Intelligent management scheduling system for charging pile group

Abstract

The invention discloses a charging pile group intelligent management scheduling system, which belongs to the technical field of charging pile group management scheduling and comprises a real-time monitoring data acquisition module, a battery state joint estimation algorithm engine module, an intelligent scheduling decision module, a charging control and order management module and a data management analysis module. The real-time monitoring data acquisition module is connected with each charging pile in a charging pile group and is used for acquiring and uploading charging pile operation state data and high-precision charging process electric parameter data in real time, and the battery state joint estimation algorithm engine module receives the high-precision charging process electric parameter data; a joint estimation algorithm based on electrochemical model and data driving fusion is built in the controller, and core state parameters of a power battery of a currently connected vehicle are estimated and output in real time. According to the invention, on the basis of realizing intelligent management scheduling of the charging pile group, independent estimation based on the battery state can be realized, and accurate intelligent scheduling can be realized.

Description

Technical Field

[0001] This invention relates to the field of charging pile group management and scheduling technology, and more specifically, to an intelligent management and scheduling system for charging pile groups. Background Technology

[0002] Charging piles are energy-charging devices that provide charging services for electric vehicles. Charging piles are mainly divided into ground-mounted charging piles and wall-mounted charging piles, primarily using time-based, energy-based, and fee-based charging methods. Charging piles can be fixed to the ground or walls and installed in public buildings (public buildings, shopping malls, public parking lots, etc.) and residential parking lots or charging stations. They can charge various models of electric vehicles according to different voltage levels. The input end of the charging pile is directly connected to the AC power grid, and the output end is equipped with a charging plug for charging electric vehicles. Charging piles are mainly public charging piles, generally providing both regular charging and fast charging methods. Users can use a specific charging card to swipe on the human-machine interface provided by the charging pile to perform operations such as selecting the charging method, charging time, and printing fee data. The charging pile display screen can display data such as charging amount, fee, and charging time. Current charging station network scheduling systems primarily focus on the physical matching of charging stations and vehicles and queue management. However, the most critical basis for scheduling decisions—the real-time status of vehicle batteries and accurate prediction of charging time—heavily relies on potentially inaccurate data uploaded by the vehicle's Battery Management System (BMS). This often leads to scheduling errors: for example, assigning a vehicle that actually needs a long charging time to a temporarily idle charging station, resulting in a decrease in parking space turnover; or assigning high power to a vehicle that is about to be fully charged, while vehicles urgently needing charging are left waiting, resulting in overall inefficiency and user complaints about inaccurate waiting times. Summary of the Invention

[0003] To address the problems existing in the prior art, the purpose of this invention is to provide an intelligent management and scheduling system for charging pile groups. In addition to realizing intelligent management and scheduling of charging pile groups, this invention can also independently estimate based on battery status and achieve precise intelligent scheduling.

[0004] To solve the above problems, the present invention adopts the following technical solution: The intelligent management and scheduling system for charging pile groups includes: a real-time monitoring data acquisition module, a battery status joint estimation algorithm engine module, an intelligent scheduling decision module, a charging control and order management module, and a data management and analysis module; The real-time monitoring data acquisition module is connected to each charging pile in the charging pile group and is used to collect and upload charging pile operation status data and high-precision charging process electrical parameter data in real time. The battery state joint estimation algorithm engine module receives the high-precision charging process electrical parameter data. It has a built-in joint estimation algorithm based on the fusion of electrochemical model and data-driven approach to estimate and output the core state parameters of the power battery of the currently connected vehicle in real time. The intelligent scheduling decision module is used to receive the real-time status of the charging pile group, the requests of vehicles waiting to be charged, and the core state parameters of the battery and the prediction of the charging curve. Based on a multi-objective optimization algorithm, the module dynamically generates a scheduling scheme with the core state parameters of the battery and the prediction of the charging curve as the key decision basis. The charging control and order management module connects the intelligent scheduling decision module and each charging pile. It is used to execute scheduling plans, generate charging orders, and send control commands to designated charging piles to start or adjust the charging process. At the same time, it manages the entire process of order payment, settlement and exception handling. The data management and analysis module is used to store all historical data and to optimize the model parameters of the battery state joint estimation algorithm engine and the scheduling strategy of the intelligent scheduling decision module through analysis.

