willingness to pay data, so that the willingness to pay of users of high-premium brands gains an advantage in the model.

[0069] After weighting, the data acquisition and standardization module applies the hyperbolic tangent function to map all data to a dimensionless interval from negative one to positive one. To eliminate the interference of abnormal payment amounts, the data acquisition and standardization module performs robust scaling operations. Specifically, it sets a local sliding window with a length of thirty to sixty sampling points, calculates the median and interquartile range of the data within the window, and generates a semantic hybrid standardization tensor that is resistant to outliers by subtracting the median from the observed value and then dividing by the interquartile range.

[0070] Finally, the data acquisition and standardization processing module executes a kernel-space-oriented feature tensor stacking and pre-mapping procedure to calculate the covariance matrix of the semantic mixture standardization tensor. Based on the distribution characteristics of the covariance matrix, the bandwidth parameter of the Gaussian kernel function is determined. Tensor concatenation operations are performed on the semantic mixture standardization tensors of different vehicles along the sample dimension to construct a global feature matrix. Eigenvalue decomposition is performed on the kernel matrix corresponding to the global feature matrix to generate an eigenvalue diagonal matrix. The inverse square root operation is performed on the eigenvalue diagonal matrix to generate a whitening transformation matrix. The global feature matrix is ​​multiplied by the whitening transformation matrix to output a standardized multi-source heterogeneous dataset.

[0071] In practice, the data acquisition and standardization module first traverses the semantic mixture standardization tensor to calculate the covariance between each feature dimension, and sets the bandwidth parameter of the Gaussian kernel function based on the average Euclidean distance between sample points.

[0072] In this embodiment, the bandwidth parameter is set to 0.5 to 1 times the average Euclidean distance. This value follows the bandwidth selection criterion in statistical learning theory and aims to control the smoothness of the feature mapping to prevent overfitting.

[0073] Subsequently, the data acquisition and standardization processing module performs tensor concatenation operations to concatenate the spatiotemporal feature tensors of different vehicles into a high-dimensional global feature matrix.

[0074] To eliminate the linear correlation between features, the data acquisition and standardization processing module performs eigenvalue decomposition on the kernel matrix after mapping the global feature matrix, and extracts the eigenvalue sequence to form an eigenvalue diagonal matrix.

[0075] The data acquisition and standardization processing module performs the square root operation on the reciprocal of each diagonal element in the eigenvalue diagonal matrix to generate a whitening transformation matrix.

[0076] Finally, the data acquisition and standardization module multiplies the global feature matrix by the whitening transformation matrix to output an isotropic and decorrelated standardized multi-source heterogeneous dataset, which can be directly called by subsequent kernel principal component analysis algorithms.

[0077] In a specific embodiment of the present invention, the multidimensional feature extraction and carbon flow coupling module is configured to receive a standardized multi-source heterogeneous dataset, extract core behavioral feature vectors using an improved kernel principal component analysis algorithm, and simultaneously combine the core behavioral feature vectors and real-time carbon emission factor data of the power grid to generate full life cycle carbon footprint data through a dynamic carbon flow coupling model calibrated by a genetic algorithm. The core behavioral feature vectors are then transmitted to the user behavior classification and queuing estimation module, and the full life cycle carbon footprint data is transmitted to the multi-agent collaborative scheduling module.

[0078] The multidimensional feature extraction and carbon flow coupling module, as a logical operation unit configured on the server side, is configured to strictly follow the steps below to execute the computational process from high-dimensional data feature extraction to physical entity carbon emission modeling.

[0079] The multidimensional feature extraction and carbon flow coupling module uses an improved kernel principal component analysis algorithm to extract core behavioral feature vectors. The specific operations are as follows: A kernel matrix evolution construction program based on semantic weights is executed. The multidimensional feature extraction and carbon flow coupling module receives a standardized multi-source heterogeneous dataset output by the data fusion and feature extraction module. Brand preference and willingness-to-pay components are separated from the standardized multi-source heterogeneous dataset. Vector synthesis operations are performed on the brand preference and willingness-to-pay components to generate semantic weight factors. The square of the Euclidean distance between any two sample points in the standardized multi-source heterogeneous dataset is calculated in the regenerating kernel Hilbert space. The square of the Euclidean distance is divided by the square of twice the preset kernel bandwidth parameter and the result is negative. An exponential operation is performed on the calculation result to generate the basic Gaussian kernel function value. Simultaneously, an inner product operation is performed on the semantic weight factors of the two sample points. The result of the inner product operation is subjected to a hyperbolic tangent nonlinear mapping operation and incremented by one to generate a semantic gain coefficient. The basic Gaussian kernel function value and the semantic gain coefficient are multiplied to generate the matrix elements of the semantically weighted kernel evolution matrix.

[0080] In practice, the multidimensional feature extraction and carbon flow coupling module first identifies the feature dimensions corresponding to the brand and payment amount in the standardized multi-source heterogeneous dataset, and generates semantic weight factors through a weighted linear combination.

[0081] The semantic weight factor represents the value density of a user in the charging network, characterizing the user's sensitivity to time costs. When calculating the basic Gaussian kernel function value, the multidimensional feature extraction and carbon flow coupling module calls the preset kernel bandwidth parameter.

[0082] The preset kernel bandwidth parameter is the smoothing radius of sample similarity in the feature space.

[0083] In this embodiment, the preset kernel bandwidth parameter is set to a value of 0.5 to 1 times the average Euclidean distance of all sample points. The basis for setting this range is the cross-validation results in statistical learning. The purpose of this range is to control the sensitivity of the nonlinear mapping and prevent the model from overfitting to noise due to excessively narrow bandwidth.

[0084] When generating semantic gain coefficients, the multidimensional feature extraction and carbon flow coupling module uses the hyperbolic tangent function to map the inner product result of the semantic weight factors to the gain interval of zero to two.

