Intelligent cooperative charging station management system and method for electric vehicle

Abstract

The invention relates to the technical field of charging station management, and particularly discloses an intelligent cooperative charging station management system and method for an electric automobile, and the system comprises the steps: integrating real-time communication and honeycomb partition management through a cooperative decision module, constructing an efficient and state-aware charging network, and carrying out the state-aware management of the charging network; the decision model can accurately adapt to the actual operation state of the target charging station; the cooperative scheduling module optimizes resource allocation by using a state matrix and an aggregation space, so that the utilization efficiency of charging station resources is remarkably improved; and the intelligent charging scheduling decision module combines reinforcement learning rapid iteration and historical data driven prior knowledge and demand prediction, so that the generation of an optimal scheduling strategy is accelerated, the excessive dependence of decision on real-time data is reduced, and the response capability and operation stability of the system for responding to dynamic demands and complex scenes are greatly enhanced, and thus, the system is more intelligent. And finally, efficient, intelligent and networked collaborative management of the electric vehicle charging service is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of charging station management, in particular to an intelligent collaborative charging station management system and method for electric vehicles. BACKGROUND

[0002] The intelligent collaborative charging station management system for electric vehicles collects the state of each charging pile, vehicle battery data and power grid information in real time through Internet of Things sensors; these data are transmitted to the central cloud platform, where traditional optimization algorithms are used for static analysis to generate a preliminary charging scheduling scheme; finally, start-stop instructions or power adjustment instructions are sent to the charging piles through the communication network to achieve limited load balancing and basic services.

[0003] For example, the Chinese invention patent with publication number CN118013763A discloses a vehicle-network collaborative scheduling method and system considering the response ability of electric vehicle users, a storage medium and an electronic device, relating to the field of vehicle-network collaborative scheduling. The electric vehicle charging behavior characteristics are simulated by the Monte Carlo simulation method to generate the time of electric vehicles reaching the charging station, the time of leaving the charging station, and the required charging power; then an incentive price function is designed to encourage electric vehicles to participate in vehicle-network interaction; then, based on the time of electric vehicles reaching the charging station, the time of leaving the charging station, and the required charging power, and combined with the incentive price function, a vehicle-network collaborative scheduling optimization model is established; finally, the vehicle-network collaborative scheduling optimization model is solved to obtain the optimal charging strategy and the incentive price at each time.

[0004] For example, the Chinese invention patent with publication number CN119357486A discloses an intelligent electric vehicle charging station recommendation system and method. It includes a big data processing layer, a multi-path recall layer, a recall deduplication layer, and a sorting layer; the big data processing layer collects basic information and outputs user feature information, user context feature information, charging station feature information, and charging station context feature information; user feature information mainly includes user basic information, preference information, statistical information, and user behavior information; charging station feature information mainly includes charging station basic information and charging station statistical information; the multi-path recall layer includes recall based on user context, recall based on user interest degree, and recall based on charging station similarity; the recall based on user interest degree uses a collaborative filtering algorithm, and the recall based on charging station similarity is obtained through a deep learning recall model; the results of the multi-path recall layer are given to the recall deduplication layer, the output of the recall deduplication layer is given to the sorting layer, and the sorting layer outputs the recommendation results.

[0005] But in the process of implementing the embodiments of the application, the applicant finds that the above-mentioned technology at least has the following technical problems: the heterogeneous scene characteristics formed by the wide distribution, numerous access points and complex operating environment of the electric vehicle charging pile network lead to the contradiction between the initial model training and the scene universality requirement. This contradiction leads to the fact that the model trained based on the initial data set is difficult to fully adapt to the real dynamic operating environment of each region and each type of charging pile when it is actually deployed on a large scale, and its generalization ability and prediction accuracy face significant challenges. Therefore, it is urgent to design an effective mechanism to dynamically adjust and optimize the model during the retraining process of the model to ensure that it can accurately adapt to the constantly changing and highly complex actual charging pile application scenarios. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides an intelligent collaborative charging station management system and method for electric vehicles, which can effectively solve the problems involved in the background art.

[0007] To achieve the above object, the present application is implemented by the following technical scheme: the present application provides an intelligent collaborative charging station management system for electric vehicles, comprising: an intelligent charging collaborative decision module for building an intelligent charging network, a cellular partition management and a VANET real-time communication, realizing efficient utilization of target charging station resources and intelligent collaborative charging station management of electric vehicles, transmitting state information of the target charging station to the charging collaborative decision model, and testing data of the charging collaborative decision model, so that the charging collaborative decision model adapts to the actual operation scenario of the target charging station; a collaborative scheduling module for building a state matrix and forming an aggregated state space based on the electric vehicle parameters in the charging state in the target charging station, thereby optimizing the collaborative scheduling of the charging collaborative decision model; an intelligent charging scheduling decision module for making the reinforcement learning policy iteration algorithm converge to the optimal solution through multi-round online interaction and parameter iteration, and generating the best scheduling strategy suitable for the target charging station, wherein in the generation process of the best scheduling strategy, the prior knowledge model trained by embedding historical charging data and the future charging demand prediction algorithm are used to reduce the dependence on real-time data.

[0008] The second aspect of the present application provides a kind of intelligent collaborative charging station management method for electric vehicle, comprising: step one, build intelligent charging network, cellular partition management and VANET real-time communication, realize the efficient use of target charging station resources and the intelligent collaborative charging station management of electric vehicle, the state information of target charging station is transmitted to charging collaborative decision model, and data test is carried out to charging collaborative decision model, so that charging collaborative decision model adapts to the actual operation scene of target charging station;Step two, based on the electric vehicle parameter in the charging state in target charging station, state matrix is constructed and aggregate state space is formed, to optimize the collaborative scheduling of charging collaborative decision model;Step three, through multiple rounds of online interaction and parameter iteration, the reinforcement learning strategy iteration algorithm converges to optimal solution, generates the best scheduling strategy suitable for target charging station, wherein, in the generation process of best scheduling strategy, through embedding priori knowledge model trained by historical charging data and future charging demand prediction algorithm, the dependence on real-time data is reduced.

[0009] Compared with the prior art, the embodiments of the present application have at least the following advantages or benefits: (1) The present application provides an intelligent collaborative charging station management system and method for electric vehicles. Through the integration of real-time communication and cellular partition management by the collaborative decision module, an efficient and state-aware charging network is constructed, enabling the decision model to accurately adapt to the actual operating state of the target charging station. The collaborative scheduling module optimizes resource allocation using state matrices and aggregate spaces, significantly improving the utilization efficiency of charging station resources. The intelligent charging scheduling decision module combines reinforcement learning rapid iteration with historical data-driven prior knowledge and demand prediction, not only accelerating the generation of optimal scheduling strategies and reducing the over-reliance of decision-making on real-time data, but also significantly enhancing the system's response capability and operational stability in response to dynamic demand and complex scenarios, ultimately achieving efficient, intelligent and networked collaborative management of electric vehicle charging services.

[0010] (2) The intelligent charging collaborative decision module builds an intelligent charging network and cellular partition management, combined with VANET real-time communication technology, to achieve efficient utilization of target charging station resources and intelligent collaborative scheduling of electric vehicles. Its advantage lies in dynamically integrating the state information of charging stations and optimizing the decision model through data testing, making it more suitable for actual operating scenarios. This module benefits by improving the resource utilization rate of charging stations, reducing charging queue time, and enhancing the system's adaptability to complex environments, providing more stable and efficient charging services for electric vehicle users.

[0011] (3) The cooperative scheduling module constructs a state matrix and forms an aggregated state space based on the real-time charging parameters of electric vehicles in the target charging station, thereby optimizing the scheduling strategy of the charging cooperative decision-making model. The advantage is that through a data-driven modeling method, the dynamic relationship between charging demand and resource allocation is accurately captured, and multi-vehicle cooperative scheduling is achieved. The benefit is that it can balance the charging load, avoid local overload or idling, improve overall operational efficiency, and reduce the peak pressure of the power grid, providing support for the long-term stable operation of the charging station.

[0012] (4) The intelligent charging scheduling decision-making module uses a reinforcement learning policy iteration algorithm to generate the optimal scheduling strategy through multiple rounds of online interaction and parameter optimization, and reduces the dependence on real-time data by combining historical data and future demand forecasts. The advantage is that through prior knowledge models and prediction algorithms, the accuracy and robustness of the decision are improved. The benefit is that it can quickly respond to dynamically changing demands, optimize charging timing and power allocation, and improve user experience, making charging station management more intelligent and sustainable. BRIEF DESCRIPTION OF DRAWINGS

[0013] The invention is further illustrated by the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the invention. For those skilled in the art, other drawings can be obtained without creative effort based on the following drawings.

[0014] Figure 1 The system module connection diagram of the present invention.

[0015] Figure 2 The method step flowchart of the present invention.

[0016] Figure 3 The accuracy optimization flowchart of the charging cooperative decision-making model of the present invention.

[0017] Figure 4 The accuracy optimization flowchart of the charging cooperative decision-making model of the present invention.

[0018] Figure 5 The data concurrency optimization flowchart of the charging cooperative decision-making model of the present invention.