[0005] As a preferred embodiment of the present invention, the battery state joint estimation algorithm engine module specifically includes a data preprocessing and feature extraction unit, a model parameter online identification and correction unit, a multi-state joint estimation unit, and a charging curve dynamic prediction unit. The data preprocessing and feature extraction unit is used to receive and process real-time charging process time-series data measured by the charging pile, the time-series data including terminal voltage. Charging current and battery surface temperature Meanwhile, the unit filters and denoises the raw data, removes outliers, and extracts feature vectors for state estimation. The feature vectors include incremental capacity IC curve features, voltage relaxation curve features, and statistical features at multiple time scales. The online model parameter identification and correction unit is connected to the data preprocessing and feature extraction unit, and is used to identify key parameters of the battery equivalent circuit model online and in real time based on the processed data. This unit includes a state observer and a parameter observer. The state observer The parameter observer is used to track the internal state of a system as it changes over time. These state variables typically include state of charge, terminal voltage, and polarization voltage. This is used for online identification and updating of key parameters in the battery equivalent circuit model. The key parameters include ohmic internal resistance, polarization resistance, and polarization capacitance. The collaborative process of the state observer and parameter observer filter is as follows: the parameter observer first predicts the parameters at the current time based on the parameter estimate value at the previous time step; the state observer uses the parameter value predicted in the previous step, combined with the system input, to predict the system state at the current time step; the state observer updates the state prediction using the actual measurement value to obtain the optimal state estimate and calculates the state prediction error; the parameter observer updates the parameter prediction using the state prediction value provided by the state observer and the actual measurement value to obtain the optimal parameter estimate. The multi-state joint estimation unit is connected to both the data preprocessing and feature extraction unit and the online model parameter identification and correction unit. It is used to receive the feature vector and the updated battery model parameters. This unit integrates multiple heterogeneous joint estimation algorithms, outputs preliminary estimates of the battery's state of charge, health state and internal temperature state, and uses a fusion weighting algorithm based on improved fuzzy entropy to dynamically allocate and fuse multiple preliminary estimates, outputting a high-precision final joint estimation result. The dynamic prediction unit for the charging curve is connected to the multi-state joint estimation unit. It takes the SOC and SOH in the final joint estimation result as the core input, combines the battery's historical charging data, and uses a data-driven model to predict the optimal charging current-voltage curve from the initial SOC to the target SOC and the accurate estimated charging time in the current state. The estimated charging time is used as the core decision-making basis of the charging pile group scheduling system.

[0006] As a preferred embodiment of the present invention, the online model parameter identification and correction unit specifically adopts a parameter identifier based on the extended Kalman filter (EKF). Its state-space model includes the following state equations: The observation equation is ,in This represents the battery model parameter vector at time k. Including ohmic internal resistance Polarization resistance Polarized capacitors The open-circuit voltage (OCV) model parameters are given in units of ohms (Ω), farads (F), and volts (V). The terminal voltage observation at time k is expressed in volts (V). For the nonlinear observation function based on the Thevenin model, The SOC reference value at time k is initially provided by the ampere-hour integration method. and These are process noise and observation noise, respectively. The EKF filter iteratively calculates the Kalman gain and updates the parameter vector in real time. The estimated value is used to adapt to battery aging and temperature changes.

[0007] As a preferred embodiment of the present invention, the multiple heterogeneous joint estimation algorithms integrated in the multi-state joint estimation unit include: A joint SOC-SOH estimator based on Extended Kalman Filter (EKF); A joint SOC-SOH estimator based on unscented Kalman filter (UKF); A state estimator using a Long Short-Term Memory (LSTM) network.