[0085] The semantic gain coefficient is used to modulate the amplitude of the basic Gaussian kernel function value calculated from the physical trajectory, so that the characteristics of users with high brand value are numerically amplified in the semantic weighted kernel evolution matrix.

[0086] Following the above steps, a nonlinear projection extraction procedure for the core behavioral feature vector is then executed. The semantically weighted kernel evolution matrix is ​​centered and eigenvalue decomposition is performed. Several feature vectors whose cumulative contribution rate meets the preset threshold are selected in descending order of eigenvalues. The standardized multi-source heterogeneous dataset is projected onto a low-dimensional feature subspace spanned by several feature vectors to generate the projected feature matrix. The projected feature matrix is ​​then decomposed along the feature dimension into charging demand density components, brand preference components, and low-carbon potential value components. The charging demand density components, brand preference components, and low-carbon potential value components are combined and output as the core behavioral feature vector.

[0087] In practice, the multidimensional feature extraction and carbon flow coupling module truncates and filters feature vectors based on a preset threshold for the cumulative contribution rate.

[0088] The cumulative contribution rate preset threshold is the proportion of the original information retained in the dimensionality-reduced data. In this embodiment, the cumulative contribution rate preset threshold is set to 96%. The basis for setting this value is the position of the feature value descent inflection point determined by scree plot analysis. The purpose of setting this value is to filter out the last 4% of noise interference and ensure the purity of feature extraction.

[0089] The multidimensional feature extraction and carbon flow coupling module decouples the projected data into three orthogonal physical components: the charging demand density component represents the user's time urgency and charging frequency; the brand preference component represents the user's sensitivity to service quality and price; and the low-carbon potential component represents the user's likelihood of participating in the grid's peak shaving and valley filling response.

[0090] The multidimensional feature extraction and carbon flow coupling module combines core behavioral feature vectors and real-time carbon emission factor data from the power grid. The specific operations for generating full lifecycle carbon footprint data through a dynamic carbon flow coupling model calibrated by a genetic algorithm are as follows: A ternary coupled dynamic carbon flow differential modeling program is executed to obtain the charging demand density component and low-carbon potential value component from the core behavioral feature vector. Simultaneously, real-time carbon emission factor data from the power grid is acquired, and a dynamic carbon flow coupling differential equation describing the change of carbon emission rate over time is established. This dynamic carbon flow coupling differential equation consists of a linear superposition of energy source terms, traffic source terms, and carbon sink offset terms. The multidimensional feature extraction and carbon flow coupling module generates energy source terms by performing multiplication operations on real-time carbon emission factor data from the power grid and the charging demand density component; generates traffic source terms by performing continuous multiplication operations on the basic factor of the private car-specific carbon emission coefficient to be calibrated, the road network traffic flow density data transformed by a nonlinear power function, and the vehicle mileage data; and generates carbon sink offset terms by performing multiplication operations on the preset carbon sink coefficient and the low-carbon potential value component.

[0091] In practice, during the calculation of traffic source terms, the multidimensional feature extraction and carbon flow coupling module calls the basic factor of the private car-specific carbon emission coefficient to be calibrated.

[0092] The physical meaning of the basic factor of the private car-specific carbon emission coefficient to be calibrated is the benchmark value of the indirect carbon emission intensity caused by traffic congestion per unit mileage of new energy private cars.

[0093] In the initial settings, the value range of the basic factor of the carbon emission coefficient for private cars to be calibrated is set to 0.0010 to 0.0015 kg per kilometer. The basis for setting this value range is the energy consumption and emission bench test data of typical private cars under the working conditions in the vehicle engineering laboratory. The role of this factor is to serve as an adjustable benchmark for the model, and to adapt to the actual emission characteristics of different models through subsequent calibration.

[0094] The multidimensional feature extraction and carbon flow coupling module performs a nonlinear power function transformation with an exponent of 0.28 on the road network traffic flow density data. The value of this exponent is derived from the empirical formula of private cars’ sensitivity to congestion in urban traffic flow theory. The role of this exponent is to simulate the physical phenomenon that carbon emissions increase nonlinearly when traffic congestion intensifies.

[0095] For carbon sink offsetting items, the multidimensional feature extraction and carbon flow coupling module introduces a preset carbon sink coefficient. In this embodiment, the preset carbon sink coefficient is set to a carbon emission reduction of 0.05 to 0.1 kg per unit of low carbon potential value. The basis for setting this coefficient is the carbon emission reduction conversion standard of the grid side for off-peak charging behavior.

[0096] Finally, a closed-loop parameter calibration and integration procedure based on a genetic algorithm is executed to construct an optimization objective function. The optimization objective function is configured to calculate the cumulative absolute error between the simulated carbon footprint data generated by the dynamic carbon flow coupling differential equation and the historical measured carbon footprint data. The genetic algorithm is run to perform iterative optimization operations of selection, crossover, and mutation on the basic factors of the private car-specific carbon emission coefficient to be calibrated until the value of the optimization objective function converges to the preset minimum value. The optimal private car-specific carbon emission coefficient is output. The optimal private car-specific carbon emission coefficient is substituted into the dynamic carbon flow coupling differential equation, and a definite integral operation is performed on the dynamic carbon flow coupling differential equation in the entire life cycle time domain to output the full life cycle carbon footprint data.

[0097] In practice, the multidimensional feature extraction and carbon flow coupling module is configured with the following parameters for the genetic algorithm: the population size is set to fifty to one hundred individuals, and the maximum number of iterations is set to two hundred generations. The basis for setting these parameters is to balance the consumption of computing resources with the ability to find global optimization.

[0098] The multidimensional feature extraction and carbon flow coupling module determines that the objective function converges when the function value is lower than the preset minimum value. In this embodiment, the preset minimum value is set to 10 to the power of negative 4. The physical meaning of setting this value is that when the simulation error is lower than this threshold, it is determined that the model parameters have approximated the real physical laws.