[0019] Figure 6 The data concurrency optimization flowchart of the charging cooperative decision-making model of the present invention. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, and not all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the scope of the present application.

[0021] Referring to Figure 1 The first aspect of the present application provides an intelligent cooperative charging station management system for an electric vehicle, comprising: an intelligent charging cooperative decision module, a cooperative scheduling module, an intelligent charging scheduling decision module, and a database.

[0022] The intelligent charging cooperative decision module is connected with the cooperative scheduling module, the cooperative scheduling module is connected with the intelligent charging scheduling decision module, and the intelligent charging cooperative decision module, the cooperative scheduling module, and the intelligent charging scheduling decision module are all connected with the database.

[0023] The intelligent charging cooperative decision module is used to build an intelligent charging network, a honeycomb partition management, and a VANET real-time communication, to realize efficient utilization of target charging station resources and intelligent cooperative charging station management of electric vehicles, to transmit state information of the target charging station to a charging cooperative decision model, and to test data of the charging cooperative decision model, so that the charging cooperative decision model adapts to the actual operation scene of the target charging station.

[0024] The scene is arranged in a city environment and is composed of electric vehicles, charging stations, a global charging controller (GCC), and a roadside unit (RSU). The GCC is responsible for collecting interaction information of each charging station and the RSU. The RSU is uniformly deployed along the city roads and has a certain computing capability, and is responsible for migration calculation and distribution of service information including electric vehicle charging requests. A GPS device is arranged on each electric vehicle to guide the vehicle to travel according to the returned instructions. In order to facilitate city environment planning, the city map is divided into a honeycomb network and is uniformly divided into M*N units, wherein each honeycomb unit represents an area. The problem of vehicle searching for a charging station is described as follows: the vehicle in each area travels randomly without any prior information interaction, and when the electric quantity of the unmanned vehicle is low to a threshold value, the vehicle starts to make a charging request. After sending the charging request, the vehicle uses VANET to transmit information and interacts with the global controller, that is, the request is sent to the RSU and uploaded to the global controller. At the same time, the charging station transmits its own information to the global controller at a certain time interval, and after a series of calculations, the charging decision is transmitted back to the vehicle via VANET (Internet of Vehicles) and guides the vehicle to travel to the charging station.

[0025] The city map is divided according to a cellular network structure, and is evenly divided into a grid layout containing M rows and N columns. Here, M represents the number of units in the vertical direction (rows) of the grid, and N represents the number of units in the horizontal direction (columns) of the grid. Therefore, the entire urban area is divided into M*N independent cellular units, each of which represents a specific geographic area. This division provides a spatial management framework for subsequent vehicle positioning, information collection (such as charging requests), and global controller (GCC) charging decision-making and path guidance.

[0026] Specifically, the charging coordination decision model is tested with data, and the specific analysis process is as follows: the state information of the target charging station includes a first data set of the target charging station and a second data set of the target charging station. The first data set of the target charging station is used to test the prediction accuracy of the charging coordination decision model, and the second data set of the target charging station is used to test the prediction accuracy and data concurrency capability of the charging coordination decision model.

[0027] The first data set of the target charging station is transmitted to the charging coordination decision model, the parameter prediction data set of the charging coordination decision model is collected, and the reference parameter prediction data set corresponding to the first data set is compared to determine whether to optimize the accuracy of the charging coordination decision model. The parameter prediction data set refers to the set of predicted output parameters generated by the charging coordination decision model after processing the input information in the first data set of the target charging station.

[0028] After the accuracy optimization is completed, the second data set of the target charging station is transmitted to the charging coordination decision model, the performance prediction data set of the charging coordination decision model is collected, and the performance prediction data set corresponding to the second data set is compared to determine whether to optimize the data concurrency capability of the charging coordination decision model. The performance prediction data set refers to the set of performance prediction output parameters generated by the charging coordination decision model after processing the input information in the second data set of the target charging station.

[0029] The first data set and the second data set are both derived from the target charging station and are used together to test the prediction accuracy of the charging coordination decision model; the first data set is specifically and only used to verify the basic prediction accuracy of the model, and is usually of low scale or complexity, while the second data set, in addition to verifying the prediction accuracy, additionally and emphatically tests the data concurrency capability of the model in high-load, multiple-request simultaneous processing scenarios and its accuracy under high pressure, so its scale is larger or simulates a high-concurrency scenario, and is a more stringent comprehensive performance test set. Therefore, the second data set test can only be performed after the model passes the first data set accuracy test in the use process.

[0030] The charging coordination decision model is an artificial intelligence model and is constructed based on machine learning. The core function of the model is to comprehensively analyze multi-source information in a charging station (such as charging pile state, vehicle demand, power grid load, electricity price, queue, etc.), predict key charging state parameters (such as future occupancy state, charging power, completion time, etc.), and implicitly or explicitly generate optimization decision suggestions (such as power allocation, scheduling strategy), so as to realize the coordinated optimization of charging efficiency, power grid balance, user experience and operation benefit.

[0031] In another example embodiment, the charging coordination decision model is constructed based on reinforcement learning. First, a communication and computing architecture is deployed in a discretized urban scenario of electric vehicles-roadside unit (RSU)-global charging controller (GCC)-charging station. The GCC collects interaction information of each station and RSU, and divides the urban map into a cellular grid to limit the charging of vehicles in the current area or adjacent area to balance accessibility and scheduling. On this basis, the time is discretized by granularity with the double parameters of the stay time in the area and the required charging time, and the aggregated state space independent of the vehicle scale is obtained by binning classification. According to this, the action decision space and state-action mapping are designed, so that large-scale solving is converted into a small-scale computable problem. Then, the least squares policy iteration (ER-LSPI) based on experience replay is used to evaluate and improve the online policy until convergence, so as to obtain the optimal scheduling strategy. The PySNS3 platform and V2X joint debugging environment are used to carry out sample collection, simulation and field measurement to verify the effectiveness and convergence of the model.

[0032] With the stay time in the area δ_soj and the required charging time δ_cha as two-dimensional indicators, the state matrix of H_m×H_m is formed by discretizing the time granularity T. At time t, the scheduling system makes a decision value as according to the state space diagonal variable set (Λ[0], Λ[H_m-1]) to decide which vehicles to charge and which vehicles to continue to delay. All possible decision actions constitute the decision space As. The purpose of the current layer scheduling is to balance the load in the charging area, while ensuring that the charging of each electric vehicle is completed before leaving. Therefore, the optimization objective function is defined as: ; wherein h is the position number of the main diagonal grid index, h=0, 1, 2,..., , is the number of vehicles occupying the state s (occupancy matrix of the current time slot), is the decision value, E sp is the instantaneous state of the charging vehicle, is the minimum value selected from to as, It refers to the product of the minimum sum of the values ​​from 0 to the minimum value of the dwell time minus the charging time in all states and the amplification factor β (β is the state transition variable), where k is the state, and state s includes, but is not limited to, the state of the car being fully charged and the state of the car being charged.

[0033] The optimization objective is to find the optimal policy μ* = min( This minimizes the mean of the above equation. Since the state transition probability in the above equation is unknown, μ* cannot be solved precisely, so other learning methods must be used to obtain the optimal policy. We propose using an Experience-Based Least Squares Policy Iteration Algorithm (ER-LSPI), which is based on a linear least squares function approximation structure, combining online policy iteration with the ER algorithm to search for the optimal policy. This algorithm framework mainly consists of two parts: an optimal policy learning algorithm and a sample collection and utilization algorithm. The core of the former is least squares policy iteration. Its basic idea is to start with any initial policy μ0, and in each iteration I (I≥0), use a policy evaluation algorithm to interact with the environment to collect samples online to evaluate the current policy and calculate the Q-value function. Then, through CH+1=argmax[ The strategy is improved until the algorithm finally converges to the optimal strategy μ*.

[0034] Q is a stable function value. The Q value from the i-th state s to the decision state as, CH+1 is the optimal decision obtained by CH+1 iterations.

[0035] The model's input consists of multi-dimensional information from the target charging station dataset (either the first or second dataset), primarily including the real-time status of charging piles (power, charge level, status), vehicle / user request information (battery SOC (current remaining charge) / SOH (battery health), demand), grid / energy information (load, electricity price, renewable energy forecast), time environment information, and possible queue / historical data. The model's output is a set of predicted key charging state parameters (such as future state, power, completion time, and load prediction), used to evaluate the accuracy index of charging state prediction. Under high-concurrency testing (the second dataset), the output results themselves or accompanying performance indicators also constitute performance prediction output parameters, used to evaluate concurrency capability and accuracy under high voltage.