[0008] As a preferred embodiment of the present invention, the fusion weighting algorithm based on improved fuzzy entropy has the following steps for calculating the fusion weights: Calculate the residual of the terminal voltage estimation for each estimator at time k. The unit is volts (V). Build with The membership function is used as the input, and the fuzzy entropy corresponding to each estimator is calculated. This is used to measure the uncertainty of the estimator's output; Set adjustment factor and create a judgment vector. ,when When this happens, the corresponding output of the estimator is considered bad data and suppressed in the weight calculation; Calculate each estimator for the state Fusion weights of (SOC or SOH) , The final fusion estimate is calculated using the following formula: ,in This is the initial estimate for the m-th estimator; The dynamic prediction unit for the charging curve employs a long short-term memory network model based on an attention mechanism. The input sequence of this model consists of a joint estimated state sequence from historical and current times, a charging current sequence, a voltage sequence, and a battery model code. The output of the model is the optimal charging current sequence for multiple future time steps. The estimated filling time The calculation formula is: ,in The target state of charge is 100% or a user-defined value. This represents the final fused SOC estimate at the current moment. The actual usable capacity of the current battery is obtained by multiplying the final fused SOH estimate by the rated capacity, in ampere-hours (Ah). The value is the average of the optimal charging current sequence predicted by the attention long short-term memory network model, in amperes.

[0009] As a preferred embodiment of the present invention, the intelligent scheduling decision module includes a data fusion and knowledge construction unit, a multi-objective optimization decision unit, and an adaptive strategy execution and evaluation unit that are sequentially connected and form a closed loop; The data fusion and knowledge construction unit is used to receive and fuse multi-source heterogeneous real-time data and historical data, and to construct and dynamically update the scheduling domain knowledge graph. The multi-objective optimization decision unit is connected to the data fusion and knowledge construction unit, and is used to call the hybrid algorithm library to perform multi-objective collaborative optimization calculations based on the semantic information and real-time status of the knowledge graph, and generate the optimal or near-optimal scheduling strategy set. The adaptive strategy execution and evaluation unit is connected to the multi-objective optimization decision-making unit. It is used to distribute the scheduling strategy to the execution layer and perform online reinforcement learning and iterative optimization of the decision model and knowledge graph based on the reward signals and key performance indicators (KPIs) from the environmental feedback.

[0010] As a preferred embodiment of the present invention, the data fusion and knowledge construction unit includes a multi-model data governance subunit, a dynamic knowledge graph construction subunit, and an interpretable feature extraction subunit; The multi-model data governance subunit is used to clean, standardize, and correlate and fuse the input power grid flow, equipment status, traffic conditions, and order demand data to establish a data pool with a unified spatiotemporal benchmark. The dynamic knowledge graph construction subunit is based on expert rules and data-driven approach. It represents entities and their relationships in a graph form and uses a graph neural network (GNN) model to dynamically mine potential associations and state evolution patterns. Its graph update is represented as an incremental learning process, where entities are substations, charging piles, production lines and vehicles, and relationships are connections, memberships or loads. The interpretable feature extraction subunit employs deep learning model technology based on attention mechanism to quantify the contribution of each input feature to subsequent decision-making and generate a decision basis report.

[0011] As a preferred embodiment of the present invention, the multi-objective optimization decision unit includes a hybrid algorithm library, a multi-objective solution and trade-off subunit, and a real-time dynamic rescheduling trigger; The hybrid algorithm library integrates heuristic and metaheuristic algorithms, reinforcement learning algorithms, and prediction and sequence modeling algorithms. The heuristic and metaheuristic algorithms are used to quickly obtain high-quality feasible solutions. These heuristic and metaheuristic algorithms are genetic algorithms (GA), and the fitness function of a genetic algorithm can be defined as a weighted sum of multiple objectives. ,in These represent the objective function values ​​for cost, time, and energy consumption, respectively. The weights are dynamically adjustable. The reinforcement learning algorithm, including Deep Q-Network (DQN), is used to learn long-term optimal policies in dynamic environments. The core Q-learning update formula is: ,in Let $\begin{case}{a}{a}$ be the expected value of the cumulative reward per unit of state action value function. The state at time t, For the action to be taken, For immediate rewards, the unit's scalar value is related to the target. For learning rate, The discount factor is dimensionless and used to weigh current and future rewards. The prediction and sequence modeling algorithms are used for predicting load, demand, and ETA, including time-series-based neural networks, and their output formula can be expressed as follows: ,in The predicted value for time t. Given historical sequence data, W and b are the weight matrix and bias vector, respectively. For activation functions; The multi-objective solution and trade-off subunit uses the Pareto optimal front search algorithm to generate a non-dominated solution set and provides a visual trade-off analysis through a human-computer interaction interface for decision-makers to select the final solution. The real-time dynamic rescheduling trigger is based on an event-driven mechanism. When a key state variable is detected to deviate from a preset threshold, it automatically triggers a local or global recalculation of the decision-making process. The driving mechanism includes new order insertion, equipment failure, and traffic congestion.