[0099] The time domain of the definite integral operation performed by the multidimensional feature extraction and carbon flow coupling module covers the entire life cycle of the vehicle from the moment it connects to the charging pile to the moment it leaves. By continuously integrating the instantaneous carbon emission rate, the full life cycle carbon footprint data output by the multidimensional feature extraction and carbon flow coupling module is a dynamic cumulative total that integrates real-time grid status, traffic conditions and user behavior characteristics. The calculation logic of this data can control the carbon accounting error to within 8.3%.

[0100] In a specific embodiment of the present invention, the user behavior classification and queuing estimation module is configured to receive core behavior feature vectors, run a hidden Markov model to calculate user type weight coefficients, call an improved queuing theory model based on user type weight coefficients, use the Weibull distribution function to fit the probability density of charging time to replace the exponential distribution assumption, and calculate and output the predicted value of queuing waiting time.

[0101] The user behavior classification and queuing estimation module is configured as a logical operation unit on the server side. The user behavior classification and queuing estimation module strictly follows the following steps to perform the calculation process from unstructured behavioral characteristics to quantified time costs.

[0102] The specific operation of the user behavior classification and queuing estimation module in calculating the user type weight coefficients using the Hidden Markov Model is as follows: The user hidden state fuzzy decoding program is executed. The user behavior classification and queuing estimation module obtains the core behavior feature vector output by the multi-dimensional feature extraction and carbon flow coupling module. A hidden state space containing home commuting, long-distance travel, and local transportation states is constructed in the Hidden Markov Model. The core behavior feature vector is mapped to the hidden state space. The home commuting state is configured as representing the physical state where the vehicle is charging during weekday morning and evening peak hours and has a remaining mileage greater than 50 kilometers. The long-distance travel state is configured as representing the physical state where the vehicle's single charge exceeds 80% of the total battery capacity and the geographical coordinates are located near the highway network. The local transportation state is configured as representing the physical state where the vehicle's remaining mileage is less than 30 kilometers and a slow charging service is requested.

[0103] In practice, the user behavior classification and queuing estimation module parameterizes the physical boundaries of each state when constructing the implicit state space.

[0104] For home commuting scenarios, the user behavior classification and queue estimation module sets the logical gating conditions for weekday morning and evening peak hours to 7:00-9:00 AM and 5:00-7:00 PM.

[0105] The user behavior classification and queuing estimation module sets the threshold for judging whether the remaining mileage data is greater than 50 kilometers to a value derived from the statistical data of the average daily round-trip mileage of urban commuter vehicles. The purpose of setting this threshold is to exclude random behavior of being forced to charge nearby due to extreme power depletion and to lock in planned charging behavior.

[0106] For long-distance travel, the user behavior classification and queuing estimation module sets the power threshold exceeding 80% of the total battery capacity as the logical judgment value representing the need for deep fast charging.

[0107] The user behavior classification and queuing estimation module sets the spatial determination range around the highway network to a geofenced area within five kilometers of highway entrances / exits or service areas.

[0108] For local commuter vehicles, the user behavior classification and queuing estimation module establishes the condition of remaining mileage data of less than 30 kilometers as typical operating data of old residential areas or short-distance commuter vehicles. The identification of this condition is used to trigger the system to schedule such vehicles for slow charging during the off-peak period of the power grid load.

[0109] Following the above steps, the user behavior classification and queuing estimation module performs Gaussian mixture model fitting operations to calculate the observation probability density values ​​of the core behavior feature vector in each state in the hidden state space. It then calls the forward and backward iterative algorithm to calculate the posterior probability values ​​of the current vehicle observation sequence in the home commuting state, long-distance travel state, and local transportation state, respectively. The user behavior classification and queuing estimation module performs normalization operations on the posterior probability values ​​to generate three normalized probability components corresponding to the above three states, and combines the three normalized probability components to output the user type weight coefficient.

[0110] In practice, the user behavior classification and queuing estimation module uses a Gaussian mixture model containing three to five Gaussian components to fit the distribution of the observed data.

[0111] The user behavior classification and queuing estimation module uses a forward-backward iterative algorithm to calculate soft probabilities. The posterior probability value quantifies the mixed attributes of the current vehicle behavior pattern.

[0112] After normalization, the user type weight coefficients are output as a three-dimensional vector with a sum of 1 (e.g., 0.6, 0.3, 0.1). As a linear mapping parameter connecting discrete states and continuous distribution parameters, the user type weight coefficients enable a numerical characterization of mixed-type user behavior.

[0113] The user behavior classification and queuing estimation module calls an improved queuing theory model based on user type weight coefficients. It uses the Weibull distribution function to fit the probability density of charging time to replace the exponential distribution assumption. The specific operation of calculating and outputting the predicted queuing waiting time is as follows: The distribution parameter dynamic nonlinear synthesis program is executed. The user behavior classification and queuing estimation module obtains the user type weight coefficients and calls the pre-stored baseline scale parameters corresponding to the three user types. The user type weight coefficients are used as the exponential term to perform a weighted geometric mean operation on the baseline scale parameters to generate the synthetic scale parameters specific to the current vehicle. The value of the synthetic shape parameter of the Weibull distribution function is fixed at 1.8.

[0114] In practice, the user behavior classification and queuing estimation module pre-stores baseline scale parameters derived from historical data statistics: for home commuting, the baseline scale parameter is set to 1.2 hours; for long-distance travel, the baseline scale parameter is set to 0.8 hours; and for local commuting, the baseline scale parameter is set to 2.0 hours.

[0115] The user behavior classification and queuing estimation module uses weighted geometric mean calculation because the distribution of time data follows a log-normal law, and geometric mean can suppress the influence of extreme values.