[0036] In data preparation and preprocessing, firstly, comprehensive data is collected from the historical database of charging stations, including real-time status of charging piles, vehicle battery parameters, grid load curves, timestamps, weather records, and actual output results (such as charging completion time). Then, data cleaning is performed, handling missing values ​​using imputation or deletion strategies; outliers (such as negative power values ​​or excessively long charging times) are corrected. Next, feature engineering is performed: derived features are constructed using domain knowledge, and numerical features are standardized (to eliminate the influence of dimensions), while categorical features are one-hot encoded. Model architecture design then begins. Based on the dual objectives of time-series prediction and optimization decision-making in charging collaborative decision-making, an LSTM neural network is used in the prediction layer to handle time-series dependencies. The number of neurons in the input layer of the LSTM layer equals the preprocessed feature dimension. The decision layer connects to a fully connected network (FCN) to output predicted values ​​(such as SOC, completion time), and integrates a reinforcement learning module to generate scheduling strategies. This architecture employs a composite loss function: smoothed L1 loss for prediction tasks and a custom reward function for decision-making tasks, with AdamW selected as the optimizer. During model training, the system initiates iterative training. Input batch data is processed by an LSTM layer to capture temporal patterns, and fully connected layers output predicted values. Simultaneously, the reinforcement learning module generates action probability distributions. The training process calculates the composite loss: the prediction part calculates the smoothed L1 error between the predicted value and the true label, and the decision part calculates the policy gradient loss based on a custom reward function. Automatic differentiation is used to calculate the gradient of the loss with respect to millions of model parameters. The AdamW optimizer applies gradient pruning to update weights while using L2 regularization to constrain weight magnitudes. During training, some neurons in the fully connected layers are randomly dropped out to force the network to learn redundant feature representations. After each epoch, model performance is evaluated on the validation set, and an early stopping mechanism is triggered to prevent overfitting. Finally, validation and tuning are performed, with real-time monitoring of the training and validation loss curves. If significant divergence occurs, tuning measures are taken, such as increasing the dropout rate and using Bayesian optimization to explore the optimal combination of hyperparameters on the validation set. If the validation metrics stagnate, the model structure is improved, and the model is retrained based on the tuning results until the validation loss converges stably, indicating that the model performance has reached expectations.

[0037] The testing method significantly improves the reliability of the intelligent collaborative charging station management system by verifying the predictive accuracy and data concurrency capabilities of the charging collaborative decision-making model in stages. First, the model's prediction accuracy is optimized using a first dataset to ensure the accuracy of decisions regarding key parameters such as charging demand and electricity price fluctuations, providing a scientific basis for dynamic pricing and load balancing. Then, stress testing with a second dataset strengthens the model's high-concurrency processing capabilities, enabling the system to respond in real-time to the collaborative scheduling needs of multiple charging stations and electric vehicles. This two-stage verification mechanism not only reduces the risk of resource allocation imbalance caused by prediction errors but also enhances stability in complex scenarios by increasing system throughput, ultimately optimizing the operational efficiency of the charging station cluster and meeting the dual requirements of accurate decision-making and efficient response for collaborative charging systems in a smart grid environment.

[0038] The collaborative scheduling module is used to construct a state matrix and form an aggregated state space based on the parameters of electric vehicles in the charging state at the target charging station, thereby optimizing the collaborative scheduling of the charging collaborative decision-making model.

[0039] A scalable state model is constructed by aggregating the dwell time of vehicles within the charging area with the required charging time. The two duration parameters are discretized into integer cells at fixed time intervals, and the number of cells is determined based on the longest dwell time, forming a two-dimensional state matrix with fixed dimensions. A scheduling action space is designed based on this matrix, optimizing decisions with load balancing as the goal, thereby solving the curse of dimensionality problem caused by the increase in vehicle scale. Simultaneously, the choice of time interval needs to be balanced: smaller intervals reduce discretization errors but increase matrix dimensions, while larger intervals improve computational efficiency but introduce approximation errors. Therefore, the core research will quantify the impact of this error on scheduling performance to seek the optimal time granularity to achieve a balance between accuracy and scalability.

[0040] The construction of the state matrix and the formation of the aggregated state space refer to the following steps: In terms of system architecture, a vehicle-RSU-Global Controller (GCC)-charging station vehicle-to-grid communication link is adopted, and the charging station reports its status to the GCC at fixed time slots. In terms of modeling method, the dwell time δ_soj and the required charging time δ_cha are used as two-dimensional indicators, which are discretized at the time granularity T to form an H_m×H_m state matrix. The matrix elements are the vehicle counts at the same discrete grid point, thus obtaining an aggregated state space that is independent of vehicle size and only related to time granularity. Based on this aggregated state and the set of diagonal variables, action decisions are made. The objective function is mainly based on regional load balancing and supplemented by a penalty for incomplete charging. In terms of solution, the least squares strategy iteration with empirical replay (ER-LSPI) is used to converge online to the optimal strategy.

[0041] A large number of vehicles arrive at the charging area for orderly charging. Their arrival times differ, but their maximum tolerance time is the same. Using the departure time as a benchmark, the required charging time δ_cha can be effectively estimated based on the State of Charge (SOC) and charging rate. This, together with the waiting time δ_wait within the area, constitutes the vehicle's dwell time δ_soj within the area. Assuming all vehicles in the same area have the same charging rate, the charging energy required by each vehicle can be converted into the required charging time δ_cha. Let there be NM vehicles in region Z at time t. Let the i-th charging vehicle EV in region Z be... i The state is represented as s=(δ_soj i , δ_cha i ), where i represents the number of the charging vehicle, i∈{1,NM}. Using the time slot interval T, δ_soj i and δ_cha i Discretization, i.e., [δ_soj i / T] and [δ_cha i A state matrix of H_m × H_m is formed, where each element represents the number of vehicles in the same state. Assuming the longest dwell time of an electric vehicle within the charging area is Tm, then H_m = [Tm / T]. Therefore, the dimension of the state space is independent of the number of charging vehicles within the area, and only depends on the time granularity of each grid division. However, it was also found that the time granularity introduces a certain approximation error during discretization. Reducing this error would inevitably affect the model's scalability, while increasing the matrix dimension would fail to reduce the curse of dimensionality, creating a contradiction. Therefore, this application will analyze the impact of this approximation error on the overall scheduling decision and discuss the possibility of an optimal time granularity.

[0042] Figure 3The flowchart for optimizing the accuracy of the charging collaborative decision-making model of this invention is as follows: The system collects a parameter prediction dataset and compares it with a reference dataset to calculate the charging state prediction accuracy index. The system then determines whether the calculated charging state prediction accuracy index is greater than or equal to a preset charging state prediction accuracy threshold. If the charging state prediction accuracy index is greater than or equal to the charging state prediction accuracy threshold, the system determines that the prediction accuracy meets the requirements, and the process proceeds to the second dataset testing and prediction accuracy and concurrency capability stage, that is, using a second independent dataset to test the model's prediction accuracy and its ability to handle concurrent requests. If the charging state prediction accuracy index is less than the charging state prediction accuracy threshold, the system determines that accuracy optimization is needed. To perform optimization, the system increases the model regularization coefficient. After completing the optimization operation of increasing the regularization coefficient, the system recalculates and obtains the optimized charging state prediction accuracy index. The system then enters a decision point to determine whether to perform deeper optimization, deciding whether further optimization measures are needed based on the current situation.

[0043] Figure 4 The flowchart below illustrates the accuracy-focused deep optimization process for the charging collaborative decision-making model of this invention. Whether deep optimization is needed depends on whether the optimized charging state prediction accuracy index is greater than or equal to a charging state prediction accuracy threshold. If the charging state prediction accuracy index is greater than or equal to this threshold, the requirements are met and no deep optimization is needed. Subsequently, a second dataset is used to test prediction accuracy and concurrency capabilities. If the charging state prediction accuracy index is less than the charging state prediction accuracy threshold, the deep optimization process begins. In this process, the model is optimized by increasing model dropout to obtain a deeply optimized charging state prediction accuracy index. Next, the deeply optimized charging state prediction accuracy index is evaluated to determine if it is greater than or equal to the charging state prediction accuracy threshold. If it is, the requirements are met and no warning is triggered. The second dataset is used again to test prediction accuracy and concurrency capabilities. If it is still less than the threshold, an accuracy warning is triggered.

[0044] Specifically, to determine whether to optimize the accuracy of the charging collaborative decision-making model, the specific judgment process is as follows: collect the parameter prediction dataset of the charging collaborative decision-making model, compare it with the reference parameter prediction dataset corresponding to the first dataset, obtain the charging state prediction accuracy index, and compare it with the charging state prediction accuracy threshold.

[0045] If the charging state prediction accuracy index is greater than or equal to the charging state prediction accuracy threshold, the prediction accuracy of the charging collaborative decision-making model is deemed to meet the requirements, and no accuracy optimization is required. The model then proceeds to test the prediction accuracy and data concurrency capability using the second dataset from the target charging station.

[0046] The charging state prediction accuracy threshold represents the minimum value within the allowable range of the charging state prediction accuracy index.

[0047] Accurate thresholds for predicting state of charge are stored in the database.

[0048] If the charging state prediction accuracy index is less than the charging state prediction accuracy threshold, it is determined that the accuracy of the charging collaborative decision model needs to be optimized. This involves exporting and saving the abnormal state data of the charging collaborative decision model as an independent raw data file package, while increasing the regularization coefficient of the charging collaborative decision model. After optimization, the parameter prediction dataset of the charging collaborative decision model is re-collected and compared with the reference parameter prediction dataset corresponding to the first dataset. The charging state prediction accuracy index after optimization is obtained, and it is determined whether to perform in-depth accuracy optimization of the charging collaborative decision model.