[0116] The synthetic shape parameter was fixed at 1.8. This value was set based on the fitting analysis results of the charging order duration of new energy private cars. This value represents the single-peak characteristic of the charging service completion rate showing an initial increase followed by a decrease over time, and it has a long tail on the right. The purpose of this value is to correct the underestimation of the probability of long-term charging by the traditional exponential distribution model.

[0117] Subsequently, an improved queuing time calculation program based on the second-order central moment is executed. The user behavior classification and queuing estimation module calls the gamma function to calculate the second-order central moment of the Weiber distribution function. The specific calculation logic is to calculate the gamma function value with the independent variable being one plus twice the reciprocal of the composite shape parameter, subtract the square of the gamma function value with the independent variable being one plus twice the reciprocal of the composite shape parameter, and multiply the result of the subtraction operation by the square of the composite scale parameter to generate the variance of the charging service duration.

[0118] In practice, the user behavior classification and queuing estimation module performs this step to obtain the volatility characteristics of charging service duration.

[0119] The variance of charging service duration quantifies the degree of uncertainty in service time. The user behavior classification and queuing estimation module transforms the dimensionless composite shape parameter into a statistical variance term through a gamma function.

[0120] The variance of charging service duration is used as a core variable input into the subsequent Polachek-Shinchin formula, and the variance of charging service duration directly determines the nonlinear growth rate of the predicted waiting time.

[0121] Finally, the user behavior classification and queuing estimation module obtains the current average vehicle arrival rate data and system service intensity data. It performs calculations using the extended Polachek-Shinchin formula, multiplying the square of the system service intensity data by the square of the product of the average vehicle arrival rate data and the composite scale parameter by the variance coefficient, and then dividing by twice the average vehicle arrival rate data multiplied by one minus the difference in system service intensity data to obtain the physical average waiting time under steady-state conditions. A queuing patience correction factor is then introduced to perform a multiplicative correction operation on the physical average waiting time, and the predicted queuing waiting time value is output.

[0122] In practice, the system service intensity data represents the busyness of the charging station, and the value of the system service intensity data must be less than one to ensure the steady state of the system.

[0123] The average arrival rate data represents the number of vehicles arriving per unit time. The extended Polachek-Shinchin formula takes into account the degree of congestion and the volatility of service time.

[0124] In order to convert physical time into subjective decision-making time perceived by users, the user behavior classification and queuing estimation module introduces a queuing patience correction factor. In this embodiment, for users identified as having a higher weight for long-distance travel, the queuing patience correction factor is set to 1.2 to 1.5; for users identified as having a higher weight for local transportation, the queuing patience correction factor is set to 0.7 to 0.9.

[0125] The user behavior classification and queuing estimation module outputs a predicted queuing time value through multiplication correction operation. The predicted queuing time value incorporates the user's subjective psychological characteristics, providing a decision-making basis for subsequent multi-agent game scheduling.

[0126] In a specific embodiment of the present invention, the multi-agent collaborative scheduling module is configured to receive full lifecycle carbon footprint data and queuing time prediction values, solve the Nash equilibrium point based on a two-layer game architecture, and generate the optimal charging power allocation strategy.

[0127] The multi-agent cooperative scheduling module is configured as a logical operation unit on the server side. The multi-agent cooperative scheduling module strictly follows the following steps to execute the calculation process from multi-dimensional decision data to physical power control instructions.

[0128] The specific operation of the multi-agent cooperative scheduling module in solving the Nash equilibrium point based on the two-layer game architecture is as follows: Execute the multi-objective coupled heterogeneous utility functional construction program. The multi-agent cooperative scheduling module obtains the full life cycle carbon footprint data output by the multi-dimensional feature extraction and carbon flow coupling module, and at the same time obtains the queuing waiting time prediction value output by the user behavior classification and queuing estimation module. Construct a generalized heterogeneous utility functional for each site-level agent participating in the game. The generalized heterogeneous utility functional is configured as a linear superposition combination of the charging service logarithmic gain term, the full life cycle carbon emission index penalty term, and the queuing time nonlinear loss term.

[0129] In practice, the multi-agent cooperative scheduling module parameterizes each component when constructing the generalized heterogeneous utility functional.

[0130] For the logarithmic gain term of the charging service: the multi-agent collaborative scheduling module performs a natural logarithmic operation on the function with the optimal charging power allocation strategy to be solved as the independent variable and multiplies it by a preset revenue sensitivity coefficient.

[0131] The revenue sensitivity coefficient represents the degree to which charging operators value economic benefits. In this embodiment, the value of the revenue sensitivity coefficient is set to 0.8 to 1.0. The basis for setting this value range is the economic principle of diminishing marginal returns. The purpose of setting this value range is to ensure that the utility increase brought about by the increase in power gradually slows down as the power base increases, and to prevent the algorithm from outputting a strategy that excessively demands grid resources.

[0132] For the penalty item of the carbon emission index throughout the entire life cycle: the multi-agent collaborative scheduling module performs natural index calculation on the carbon footprint data throughout the entire life cycle and multiplies it by a preset carbon sensitivity coefficient.

[0133] The carbon sensitivity coefficient characterizes the system's penalty for environmental impact. In this embodiment, the carbon sensitivity coefficient is set to 1.5 to 2.0. The basis for setting this value is the non-linear penalty mechanism of regional environmental protection regulations for carbon emission overruns. The purpose of setting this value is to drastically reduce the total utility value through exponential calculation when carbon emissions increase, thereby forcing the system logic to prioritize charging during low-carbon periods.

[0134] To address the nonlinear loss term in queuing time: the multi-agent collaborative scheduling module performs a power operation on the predicted queuing time and multiplies it by a preset time sensitivity coefficient. The time sensitivity coefficient represents the risk of user churn. In this embodiment, the exponent for the power operation is set to 1.5. The purpose of setting this exponent is to simulate the nonlinear growth of user waiting anxiety over time.

[0135] Following the above steps, a dynamic interaction program based on a two-layer Stackelberg game architecture is executed. The multi-agent collaborative scheduling module establishes a game topology structure that includes upper-level regional agents and lower-level site-level agents. The regional agents generate shadow electricity price signals based on the total load constraints of the regional power grid. The site-level agents receive the shadow electricity price signals and adjust their own power demand with the goal of maximizing the generalized heterogeneous utility functional. The multi-agent collaborative scheduling module constructs a game-coupled variational inequality describing the coupling relationship between the upper-level regional agents and the lower-level site-level agents. This variational inequality characterizes the zero-sum equilibrium state of the utility gradients of each agent under the premise of satisfying the physical constraints of the power grid.

[0136] In practice, the multi-agent collaborative scheduling module defines the shadow electricity price signal as a Lagrange multiplier, and the physical essence of the shadow electricity price signal is the pricing of the scarcity of power grid resources.

[0137] When the total load in the region approaches the upper limit of the transformer capacity, the value of the shadow electricity price signal shows an exponential upward trend, thereby suppressing the power demand of the lower-level site-level agents. The game-coupled variational inequality, as the mathematical expression of the existence of Nash equilibrium, characterizes that under the current combination of electricity price and strategy, no site-level agent can obtain a higher utility value by unilaterally changing its own power strategy.

[0138] The specific operations of the multi-agent cooperative scheduling module in generating the optimal charging power allocation strategy are as follows: A Nash equilibrium solution program based on distributed momentum gradient is executed. The multi-agent cooperative scheduling module uses a distributed projection gradient algorithm to iteratively solve the game-coupled variational inequality. In each iteration cycle, each site-level agent only communicates with its neighboring nodes in the topology to exchange power state information. The multi-agent cooperative scheduling module calculates the first-order gradient value of the generalized heterogeneous utility functional with respect to the optimal charging power allocation strategy. A momentum acceleration factor is introduced to perform inertial correction operations on the power increment generated in the previous iteration cycle. The current power strategy is updated by combining the first-order gradient value with the consensus protocol data of the neighboring nodes. The updated power strategy is then mapped back to the allowed physical power range using the projection operator. The above iterative steps are repeated until the power strategies of all site-level agents converge to a steady state to obtain the Nash equilibrium point.

[0139] In practical implementation, the multi-agent collaborative scheduling module adopts a distributed architecture to distribute the computational load. Regarding the momentum acceleration factor: the momentum acceleration factor characterizes the inertia during the iteration process.

[0140] In this embodiment, the momentum acceleration factor is set to 0.01 to 0.1. The basis for setting this value is the convergence rate test results of large-scale convex optimization problems. The purpose of setting this value is to suppress numerical oscillations during the iteration process and enable the algorithm to converge within milliseconds.

[0141] Regarding the projection operator: The purpose of the projection operator is to force the calculated theoretical power value to be within the physical range allowed by the hardware (e.g., between zero and 120 kilowatts), preventing the algorithm from outputting unexecutable negative power or overload power.

[0142] Regarding the Nash equilibrium point: The Nash equilibrium point is the theoretically optimal power reference value after compromise of the interests of all parties.

[0143] Finally, the constrained model predictive control physical execution program is executed. The multi-agent cooperative scheduling module sets the Nash equilibrium point as the reference trajectory and constructs a rolling optimization objective function. The rolling optimization objective function is configured to minimize the weighted sum of the square of the tracking error between the actual output power and the reference trajectory and the square of the power change rate. The multi-agent cooperative scheduling module introduces a power change rate constraint to limit the ramp-up rate of the charging pile power. Under the premise of satisfying the power change rate constraint, the control sequence in the future finite prediction time domain is solved. The first time step element in the control sequence is selected as the physical control command, and the final optimal charging power allocation strategy is output.

[0144] In practice, the multi-agent cooperative scheduling module introduces a rolling optimization objective function to resolve the contradiction between the theoretical optimal solution and hardware security constraints.

[0145] Implementation details of the power change rate constraint: The power change rate constraint represents the physical threshold that limits the rate of increase or decrease of the current. In this embodiment, the power change rate constraint is set to no more than five kilowatts per minute. The basis for setting this threshold is the electrochemical safety characteristics of the power battery pack and the grid connection specifications of the grid side for impact loads. The purpose of setting this threshold is to prevent lithium plating of the battery or grid voltage flicker caused by drastic fluctuations in the Nash equilibrium point.

[0146] Output of physical control commands: The multi-agent cooperative scheduling module solves the control sequence for the next five to ten minutes in each control cycle (e.g., per second), but only executes the first time step element in the control sequence.

[0147] This rolling mechanism ensures that the system can respond in real time to the latest grid conditions and vehicle demands, and the final output of the optimal charging power allocation strategy is a smooth, safe, and efficient physical control curve that balances low carbon emissions and efficiency.

[0148] Example 2

[0149] like Figure 2 As shown, this invention also discloses a multi-source coupling dynamic optimization method for new energy vehicle charging networks, including the following steps:

[0150] Step S1: Collect driving trajectory data, brand preference data, payment willingness data and real-time carbon emission factor data of new energy charging vehicles, and preprocess the above data to output a standardized multi-source heterogeneous dataset.

[0151] Step S2: Receive a standardized multi-source heterogeneous dataset, extract core behavioral feature vectors using an improved kernel principal component analysis algorithm, and combine the core behavioral feature vectors with real-time carbon emission factor data from the power grid. Generate full life cycle carbon footprint data using a dynamic carbon flow coupling model calibrated by a genetic algorithm, and transmit the core behavioral feature vectors to step S3 and the full life cycle carbon footprint data to step S4.

[0152] Step S3: Receive the core behavior feature vector, run the hidden Markov model to calculate the user type weight coefficient, call the improved queuing theory model based on the user type weight coefficient, use the Weibull distribution function to fit the probability density of charging time to replace the exponential distribution assumption, and calculate and output the predicted value of queuing waiting time.