[0049] Based on the increase of the regularization coefficient of the charging state prediction accuracy index, the charging state prediction accuracy index is divided into different level intervals in the database. Each level interval corresponds to a different increase value of the regularization coefficient. The current regularization coefficient plus the increase value of the regularization coefficient is used to obtain the regularization coefficient for the next step.

[0050] The main benefit of increasing the regularization coefficient in the accuracy index for state-of-charge prediction is that it effectively suppresses the model's overfitting tendency, especially when the data contains noise or outliers. Strengthening the regularization constraint improves the model's generalization ability. This adjustment reduces the model's sensitivity to abnormal fluctuations in the training data, making parameter estimation more robust and thus improving prediction consistency on new data. For the outlier index in the data fusion process, the increased regularization coefficient reduces the impact of outlier data on the overall model by compressing the weights corresponding to outlier features, thereby lowering the outlier index value. This adjustment maintains the model's core learning capabilities while systematically filtering out local biases caused by outliers during the fusion process, making the data fusion results more stable and reliable.

[0051] The advantage of this adjustment mechanism lies in its ability to significantly improve the reliability and efficiency of the intelligent collaborative charging station management system by dynamically evaluating and optimizing the prediction accuracy of the charging collaborative decision-making model. Specifically, this method calculates the charging state prediction accuracy index by comparing the model's prediction data with reference results in real time, and then compares it with the charging state prediction accuracy threshold to quickly determine whether the model needs optimization. If the charging state prediction accuracy index fails to meet the standard, the system automatically exports abnormal data and increases the regularization coefficient to suppress overfitting, and then re-verifies after optimization. This adjustment mechanism not only ensures high accuracy in charging state prediction but also lays a data foundation for subsequent concurrent capability testing. Its core benefits are: through threshold determination and iterative optimization, the accuracy of the charging strategy is guaranteed, avoiding imbalances in charging resource allocation due to prediction deviations; abnormal data is independently saved to provide traceability support for subsequent in-depth model optimization; and the dynamic adjustment of the regularization coefficient enhances the model's generalization ability, enabling the system to adapt to complex and ever-changing charging demand scenarios, ultimately achieving efficient collaborative management of charging station resources.

[0052] Furthermore, the charging state prediction accuracy index is analyzed as follows: the parameter prediction dataset of the charging collaborative decision-making model is collected and compared with the reference parameter prediction dataset corresponding to the first dataset to obtain the attribute parameters of the dataset comparison. The attribute parameters of the dataset comparison include the overlap between the parameter prediction dataset and the reference parameter prediction dataset, the consistency correlation coefficient between the parameter prediction dataset and the reference parameter prediction dataset, and the distribution distance between the parameter prediction dataset and the reference parameter prediction dataset.

[0053] The overlap between the parameter prediction dataset and the reference parameter prediction dataset is quantified by calculating the proportion of sample points where the absolute difference between the predicted and reference values ​​falls within a defined error threshold, reflecting the precise matching degree of individual predictions. The consistency correlation coefficient between the parameter prediction dataset and the reference parameter prediction dataset comprehensively evaluates the correlation and accuracy of the two sets of data. Its calculation combines the mean and standard deviation of the Pearson correlation coefficients of the predicted and reference values; a value closer to 1 indicates a high degree of consistency between the predictions and the reference values ​​in terms of mean, variance, and linearity. Distribution distance focuses on the differences in overall statistical characteristics. By treating the two sets of data as sample distributions, the Wasserstein distance metric is calculated to assess the overall difference between the distribution patterns of the predicted and reference values; a smaller distance value indicates a more similar distribution.

[0054] The calculation of the charging state prediction accuracy index is based on the proportional relationship between the overlap and the bounded overlap between the parameter prediction dataset and the reference parameter prediction dataset, the proportional relationship between the consistency correlation coefficient and the bounded consistency correlation coefficient between the parameter prediction dataset and the reference parameter prediction dataset, and the proportional relationship between the bounded distribution distance and the distribution distance between the parameter prediction dataset and the reference parameter prediction dataset. Specifically, the bounded overlap, bounded consistency correlation coefficient, and bounded distribution distance contrast are used as benchmarks. The proportional relationships between the parameter prediction dataset and the reference parameter prediction dataset, the consistency correlation coefficient, and the distribution distance between the parameter prediction dataset and the reference parameter prediction dataset are calculated and then compared with the corresponding bounded values. Different metric ratios are assigned according to the degree of influence of each proportional relationship on the charging state prediction accuracy index, and the results are summarized to finally obtain the charging state prediction accuracy index.

[0055] The state-of-charge prediction accuracy index is used to quantify the accuracy of model evaluation in predicting the state of charge of batteries.

[0056] ; CYZ is the charging state prediction accuracy index, CZW is the overlap between the parameter prediction dataset and the reference parameter prediction dataset, CZP is the consistency correlation coefficient between the parameter prediction dataset and the reference parameter prediction dataset, CZJ is the distribution distance between the parameter prediction dataset and the reference parameter prediction dataset, CZW_L is the preset defined overlap in the database, CZP_L is the preset defined consistency correlation coefficient in the database, CZJ_L is the preset defined distribution distance in the database, A1 is the measurement ratio corresponding to the preset overlap in the database, A2 is the measurement ratio corresponding to the preset consistency correlation coefficient in the database, and A3 is the measurement ratio corresponding to the preset distribution distance in the database.

[0057] Define the overlap degree, which is the lower limit of the overlap between the parameter prediction dataset and the reference parameter prediction dataset; define the consistency correlation coefficient, which is the lower limit of the consistency correlation coefficient between the parameter prediction dataset and the reference parameter prediction dataset; define the distribution distance, which is the upper limit of the distribution distance between the parameter prediction dataset and the reference parameter prediction dataset.

[0058] A high degree of overlap between the predicted parameter dataset and the reference predicted parameter dataset is usually accompanied by a high correlation coefficient between them, because precise point matching often reflects a strong linear correlation. However, if systematic bias exists, the correlation coefficient between the predicted parameter dataset and the reference predicted parameter dataset may remain high while the overlap between them decreases. The distribution distance between the predicted parameter dataset and the reference predicted parameter dataset indicates that the predicted distribution is generally similar to the reference distribution. This may support a high correlation coefficient between the predicted parameter dataset and the reference predicted parameter dataset, but it does not necessarily lead to a high degree of overlap between them. Conversely, optimizing the correlation coefficient between the predicted parameter dataset and the reference predicted parameter dataset may ignore distribution differences, leading to an increase in the distribution distance between them.

[0059] Different metric ratios are assigned to the degree of influence of each proportional relationship on the accuracy index of state of charge prediction. Specifically, these include the metric ratio corresponding to the overlap degree in the accuracy index, the metric ratio corresponding to the consistency correlation coefficient between the parameter prediction dataset and the reference parameter prediction dataset, and the metric ratio corresponding to the distribution distance between the parameter prediction dataset and the reference parameter prediction dataset. The metric ratio corresponding to the overlap degree in the accuracy index represents the relative proportional relationship between the parameter prediction dataset and the reference parameter prediction dataset and the predefined overlap degree, and its influence on the prediction accuracy index. This metric value quantifies the contribution weight of the overlap degree in the overall evaluation, reflecting the relative importance of parameter dataset overlap in battery state prediction; the consistency correlation coefficient between the parameter prediction dataset and the reference parameter prediction dataset... The correlation coefficient represents the relative proportion between the consistency correlation coefficient between the parameter prediction dataset and the reference parameter prediction dataset and the predefined consistency correlation coefficient, indicating the degree of influence on the prediction accuracy index. This value determines the key role of consistency in the comprehensive evaluation and is used to assess the strength of the impact of data correlation on prediction accuracy. The distribution distance between the parameter prediction dataset and the reference parameter prediction dataset represents the relative proportion between the distribution distance between the parameter prediction dataset and the reference parameter prediction dataset and the predefined distribution distance, indicating the degree of influence on the prediction accuracy index. It is used to measure the sensitivity of distribution differences in charging state prediction, highlighting the diagnostic value of data distribution deviation in the overall quantification, thus comprehensively reflecting the weight allocation and contribution of each indicator in the model accuracy evaluation.

[0060] The database stores preset benchmark parameters for evaluating the state of charge prediction model, including the defined overlap, defined consistency correlation coefficient, and defined distribution distance. These parameters are dynamically bound to the model's output parameter prediction dataset and the reference parameter prediction dataset through a structured parameter mapping table. These parameters, along with the defined overlap, consistency correlation coefficient, and distribution distance, as well as preset metric weights, constitute the calculation system for the state of charge prediction accuracy index. When calculating the model's state of charge prediction accuracy index, the system extracts the current model's index value; compares the index value with the parameter mapping table in the database based on a preset rule base or historical data matching; and finally dynamically outputs the defined overlap, defined consistency correlation coefficient, defined distribution distance, and corresponding metric weight coefficients applicable to the model. The metric weight coefficients all range from 0 to 1.

[0061] Specifically, the process for determining whether to perform in-depth accuracy optimization on the charging collaborative decision-making model is as follows: if the accuracy index of the charging state prediction is greater than or equal to the accuracy threshold of the charging state prediction after optimization, then the prediction accuracy of the charging collaborative decision-making model is deemed to meet the requirements, and there is no need to perform in-depth accuracy optimization on the charging collaborative decision-making model. The process then proceeds to test the prediction accuracy and data concurrency capability using the second dataset of the target charging station.

[0062] If the accuracy index of charging state prediction is less than the accuracy threshold of charging state prediction after optimization, then the accuracy of the charging collaborative decision model is deeply optimized. This involves exporting and saving the abnormal state data stream generated during the charging process of the charging pile as an independent raw data file package, while increasing the Dropout of the charging collaborative decision model and retransmitting the first dataset of the target charging station into the charging collaborative decision model, thereby obtaining the deeply optimized accuracy index of charging state prediction.

[0063] If the accuracy index of the deeply optimized charging state prediction is greater than or equal to the accuracy threshold of the charging state prediction, the prediction accuracy of the charging collaborative decision-making model is determined to meet the requirements. There is no need to trigger an accuracy warning, and the model will switch to using the second dataset of the target charging station to test the prediction accuracy and data concurrency capability.

[0064] If the accuracy index of the deeply optimized charging state prediction is less than the accuracy threshold of the charging state prediction, an accuracy warning will be triggered.

[0065] In the accuracy optimization phase, increasing the Dropout ratio of the charging collaborative decision-making model refers to increasing the original Dropout ratio in the hidden layers of the model by a predetermined amount. This operation should be implemented first in the hidden layers and retrained using the exported original data package of abnormal states. Simultaneously, a dynamic monitoring mechanism must be established: after each increase, the effect should be evaluated using the validation set. If a sharp drop in training set accuracy or an abnormally high validation loss is detected, the Dropout ratio should be immediately adjusted back or optimization should be stopped. Appropriately increasing Dropout, by randomly dropping more neurons during training, reduces the coupling dependency between neurons, enhances the model's generalization ability to training data, and suppresses overfitting, potentially improving prediction accuracy on unknown data.

[0066] The advantage of this adjustment mechanism lies in its ability to significantly improve the reliability and efficiency of the intelligent collaborative charging station management system by dynamically evaluating and optimizing the prediction accuracy of the charging collaborative decision-making model. Specifically, this method calculates the charging state prediction accuracy index by comparing the model's prediction data with reference results in real time, and then compares it with the charging state prediction accuracy threshold to quickly determine whether the model needs optimization. If the charging state prediction accuracy index fails to meet the standard, the system automatically exports abnormal data and increases the regularization coefficient to suppress overfitting, and then re-verifies after optimization. This adjustment mechanism not only ensures high accuracy in charging state prediction but also lays a data foundation for subsequent concurrent capability testing. Its core benefits are: through threshold determination and iterative optimization, the accuracy of the charging strategy is guaranteed, avoiding imbalances in charging resource allocation due to prediction deviations; abnormal data is independently saved to provide traceability support for subsequent in-depth model optimization; and the dynamic adjustment of the regularization coefficient enhances the model's generalization ability, enabling the system to adapt to complex and ever-changing charging demand scenarios, ultimately achieving efficient collaborative management of charging station resources.

[0067] Figure 5 The flowchart for optimizing the data concurrency capability of the charging collaborative decision-making model of this invention is as follows: It begins by testing the model's prediction accuracy and concurrency capability using a second dataset; then, it determines whether the model's prediction anomaly index is less than or equal to the model's prediction anomaly threshold. If the model's prediction anomaly index is less than or equal to the model's prediction anomaly threshold, the current model configuration is maintained, and this version is marked as a version that has passed the concurrency capability verification. If the model's prediction anomaly index is greater than the model's prediction anomaly threshold, the training batch size of the model is increased, and then the optimized model is verified to obtain the optimized model's prediction anomaly index; finally, based on the optimized model's prediction anomaly index, a decision is made on whether to perform deep optimization.

[0068] Figure 6The flowchart below illustrates the process of deeply optimizing the data concurrency capability of the charging collaborative decision-making model according to this invention. The process involves several steps: If the predicted anomaly index of the optimized model is less than or equal to the predicted anomaly threshold, the current model configuration is maintained, and this version is marked as having passed concurrency capability verification. If the predicted anomaly index of the optimized model is greater than the predicted anomaly threshold, the dataset batch size is increased, and the model is re-verified to obtain the predicted anomaly index of the deeply optimized model. Subsequently, it is evaluated whether the predicted anomaly index of this deeply optimized model is less than or equal to the predicted anomaly threshold. If the predicted anomaly index of the deeply optimized model is less than or equal to the predicted anomaly threshold, the current model configuration is maintained again, and it is marked as having passed concurrency capability verification. If the predicted anomaly index of the deeply optimized model is greater than the predicted anomaly threshold, a data concurrency warning operation is finally executed.

[0069] Specifically, the performance prediction dataset of the charging collaborative decision-making model was collected, and the specific analysis process is as follows: The prediction process parameters of the parameter prediction dataset and the matching degree between the parameter prediction dataset and the reference performance prediction dataset are collected. The prediction process parameters of the parameter prediction dataset include the total model processing time and the model throughput of the prediction process.

[0070] The total processing time of the model during the prediction process is recorded using a high-precision timer (time.perf_counter()), which starts before the model receives a single sample input and stops after the prediction output is completed. The time taken for that sample is recorded. Finally, the average time taken for all samples in the entire representative test set is reported.

[0071] Model throughput in the prediction process: After the model warms up, the entire test set samples are continuously and seamlessly input, and the total time T_total from processing the first sample to outputting the last sample is measured; the total time is then substituted into the formula for calculating throughput to obtain the throughput.

[0072] The process involves collecting prediction parameters from the parameter prediction dataset and assessing the matching degree between the parameter prediction dataset and the reference performance prediction dataset. The entire parameter prediction dataset is processed using a charging collaborative decision-making model to obtain predicted performance values, which are then compared with the actual performance values ​​in the reference performance prediction dataset that are strictly aligned using unique identifiers (such as timestamps). Key statistical indicators such as mean absolute error, root mean square error, coefficient of determination, and Pearson correlation coefficient are calculated, and visualization using scatter plots comparing predicted and actual values ​​is employed to comprehensively evaluate the matching degree between the predicted and actual values. All measurements must be performed in a consistent hardware and software environment.

[0073] Match assessment is a comprehensive quantitative and visual analysis process that aligns predicted and actual values, calculates key statistical indicators reflecting the magnitude of error, correlation, and explanatory power, and supplements this with scatter plots that visually demonstrate the relationship between predicted and actual values.

[0074] The second charging state prediction accuracy index is obtained through the following process: Based on the performance prediction dataset output by the charging collaborative decision-making model, this dataset is compared and analyzed with the reference performance prediction dataset corresponding to the second dataset. Finally, the quantitative index calculated by the preset evaluation criteria is used to reflect the accuracy of the model in predicting the charging state in the performance prediction dimension.

[0075] The charging state prediction accuracy index is used to quantify the accuracy of the charging state prediction results. The second charging state prediction accuracy index has the same meaning as the charging state prediction accuracy index, only the acquisition method is different. The second charging state prediction accuracy index is based on the performance prediction dataset output by the charging collaborative decision model, while the charging state prediction accuracy index is based on the parameter prediction dataset of the charging collaborative decision model.

[0076] The calculation of the model prediction anomaly index is based on the proportional relationship between the defined throughput value and the model throughput of the prediction process, the proportional relationship between the total model processing time of the prediction process and the defined total processing time, and the proportional relationship between the defined matching degree and the matching degree between the parameter prediction dataset and the reference performance prediction dataset. Specifically, the defined throughput value, defined total processing time, and defined matching degree are used as benchmarks. The proportional relationships between the model throughput of the prediction process, the total model processing time of the prediction process, and the matching degree between the parameter prediction dataset and the reference performance prediction dataset and the corresponding defined values ​​are calculated. Different metric ratios are assigned to the degree of influence of each proportional relationship on the model prediction anomaly index, and the results are summarized. At the same time, the reciprocal of the second charging state prediction accuracy index is added to obtain the model prediction anomaly index.

[0077] The model prediction anomaly index is used to quantify the degree of anomaly in a model's concurrency capability after its accuracy has been improved.

[0078] ; CMY is the model prediction anomaly index, CML is the model throughput during the prediction process, CMS is the total model processing time during the prediction process, CMC is the matching degree between the parameter prediction dataset and the reference performance prediction dataset, G is the second charging state prediction accuracy index, CML_L is the preset defined throughput value in the database, CMS_L is the preset defined total processing time in the database, CMC_L is the preset defined matching degree in the database, C1 is the metric ratio corresponding to the preset throughput value in the database, C2 is the metric ratio corresponding to the preset total processing time in the database, and C3 is the metric ratio corresponding to the preset accuracy in the database.

[0079] Define the throughput value, which is the lower limit of the model throughput in the prediction process; define the total processing time, which is the upper limit of the model processing time in the prediction process; define the matching degree, which is the lower limit of the matching degree between the parameter prediction dataset and the reference performance prediction dataset.