[0153] Step S4: Receive full lifecycle carbon footprint data and queuing time prediction values, solve the Nash equilibrium point based on a two-layer game architecture, and generate the optimal charging power allocation strategy.

[0154] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-source coupling dynamic optimization system for new energy vehicle charging networks, characterized in that, include: The module includes data acquisition and standardization processing, multi-dimensional feature extraction and carbon flow coupling, user behavior classification and queuing estimation, and multi-agent collaborative scheduling. The data acquisition and standardization processing module is configured to collect driving trajectory data, brand preference data, payment willingness data, and real-time carbon emission factor data of new energy charging vehicles, and preprocess the above data to output a standardized multi-source heterogeneous dataset. The multidimensional feature extraction and carbon flow coupling module is configured to receive standardized multi-source heterogeneous datasets, extract core behavioral feature vectors using an improved kernel principal component analysis algorithm, and generate full life cycle carbon footprint data by combining the core behavioral feature vectors and real-time carbon emission factor data of the power grid through a dynamic carbon flow coupling model calibrated by a genetic algorithm. The core behavioral feature vectors are then transmitted to the user behavior classification and queuing estimation module, and the full life cycle carbon footprint data is transmitted to the multi-agent collaborative scheduling module. The user behavior classification and queuing estimation module is configured to receive core behavior feature vectors, run a hidden Markov model to calculate user type weight coefficients, call an improved queuing theory model based on user type weight coefficients, use the Weibull distribution function to fit the probability density of charging time to replace the exponential distribution assumption, and calculate and output the predicted value of queuing waiting time. The multi-agent collaborative scheduling module is configured to receive full lifecycle carbon footprint data and queuing time predictions, solve for the Nash equilibrium point based on a two-layer game architecture, and generate the optimal charging power allocation strategy.

2. The multi-source coupling dynamic optimization system for new energy vehicle charging networks according to claim 1, characterized in that, The data acquisition and standardization module performs the following preprocessing operations to output a standardized multi-source heterogeneous dataset: A1 executes the heterogeneous data dynamic spatiotemporal anchor point synchronization mapping program, calls the timestamp sequence in the driving trajectory data of new energy charging vehicles as the reference clock, establishes a dynamic spatiotemporal anchor point mapping field, obtains real-time carbon emission factor data of the power grid, performs time decay operation based on exponential function on the real-time carbon emission factor data of the power grid, generates carbon emission factors with time weight, and then performs tensor product expansion operation to map the carbon emission factors with time weight to the vector space of driving trajectory data, generating a preliminary synchronization spatiotemporal matrix; A2 executes a manifold completion program based on topological graph Laplacian regularization to obtain the preliminary synchronized spatiotemporal matrix and pre-stored road network topology data. It constructs a regularization objective function containing the graph Laplacian matrix, which consists of a data fidelity term and a manifold constraint term. Iterative minimization is performed on the regularization objective function to calculate the Frobenius norm between the preliminary synchronized spatiotemporal matrix and the ideal complete data matrix to be solved. At the same time, it calculates the trace of the transpose of the ideal complete data matrix, the graph Laplacian matrix, and the ideal complete data matrix. The ideal complete data matrix is ​​updated using the gradient descent algorithm until convergence, and the topological complete reconstruction matrix is ​​output.

3. The multi-source coupling dynamic optimization system for new energy vehicle charging networks according to claim 2, characterized in that, The data acquisition and standardization module preprocesses data to output a standardized multi-source heterogeneous dataset. Specific operations include: A3 executes a semantically driven hybrid dimensional nonlinear modulation program to obtain a topologically complete reconstruction matrix, brand preference data, and willingness-to-pay data. It constructs a brand feature mapping matrix, transforms discrete brand preference data into a continuous semantic potential function, performs dot product weighting, modulates the amplitude of the willingness-to-pay data using the semantic potential function, and generates a semantically weighted payment vector. Subsequently, it performs hyperbolic tangent nonlinear mapping on the topologically complete reconstruction matrix and the semantically weighted payment vector to compress the physical values ​​to a preset normalization interval. It also performs robust scaling based on the median and interquartile range statistics within a local sliding window to generate a semantically hybrid normalized tensor. A4 executes a kernel-space-oriented feature tensor stacking and pre-mapping procedure, calculates the covariance matrix of the semantic mixture normalized tensor, determines the bandwidth parameter of the Gaussian kernel function based on the distribution characteristics of the covariance matrix, performs tensor concatenation operation on the semantic mixture normalized tensors of different vehicles along the sample dimension to construct a global feature matrix, performs eigenvalue decomposition operation on the kernel matrix corresponding to the global feature matrix to generate an eigenvalue diagonal matrix, performs inverse square root operation on the eigenvalue diagonal matrix to generate a whitening transformation matrix, multiplies the global feature matrix with the whitening transformation matrix, and outputs a normalized multi-source heterogeneous dataset.