[0080] The model throughput and total processing time in the prediction process are generally negatively correlated; higher throughput often leads to shorter processing time, and vice versa. However, achieving a higher match between the parameter prediction dataset and the reference performance prediction dataset may require more complex computations or data processing, potentially increasing the total processing time or decreasing the model throughput. These three parameters collectively reflect the inherent trade-off between model efficiency and the match between the parameter prediction dataset and the reference performance prediction dataset. The second charging state prediction accuracy index has a negative impact on the model prediction anomaly index; a higher second charging state prediction accuracy index results in a lower model prediction anomaly index, indicating better data concurrency capability of the charging collaborative decision-making model.

[0081] Different metric ratios are assigned to the degree of influence of various proportional relationships on the model's predicted anomaly index. These include the metric ratio corresponding to the model throughput in the prediction process, the metric ratio corresponding to the total model processing time in the prediction process, and the metric ratio corresponding to the matching degree between the parameter prediction dataset and the reference performance prediction dataset. The metric ratio corresponding to the model throughput in the prediction process represents the relative proportion between the model throughput in the prediction process and the pre-defined throughput, and its influence on the model's predicted anomaly index. This ratio quantifies the contribution weight of throughput in the overall evaluation, reflecting the relative importance of concurrent processing capability in prediction reliability. The metric ratio corresponding to the total model processing time in the prediction process represents the relative proportion between the total model processing time in the prediction process and the pre-defined total processing time, and its influence on the model's predicted anomaly index. This value determines the importance of processing time in the comprehensive evaluation, and is used to assess the key role of model efficiency in anomaly identification. Increased processing latency directly increases the anomaly index, emphasizing the diagnostic value of response speed in concurrent processing. The matching ratio between the parameter prediction dataset and the reference performance prediction dataset represents the relative proportion between the matching degree of the parameter prediction dataset and the reference performance prediction dataset and the preset defined matching degree. It reflects the degree of influence of the model prediction anomaly index, measuring the sensitivity of the matching degree in model prediction anomaly detection. This highlights the diagnostic value of data consistency in overall quantification; insufficient matching degree exacerbates the anomaly index, reflecting the core impact of the deviation between the prediction results and the reference performance on overall reliability. The model prediction anomaly index also incorporates the reciprocal term of the second charging state prediction accuracy index. This part represents the reciprocal relationship of the improvement in prediction accuracy. When the increase is small, the anomaly index increases significantly, quantifying the supplementary contribution of charging state prediction reliability to the anomaly degree and strengthening the robustness assessment of the model in dynamic environments.

[0082] The database stores preset model performance evaluation benchmark parameters, including defined throughput, defined total processing time, and defined matching degree. These benchmark parameters are dynamically bound to the model throughput, total processing time, matching degree between the parameter prediction dataset and the reference dataset, and preset metric coefficients through a structured parameter mapping table, forming the calculation system for the model prediction anomaly index. When it is necessary to calculate the model's prediction anomaly index, the system first extracts the model throughput, total processing time, matching degree between the parameter prediction dataset and the reference dataset, and the second charging state prediction accuracy index of the current model's prediction process. The system then compares the model throughput, total processing time, and matching degree between the parameter prediction dataset and the reference dataset with the corresponding defined values ​​in the database through the mapping table, and applies the metric coefficients associated in the mapping table to calculate the degree of influence of each part. Finally, it integrates the reverse effect of the increase in the charging state prediction accuracy index and dynamically summarizes to generate an anomaly index that comprehensively reflects the model's concurrent processing capability and the reliability of the prediction results. The metric coefficients all range from 0 to 1.

[0083] Furthermore, to determine whether to optimize the data concurrency capability of the charging collaborative decision-making model, the specific judgment process is as follows: compare the model's predicted anomaly index with the model's predicted anomaly threshold.

[0084] If the model predicts anomaly index less than or equal to the model predicts anomaly threshold, the current model configuration is maintained, and there is no need to optimize the data concurrency capability of the charging collaborative decision-making model. The model version is then marked as a version whose concurrency capability has been verified.

[0085] If the model predicts an anomaly index greater than the model predicts anomaly threshold, the data concurrency capability of the charging collaborative decision-making model will be optimized. The specific optimization method is to increase the model training batch based on the model predicts anomaly index and to verify the charging collaborative decision-making model.

[0086] Collect the anomaly index predicted by the optimized model to determine whether to perform in-depth optimization of the data concurrency capability of the charging collaborative decision-making model.

[0087] The model training batch is increased based on the model prediction anomaly index. The model prediction anomaly index is divided into different level intervals in the database. Each level interval corresponds to a different increase value for the model training batch. The current model training batch is added to the increase value of the current model training batch to obtain the next model training batch.

[0088] Increasing the training batch size of the model based on its anomaly prediction index significantly improves the stability and generalization ability of the model, thereby effectively reducing the anomaly prediction index. Increasing the batch size means that each parameter update is based on the average gradient of more samples, which greatly reduces gradient estimation noise, making the optimization process smoother and the learned patterns more statistically meaningful and robust. This enhanced stability directly translates into stronger generalization ability, making the model more reliable and less biased in predicting new data or potential anomalies, ultimately resulting in improved key indicators constituting the model's anomaly prediction index.

[0089] This judgment process dynamically compares the model's predicted anomaly index with the predicted anomaly threshold, enabling intelligent optimization of the charging collaborative decision-making model's data concurrency capabilities. Its advantage lies in significantly improving the efficiency and reliability of intelligent collaborative charging station management for electric vehicles. Specifically, this mechanism triggers optimization only when the model's predicted anomaly index exceeds the predicted anomaly threshold, ensuring stable system operation under high-concurrency scenarios. Through label verification and version / depth optimization judgments, it promotes adaptive iteration of the model, reduces the risk of failure, thereby directly improving the overall service quality of charging stations and enhancing the real-time response capability of collaborative decision-making, ultimately achieving more intelligent and sustainable electric vehicle charging management.

[0090] Specifically, the process for determining whether to perform in-depth optimization of the data concurrency capability of the charging collaborative decision-making model is as follows: if the optimized model prediction anomaly index is less than or equal to the model prediction anomaly threshold, then the current model configuration is maintained, and there is no need to optimize the data concurrency capability of the charging collaborative decision-making model. The model version is then marked as a version whose concurrency capability has been verified.

[0091] The model predicts the anomaly threshold, which represents the maximum value within the allowable range of the model's anomaly index.

[0092] The model predicts anomaly thresholds, which are then stored in the database.

[0093] If the anomaly index predicted by the optimized model is greater than the anomaly threshold predicted by the model, deep optimization is performed. This involves increasing the batch size of the dataset for the charging collaborative decision-making model based on the anomaly index predicted by the optimized model, and then validating the charging collaborative decision-making model.

[0094] Obtain the model prediction anomaly index after deep optimization. If the model prediction anomaly index after deep optimization is less than or equal to the model prediction anomaly threshold, maintain the current model parameter configuration. There is no need to optimize the data concurrency capability of the charging collaborative decision model. Mark the model version as the version that has passed the concurrency capability verification.

[0095] If the anomaly index predicted by the deeply optimized model is greater than the anomaly threshold predicted by the model, a concurrent data warning will be issued.

[0096] Based on the optimized model's predicted anomaly index, the batch size of the charging collaborative decision-making model's dataset is increased. In the database, the optimized model's predicted anomaly index is divided into different level intervals. Each level interval corresponds to a different increase in the batch size of the charging collaborative decision-making model's dataset. The batch size of the charging collaborative decision-making model's dataset for the next step is obtained by adding the increase in the batch size of the charging collaborative decision-making model's dataset from the model training batch to the current batch size of the charging collaborative decision-making model's dataset.

[0097] The former merely increases the training batch passively by predicting anomaly indices to improve model stability, while the latter coordinates the optimized prediction model with the charging decision end-to-end, and its dataset batch optimization directly serves the generation quality of dynamic charging strategies.

[0098] Increasing the batch size of the dataset for the optimized charging collaborative decision-making model, based on the predicted anomaly index, improves the stability and efficiency of model validation, thereby more reliably evaluating the model's predicted anomaly index. Increasing the batch size of the dataset for the charging collaborative decision-making model increases the sample size covered in a single calculation, significantly reducing the noise impact of random data fluctuations. This makes the calculated anomaly index more closely approximate the central tendency of the model's true performance, reducing the possibility of inflated indices or misjudgments due to the randomness of small batches of data.

[0099] This decision-making process, as the core optimization mechanism for the intelligent collaborative charging station management of electric vehicles, offers advantages and benefits primarily in enhancing the system's intelligence, reliability, and resource efficiency. By dynamically comparing the model's predicted anomaly index with its predicted anomaly threshold, the system can adaptively determine the optimization needs for data concurrency capabilities, avoiding unnecessary model adjustments. This saves computational resources and shortens decision-making time, ensuring efficient operation of the charging station. Secondly, when the optimized model's predicted anomaly index exceeds the predicted anomaly threshold, increasing the dataset batch size and re-verifying allows the system to accurately optimize data concurrency processing capabilities, enhancing the model's response speed and prediction accuracy to high-concurrency charging requests, reducing charging queuing delays, and improving user experience. Finally, the introduction of an early warning mechanism can promptly issue alerts when deep optimization fails, helping maintenance personnel intervene in advance to prevent system overload or failure, ensuring the stability and safety of the charging station.

[0100] The intelligent charging scheduling decision module is used to enable the reinforcement learning strategy iteration algorithm to converge to the optimal solution through multiple rounds of online interaction and parameter iteration, and generate the best scheduling strategy suitable for the target charging station. In the process of generating the best scheduling strategy, the reliance on real-time data is reduced by embedding a prior knowledge model trained on historical charging data and a future charging demand prediction algorithm.

[0101] The charging status is determined by determining whether a vehicle can be fully charged within its expected stay time. If the expected stay time is not less than the time required to complete the expected charging demand, the vehicle will not be fully charged before leaving. This project assumes that, under normal circumstances, vehicles entering the charging area will be fully charged before leaving, barring any special or unforeseen events. Therefore, the project will ignore unforeseen circumstances where a vehicle fails to fully charge and will introduce a penalty function mechanism in the optimization design to ensure that all vehicles entering the charging area eventually complete charging.

[0102] The scheduling decision-making mechanism involves observing the current system state at each decision-making moment. This state contains a set of key state variables. Based on the current state, the system makes a decision: which vehicles should begin charging at that moment, and which vehicles should continue to wait for charging. All possible decisions constitute the system's decision space.

[0103] The core objective of this scheduling layer is to achieve load balancing within the charging area and ensure that each electric vehicle completes its expected charging needs before leaving. Therefore, the optimization objective function is defined as minimizing the sum of two parts: the load balancing term minimizes the squared deviation between the total charging power of the area at the current moment and the target expected power. The charging completion guarantee term minimizes a penalty term. This penalty term is calculated based on the difference between the vehicle's expected dwell time and the time required for charging: for each vehicle with delayed charging, if the difference (dwell time - charging time) is negative (i.e., insufficient time), a penalty proportional to the square of the difference is incurred; if the difference is positive or zero (i.e., sufficient time), no penalty is incurred.

[0104] The ultimate goal of optimization is to find an optimal scheduling policy that minimizes the long-run average of the objective function. Since the precise model of the system's state transitions is unknown, the optimal policy cannot be directly and precisely solved. Therefore, data-driven methods such as reinforcement learning are needed to learn the optimal policy.

[0105] The learning algorithm employed is an experience-replay-based least-squares policy iteration. This algorithm framework is based on a linear function approximation structure and comprises two collaborative parts: a sample collection and utilization algorithm, responsible for collecting sample data such as states, decisions, rewards, and new states during interactions with the environment, and storing these samples in an experience replay pool for learning purposes. Experience replay helps break the temporal correlation between samples, improving data utilization efficiency and learning stability. The optimal policy learning algorithm starts with an arbitrary initial scheduling policy. Under the current policy, it uses the collected samples to evaluate the corresponding state-decision value function online using the least squares method. Based on the value function evaluated in the previous step, for each state, it selects the decision that yields the maximum estimate, thus generating an improved new policy, continuously updating the policy and value function. The algorithm stops when the policy no longer improves significantly or the value function converges, yielding an approximate optimal scheduling policy.

[0106] Specifically, the charging collaborative decision-making model is validated. The validation process is as follows: historical data and real-time data are fused together. The real-time samples in the second dataset are calibrated using historical charging data. Virtual samples of future charging demand are generated based on the charging collaborative decision-making model. The calibrated real-time samples are mixed with the generated virtual samples to form an enhanced second dataset. The amount of mixed data is determined based on the model's prediction of the anomaly index.

[0107] Multi-agent collaborative verification, in a distributed scenario, involves electric vehicles or charging stations exchanging information through local communication to cross-verify the accuracy of calibration samples and virtual samples. The amount of data exchanged is determined based on the model's predicted anomaly index. Calibration samples are generated by fusing historical and real-time data to ensure consistency with the actual scenario. Virtual samples are generated by model predictions to verify the reasonableness of the predictions. The verification process involves multi-agent collaborative communication, comparing the consistency between the two types of samples. If the model's predicted anomaly index exceeds the limit, data calibration is triggered. This mechanism simultaneously ensures the reliability of real-time data and the usability of prediction results.

[0108] The amount of mixed data is determined based on the model's predicted anomaly index. In the database, the model's predicted anomaly index is divided into different intervals, and each interval corresponds to a different amount of mixed data, thus obtaining the amount of mixed data used in the fusion of historical data and real-time data.

[0109] The amount of data for exchanging information is determined based on the model-predicted anomaly index. In the database, the model-predicted anomaly index is divided into different intervals, and each interval corresponds to a different amount of data for exchanging information, thereby obtaining the amount of data for exchanging information used in multi-agent collaborative verification.

[0110] By dynamically fusing historical and real-time data to generate high-confidence calibration samples, the model input is effectively aligned with the actual charging scenario, providing a solid data foundation for decision-making. By using virtual samples generated by the model for prospective testing, combined with local communication and cross-validation among multiple agents, the rationality of the prediction results can be evaluated in real time and potential anomalies can be quickly identified.

[0111] Reference Figure 2 As shown, the second aspect of this invention provides a method for managing intelligent collaborative charging stations for electric vehicles. Step one involves constructing an intelligent charging network, implementing cellular partitioning management and VANET real-time communication to achieve efficient utilization of target charging station resources and intelligent collaborative charging station management for electric vehicles. The state information of the target charging station is transmitted to a charging collaborative decision-making model, and the model is tested with data to adapt it to the actual operating scenarios of the target charging station. Step two involves constructing a state matrix and forming an aggregated state space based on the parameters of electric vehicles charging at the target charging station, thereby optimizing the collaborative scheduling of the charging collaborative decision-making model. Step three involves converging the reinforcement learning strategy iteration algorithm to the optimal solution through multiple rounds of online interaction and parameter iteration, generating the best scheduling strategy suitable for the target charging station. During the generation of the best scheduling strategy, a prior knowledge model trained with historical charging data and a future charging demand prediction algorithm are embedded to reduce reliance on real-time data.

[0112] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A smart collaborative charging station management system for electric vehicles, characterized in that, include: The intelligent charging collaborative decision-making module is used to build an intelligent charging network, cellular partition management and VANET real-time communication, to realize the efficient utilization of target charging station resources and intelligent collaborative charging station management of electric vehicles. It transmits the status information of the target charging station to the charging collaborative decision-making model and conducts data testing on the charging collaborative decision-making model, so that the charging collaborative decision-making model can adapt to the actual operation scenario of the target charging station. The collaborative scheduling module is used to construct a state matrix and form an aggregated state space based on the parameters of electric vehicles in the charging state at the target charging station, thereby optimizing the collaborative scheduling of the charging collaborative decision-making model. The intelligent charging scheduling decision module is used to enable the reinforcement learning strategy iteration algorithm to converge to the optimal solution through multiple rounds of online interaction and parameter iteration, and generate the best scheduling strategy suitable for the target charging station. In the process of generating the best scheduling strategy, the reliance on real-time data is reduced by embedding a prior knowledge model trained on historical charging data and a future charging demand prediction algorithm. The specific analysis process for data testing of the charging collaborative decision-making model is as follows: The status information of the target charging station includes a first dataset and a second dataset of the target charging station; The first dataset of the target charging station is transmitted to the charging collaborative decision model. The parameter prediction dataset of the charging collaborative decision model is collected and compared with the reference parameter prediction dataset corresponding to the first dataset to determine whether the accuracy of the charging collaborative decision model should be optimized. The specific process for determining whether to optimize the accuracy of the charging collaborative decision-making model is as follows: Collect the parameter prediction dataset of the charging collaborative decision-making model and compare it with the reference parameter prediction dataset corresponding to the first dataset to obtain the charging state prediction accuracy index, and compare it with the charging state prediction accuracy threshold. If the charging state prediction accuracy index is greater than or equal to the charging state prediction accuracy threshold, the prediction accuracy of the charging collaborative decision-making model is determined to meet the requirements, no accuracy optimization is required, and the model will switch to using the second dataset of the target charging station to test the prediction accuracy and data concurrency capability. The charging state prediction accuracy threshold represents the minimum value within the allowable range of the charging state prediction accuracy index. If the charging state prediction accuracy index is less than the charging state prediction accuracy threshold, it is determined that the accuracy of the charging collaborative decision model needs to be optimized. This involves exporting and saving the abnormal state data of the charging collaborative decision model as an independent raw data file package, while increasing the regularization coefficient of the charging collaborative decision model. After optimization, the parameter prediction dataset of the charging collaborative decision model is re-collected and compared with the reference parameter prediction dataset corresponding to the first dataset. The charging state prediction accuracy index after optimization is obtained, and it is determined whether to perform in-depth accuracy optimization of the charging collaborative decision model.