4. The multi-source coupling dynamic optimization system for new energy vehicle charging networks according to claim 3, characterized in that, The multidimensional feature extraction and carbon flow coupling module utilizes an improved kernel principal component analysis algorithm to extract core behavioral feature vectors. The specific operations are as follows: B1 executes a kernel matrix evolution construction procedure based on semantic weights. The multidimensional feature extraction and carbon flow coupling module receives the standardized multi-source heterogeneous dataset output by the data fusion and feature extraction module. It separates the brand preference component and the willingness to pay level component from the standardized multi-source heterogeneous dataset. It performs vector synthesis operation on the brand preference component and the willingness to pay level component to generate semantic weight factors. It calculates the square of the Euclidean distance between any two sample points in the standardized multi-source heterogeneous dataset in the regenerating kernel Hilbert space. It divides the square of the Euclidean distance by twice the square of the preset kernel bandwidth parameter and takes the negative number. It performs an exponential operation on the calculation result to generate the basic Gaussian kernel function value. At the same time, it performs an inner product operation on the semantic weight factors of the two sample points. It performs a hyperbolic tangent nonlinear mapping operation on the result of the inner product operation and adds one to generate the semantic gain coefficient. It performs a multiplication operation on the basic Gaussian kernel function value and the semantic gain coefficient to generate the matrix elements of the semantic weighted kernel evolution matrix. B2, then executes the nonlinear projection extraction procedure for the core behavioral feature vector, performs centering processing and eigenvalue decomposition operation on the semantic weighted kernel evolution matrix, selects several feature vectors whose cumulative contribution rate meets the preset threshold in descending order of eigenvalues, projects the standardized multi-source heterogeneous dataset onto the low-dimensional feature subspace spanned by several feature vectors, generates the projected feature matrix, and decomposes the projected feature matrix along the feature dimension into charging demand density component, brand preference component, and low carbon potential value component, and combines the charging demand density component, brand preference component, and low carbon potential value component to output the core behavioral feature vector.

5. The multi-source coupling dynamic optimization system for new energy vehicle charging networks according to claim 4, characterized in that, The multidimensional feature extraction and carbon flow coupling module combines core behavioral feature vectors and real-time carbon emission factor data from the power grid to generate full life-cycle carbon footprint data through a dynamic carbon flow coupling model calibrated by a genetic algorithm. The specific operations are as follows: B3 executes a three-dimensional coupled dynamic carbon flow differential modeling program to obtain the charging demand density component and low-carbon potential value component in the core behavioral feature vector. At the same time, it obtains real-time carbon emission factor data of the power grid and establishes a dynamic carbon flow coupled differential equation describing the change of carbon emission rate over time. The dynamic carbon flow coupled differential equation is composed of the linear superposition of energy source term, traffic source term, and carbon sink offset term. The multidimensional feature extraction and carbon flow coupling module generates the energy source term by performing a multiplication operation on the real-time carbon emission factor data of the power grid and the charging demand density component. It generates the traffic source term by performing a continuous multiplication operation on the basic factor of the private car-specific carbon emission coefficient to be calibrated, the road network traffic flow density data after nonlinear power function transformation, and the vehicle mileage data. It generates the carbon sink offset term by performing a multiplication operation on the preset carbon sink coefficient and the low-carbon potential value component. B4. Execute a closed-loop parameter calibration and integration procedure based on a genetic algorithm to construct an optimization objective function. The optimization objective function is configured to calculate the cumulative absolute error between the simulated carbon footprint data generated by the dynamic carbon flow coupling differential equation and the historical measured carbon footprint data. Run the genetic algorithm to perform iterative optimization operations of selection, crossover, and mutation on the basic factors of the private car-specific carbon emission coefficient to be calibrated until the value of the optimization objective function converges to the preset minimum value. Output the optimal private car-specific carbon emission coefficient. Substitute the optimal private car-specific carbon emission coefficient into the dynamic carbon flow coupling differential equation and perform definite integral operations on the dynamic carbon flow coupling differential equation in the entire life cycle time domain to output the full life cycle carbon footprint data.

6. The multi-source coupling dynamic optimization system for new energy vehicle charging networks according to claim 5, characterized in that, The specific steps for the user behavior classification and queuing estimation module to run the Hidden Markov Model and calculate the user type weight coefficients are as follows: C1 executes the user hidden state fuzzy decoding program. The user behavior classification and queuing estimation module obtains the core behavior feature vector output by the multi-dimensional feature extraction and carbon flow coupling module. In the hidden Markov model, a hidden state space containing home commuting state, long-distance travel state, and local transportation state is constructed. The core behavior feature vector is mapped to the hidden state space. The home commuting state is configured to represent the physical state of the vehicle charging during the morning and evening peak hours on weekdays and the remaining mileage data is greater than 50 kilometers. The long-distance travel state is configured to represent the physical state of the vehicle charging more than 80% of the total battery capacity in a single charge and the geographical coordinates being located around the highway network. The local transportation state is configured to represent the physical state of the vehicle having less than 30 kilometers of remaining mileage data and requesting slow charging service. C2, the user behavior classification and queuing estimation module performs Gaussian mixture model fitting operation, calculates the observation probability density value of the core behavior feature vector in each state in the hidden state space, and calls the forward and backward iterative algorithm to calculate the posterior probability value of the current vehicle observation sequence in the home commuting state, long-distance travel state, and local transportation state respectively. The user behavior classification and queuing estimation module performs normalization operation on the posterior probability value, generates three normalized probability components corresponding to the above three states respectively, and combines the three normalized probability components to output the user type weight coefficient.

7. The multi-source coupling dynamic optimization system for new energy vehicle charging networks according to claim 6, characterized in that, The user behavior classification and queuing estimation module calls an improved queuing theory model based on user type weight coefficients. It uses the Weibull distribution function to fit the probability density of charging time to replace the exponential distribution assumption, and calculates and outputs the predicted queuing time as follows: C3 executes the dynamic nonlinear synthesis program of distribution parameters. The user behavior classification and queuing estimation module obtains the user type weight coefficients. At the same time, it calls the pre-stored baseline scale parameters corresponding to the three user types respectively. It uses the user type weight coefficients as the exponent to perform a weighted geometric mean operation on the baseline scale parameters to generate the current vehicle-specific synthetic scale parameters. The value of the synthetic shape parameter of the Weiber distribution function is fixed at 1.