2. The intelligent collaborative charging station management system for electric vehicles according to claim 1, characterized in that: The data testing of the charging collaborative decision-making model also includes: The first dataset of the target charging station is used to test the prediction accuracy of the charging collaborative decision-making model, and the second dataset of the target charging station is used to test the prediction accuracy and data concurrency capability of the charging collaborative decision-making model. The parameter prediction dataset refers to the set of predicted output parameters generated by the charging collaborative decision-making model after processing the input information in the first dataset of the target charging station. After the accuracy optimization is completed, the second dataset of the target charging station is transmitted to the charging collaborative decision model. The performance prediction dataset of the charging collaborative decision model is collected and compared with the reference performance prediction dataset corresponding to the second dataset to determine whether the data concurrency capability of the charging collaborative decision model should be optimized. The performance prediction dataset refers to the set of performance prediction output parameters generated by the charging collaborative decision model after processing the input information in the second dataset of the target charging station.

3. The intelligent collaborative charging station management system for electric vehicles according to claim 1, characterized in that: The specific analysis process for the charging state prediction accuracy index is as follows: Collect the parameter prediction dataset of the charging collaborative decision-making model and compare it with the reference parameter prediction dataset corresponding to the first dataset to obtain the attribute parameters of the dataset comparison. The attribute parameters of the dataset comparison include the overlap between the parameter prediction dataset and the reference parameter prediction dataset, the consistency correlation coefficient between the parameter prediction dataset and the reference parameter prediction dataset, and the distribution distance between the parameter prediction dataset and the reference parameter prediction dataset. The calculation of the charging state prediction accuracy index is based on the proportional relationship between the overlap between the parameter prediction dataset and the reference parameter prediction dataset and the defined overlap, the proportional relationship between the consistency correlation coefficient between the parameter prediction dataset and the reference parameter prediction dataset and the defined consistency correlation coefficient, and the proportional relationship between the defined distribution distance and the distribution distance between the parameter prediction dataset and the reference parameter prediction dataset. The specific process is as follows: using the defined overlap, defined consistency correlation coefficient, and defined distribution distance contrast as benchmarks, the proportional relationship between the overlap between the parameter prediction dataset and the reference parameter prediction dataset, the consistency correlation coefficient between the parameter prediction dataset and the reference parameter prediction dataset, and the distribution distance between the parameter prediction dataset and the reference parameter prediction dataset and the corresponding defined value are calculated. Different measurement ratios are assigned according to the degree of influence of each proportional relationship on the charging state prediction accuracy index, and the results are summarized to finally obtain the charging state prediction accuracy index. The state-of-charge prediction accuracy index is used to quantify the accuracy of model evaluation in predicting the state of charge of batteries.

4. The intelligent collaborative charging station management system for electric vehicles according to claim 1, characterized in that: The specific process for determining whether to perform accuracy-deep optimization on the charging collaborative decision-making model is as follows: If the accuracy index of charging state prediction is greater than or equal to the accuracy threshold of charging state prediction after optimization, it is determined that the prediction accuracy of the charging collaborative decision model meets the requirements. There is no need to perform in-depth accuracy optimization on the charging collaborative decision model, and the model is then switched to using the second dataset of the target charging station to test the prediction accuracy and data concurrency capability. If the accuracy index of charging state prediction is less than the accuracy threshold of charging state prediction after optimization, then the accuracy of the charging collaborative decision model is deeply optimized. This means exporting and saving the abnormal state data stream generated during the charging process of the charging pile as an independent raw data file package, while increasing the Dropout of the charging collaborative decision model and retransmitting the first dataset of the target charging station to the charging collaborative decision model, thereby obtaining the deeply optimized accuracy index of charging state prediction. If the accuracy index of the charging state prediction after deep optimization is greater than or equal to the accuracy threshold of the charging state prediction, it is determined that the prediction accuracy of the charging collaborative decision-making model meets the requirements, there is no need to trigger an accuracy warning, and the process will switch to using the second dataset of the target charging station to test the prediction accuracy and data concurrency capability. If the accuracy index of the deeply optimized charging state prediction is less than the accuracy threshold of the charging state prediction, an accuracy warning will be triggered.

5. The intelligent collaborative charging station management system for electric vehicles according to claim 2, characterized in that: The specific analysis process for the performance prediction dataset of the charging collaborative decision-making model is as follows: The prediction process parameters of the parameter prediction dataset and the matching degree between the parameter prediction dataset and the reference performance prediction dataset are collected. The prediction process parameters of the parameter prediction dataset include the total model processing time and the model throughput of the prediction process. Obtain the accuracy index of the second charging state prediction; The calculation of the model prediction anomaly index is based on the proportional relationship between the defined throughput value and the model throughput of the prediction process, the proportional relationship between the total model processing time of the prediction process and the total defined processing time, and the proportional relationship between the defined matching degree and the matching degree between the parameter prediction dataset and the reference performance prediction dataset. The specific process is as follows: using the defined throughput value, the defined processing time, and the defined matching degree as benchmarks, the proportional relationship between the model throughput of the prediction process, the total model processing time of the prediction process, and the matching degree between the parameter prediction dataset and the reference performance prediction dataset and the corresponding defined value is calculated. Different measurement ratios are assigned according to the degree of influence of each proportional relationship on the model prediction anomaly index, and the results are summarized. At the same time, the second charging state prediction accuracy index is integrated to obtain the model prediction anomaly index. The model prediction anomaly index is used to quantify the degree of anomaly in a model's concurrency capability after its accuracy has been improved.

6. The intelligent collaborative charging station management system for electric vehicles according to claim 2, characterized in that: The specific process for determining whether to optimize the data concurrency capability of the charging collaborative decision-making model is as follows: Compare the model's predicted anomaly index with the model's predicted anomaly threshold; If the model predicts anomaly index less than or equal to the model predicts anomaly threshold, the current model configuration is maintained, and there is no need to optimize the data concurrency capability of the charging collaborative decision model. The model version is then marked as the version whose concurrency capability has been verified. If the model predicts an anomaly index greater than the model predicts anomaly threshold, the data concurrency capability of the charging collaborative decision-making model will be optimized. The specific optimization method is to increase the model training batch based on the model predicts anomaly index and to verify the charging collaborative decision-making model. Collect the anomaly index predicted by the optimized model to determine whether to perform in-depth optimization of the data concurrency capability of the charging collaborative decision-making model.

7. The intelligent collaborative charging station management system for electric vehicles according to claim 6, characterized in that: The specific process for determining whether to deeply optimize the data concurrency capability of the charging collaborative decision-making model is as follows: If the optimized model predicts anomaly index less than or equal to the model predicts anomaly threshold, then the current model configuration is maintained, there is no need to optimize the data concurrency capability of the charging collaborative decision model, and the model version is marked as the version that has passed the concurrency capability verification. The model predicts anomaly threshold, which represents the maximum value within the allowable range of the model's anomaly index. If the anomaly index predicted by the optimized model is greater than the anomaly threshold predicted by the model, deep optimization is performed, that is, the batch size of the dataset of the charging collaborative decision-making model is increased based on the anomaly index predicted by the optimized model, and the charging collaborative decision-making model is verified. Obtain the model prediction anomaly index after deep optimization. If the model prediction anomaly index after deep optimization is less than or equal to the model prediction anomaly threshold, then keep the current model parameter configuration, do not need to optimize the data concurrency capability of the charging collaborative decision model, and mark the model version as the version that has passed the concurrency capability verification. If the anomaly index predicted by the deeply optimized model is greater than the anomaly threshold predicted by the model, a concurrent data warning will be issued.

8. The intelligent collaborative charging station management system for electric vehicles according to claim 7, characterized in that: The verification process for the charging collaborative decision-making model is as follows: The fusion of historical and real-time data involves using historical charging data to calibrate real-time samples in the second dataset, and generating virtual samples of future charging demand based on the charging collaborative decision-making model. The calibrated real-time samples are then mixed with the generated virtual samples to form an enhanced second dataset. The amount of mixed data is determined based on the model's prediction of anomaly indices. Multi-agent collaborative verification: In a distributed scenario, electric vehicles or charging stations exchange information through local communication to cross-verify the accuracy of calibration samples and virtual samples, and determine the amount of data to be exchanged based on the model's prediction of anomaly indices. The calibration sample is generated by fusing historical and real-time data to ensure that the data is consistent with the actual scenario; The virtual samples are generated by model predictions and are used to verify the rationality of the predictions. The verification process is carried out through multi-agent collaborative communication to compare the consistency between the two types of samples. If the model prediction anomaly index exceeds the standard, data calibration is triggered.

9. A method for managing intelligent collaborative charging stations for electric vehicles, characterized in that: include: Step 1: Construct an intelligent charging network, cellular partition management, and VANET real-time communication to achieve efficient utilization of target charging station resources and intelligent collaborative charging station management for electric vehicles. Transmit the status information of the target charging station to the charging collaborative decision model and conduct data testing on the charging collaborative decision model to make the charging collaborative decision model adapt to the actual operation scenario of the target charging station. Step 2: Based on the parameters of electric vehicles in the charging state at the target charging station, construct a state matrix and form an aggregated state space to optimize the collaborative scheduling of the charging collaborative decision-making model. Step 3: Through multiple rounds of online interaction and parameter iteration, the reinforcement learning strategy iteration algorithm converges to the optimal solution, generating the best scheduling strategy suitable for the target charging station. In the process of generating the best scheduling strategy, the dependence on real-time data is reduced by embedding a prior knowledge model trained with historical charging data and a future charging demand prediction algorithm.

Images