8. C4 executes an improved queuing time calculation program based on the second-order central moment. The user behavior classification and queuing estimation module calls the gamma function to calculate the second-order central moment of the Weiber distribution function. The specific calculation logic is to calculate the gamma function value of the independent variable as one plus twice the reciprocal of the composite shape parameter, subtract the square of the gamma function value of the independent variable as one plus twice the reciprocal of the composite shape parameter, and multiply the result of the subtraction operation by the square of the composite scale parameter to generate the variance of the charging service duration. C5, the user behavior classification and queuing estimation module, obtains the current average vehicle arrival rate data and system service intensity data. It performs calculations using the extended Polachek-Shinchin formula, multiplying the square of the system service intensity data by the square of the product of the average vehicle arrival rate data and the composite scale parameter by the variance coefficient, and then dividing by twice the average vehicle arrival rate data multiplied by one minus the difference in system service intensity data to obtain the physical average waiting time under steady-state conditions. A queuing patience correction factor is then introduced to perform a multiplicative correction operation on the physical average waiting time, and the predicted queuing waiting time value is output.

8. The multi-source coupling dynamic optimization system for new energy vehicle charging networks according to claim 7, characterized in that, The specific operations of the multi-agent cooperative scheduling module in solving the Nash equilibrium point based on the two-layer game architecture are as follows: D1 executes a multi-objective coupled heterogeneous utility functional construction procedure. The multi-agent collaborative scheduling module obtains the full life cycle carbon footprint data output by the multi-dimensional feature extraction and carbon flow coupling module, and simultaneously obtains the queuing wait time prediction value output by the user behavior classification and queuing estimation module. It constructs a generalized heterogeneous utility functional for each site-level agent participating in the game. The generalized heterogeneous utility functional is configured as a linear superposition combination of the charging service logarithmic gain term, the full life cycle carbon emission index penalty term, and the queuing time nonlinear loss term. The multi-agent collaborative scheduling module performs a natural logarithmic operation on the function with the optimal charging power allocation strategy to be solved as the independent variable and multiplies it by a preset gain sensitivity coefficient to generate the charging service logarithmic gain term. It performs a natural exponential operation on the full life cycle carbon footprint data and multiplies it by a preset carbon sensitivity coefficient to generate the full life cycle carbon emission index penalty term. It performs a power operation on the queuing wait time prediction value and multiplies it by a preset time sensitivity coefficient to generate the queuing time nonlinear loss term. D2 executes a dynamic interaction program based on a two-layer Stackelberg game architecture. The multi-agent collaborative scheduling module establishes a game topology structure that includes upper-level regional agents and lower-level station-level agents. Regional agents generate shadow electricity price signals based on the total load constraints of the regional power grid. Station-level agents receive shadow electricity price signals and adjust their own power demand with the goal of maximizing the generalized heterogeneous utility functional. The multi-agent collaborative scheduling module constructs a game-coupled variational inequality that describes the coupling relationship between upper-level regional agents and lower-level station-level agents. This variational inequality characterizes the zero-sum equilibrium state of the utility gradients of each agent under the premise of satisfying the physical constraints of the power grid.

9. The multi-source coupling dynamic optimization system for new energy vehicle charging networks according to claim 8, characterized in that, The specific operations of the multi-agent collaborative scheduling module in generating the optimal charging power allocation strategy are as follows: D3 executes the Nash equilibrium solution program based on distributed momentum gradient. The multi-agent cooperative scheduling module uses a distributed projection gradient algorithm to iteratively solve the game-coupled variational inequality. In each iteration cycle, each station-level agent only communicates with its neighboring nodes in the topology to exchange power state information. The multi-agent cooperative scheduling module calculates the first-order gradient value of the generalized heterogeneous utility functional with respect to the optimal charging power allocation strategy. It introduces a momentum acceleration factor to perform inertial correction on the power increment generated in the previous iteration cycle. It updates the current power strategy by combining the first-order gradient value with the consensus protocol data of the neighboring nodes, and uses the projection operator to map the updated power strategy back to the allowed physical power range. The above iterative steps are repeated until the power strategies of all station-level agents converge to a steady state to obtain the Nash equilibrium point. D4 executes the restricted model predictive control physical execution program. The multi-agent cooperative scheduling module sets the Nash equilibrium point as the reference trajectory and constructs a rolling optimization objective function. The rolling optimization objective function is configured to minimize the weighted sum of the square of the tracking error between the actual output power and the reference trajectory and the square of the power change rate. The multi-agent cooperative scheduling module introduces a power change rate constraint to limit the ramp-up rate of the charging pile power. Under the premise of satisfying the power change rate constraint, it solves the control sequence in the future finite prediction time domain, selects the first time step element in the control sequence as the physical control command, and outputs the final optimal charging power allocation strategy.

10. A multi-source coupling dynamic optimization method for new energy vehicle charging networks, applied to the multi-source coupling dynamic optimization system for new energy vehicle charging networks as described in any one of claims 1-9, characterized in that, Includes the following steps: Step S1: Collect driving trajectory data, brand preference data, payment willingness data and real-time carbon emission factor data of new energy charging vehicles, and preprocess the above data to output a standardized multi-source heterogeneous dataset. Step S2: Receive a standardized multi-source heterogeneous dataset, extract core behavioral feature vectors using an improved kernel principal component analysis algorithm, and combine the core behavioral feature vectors with real-time carbon emission factor data from the power grid. Generate full life cycle carbon footprint data using a dynamic carbon flow coupling model calibrated by a genetic algorithm, and transmit the core behavioral feature vectors to step S3 and the full life cycle carbon footprint data to step S4. Step S3: Receive the core behavior feature vector, run the hidden Markov model to calculate the user type weight coefficient, call the improved queuing theory model based on the user type weight coefficient, use the Weibull distribution function to fit the probability density of charging time to replace the exponential distribution assumption, and calculate and output the predicted value of queuing waiting time. Step S4: Receive full lifecycle carbon footprint data and queuing time prediction values, solve the Nash equilibrium point based on a two-layer game architecture, and generate the optimal charging power allocation strategy.

Images