[0012] As a preferred embodiment of the present invention, the adaptive policy execution and evaluation unit includes a policy distribution and verification subunit, a multi-dimensional feedback monitoring subunit, and an online learning and model update subunit; The strategy distribution and verification subunit converts the optimized scheduling instructions into control instructions, and then issues them after pre-verification of security and feasibility based on a digital twin or simulation environment. The multi-dimensional feedback monitoring subunit collects system feedback data in real time after the strategy is executed. The system feedback data includes physical status, business indicators and economic indicators. The physical status is voltage and location, the business indicators are fulfillment rate and waiting time, and the economic indicators are cost and revenue. The online learning and model update subunit collects feedback data and continuously updates the parameters of the value network or policy network in the decision model through a reinforcement learning framework. At the same time, successful disposal cases and corrected policies are used as new knowledge samples and integrated into the dynamic knowledge graph through case reasoning methods to achieve the self-evolution of the knowledge base.

[0013] As a preferred embodiment of the present invention, the collaborative working mechanism of the intelligent scheduling decision module is as follows: From feature extraction and knowledge reasoning to optimization and solution, each step retains and outputs readable intermediate results and logical chains, forming a complete decision tracing report; The upper-level multi-objective optimization decision-making unit handles strategic-level resource allocation and long-term planning, while the lower-level adaptive strategy execution and evaluation unit handles tactical-level real-time adjustments and anomaly recovery. The two layers collaborate through a shared knowledge graph. The module adopts a microservice architecture. Its core units and algorithm library are defined through standardized interfaces. By replacing the domain-specific knowledge graph architecture and objective function, it can be quickly adapted to different application scenarios such as power dispatching, transportation and logistics, intelligent manufacturing, and charging pile management.

[0014] Compared with the prior art, the advantages of this invention are: (1) This invention utilizes the high-precision measurement data of the charging pile itself to independently estimate the SOC, SOH and optimal charging curve in real time. Its output results serve as the gold standard for scheduling decisions, transforming the scheduling system from passively receiving potentially distorted BMS instructions to actively sensing and verifying the real needs of vehicles. The scheduling system can accurately predict the estimated charging time of each vehicle, thereby achieving true on-demand scheduling, maximizing parking space turnover, identifying battery health, implementing differentiated and protective charging strategies, extending battery life, and providing car owners with more accurate remaining power and charging time estimates than the vehicle's instrument panel, greatly improving user satisfaction and trust.

[0015] (2) This invention integrates discrete charging piles, vehicles, power grid nodes and order information into an organic site ecological knowledge network through data fusion and dynamic knowledge graph, enabling scheduling decisions to have a global perspective and deep reasoning ability. At the same time, each step of the decision-making process leaves a logical chain, forming a traceable report, which solves the trust problem of AI black box. In particular, the system's built-in reinforcement learning and online knowledge evolution mechanism enables it to continuously iterate and optimize the scheduling strategy based on historical and real-time feedback data, and has self-learning ability. The microservice architecture and standardized interface design decouple the core intelligent module of the system from the specific scenario configuration. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the modules of the intelligent management and scheduling system for charging pile groups of the present invention; Figure 2 This is a schematic diagram of the battery state joint estimation algorithm engine module in the intelligent management and scheduling system for charging pile groups of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0018] Example: Please refer to Figure 1-2 The intelligent management and scheduling system for charging pile groups includes: a real-time monitoring data acquisition module connected to each charging pile in the charging pile group, used to collect and upload charging pile operation status data and high-precision charging process electrical parameter data in real time. The charging pile operation status data includes the charging pile number, geographical location, type, power, and current status. The high-precision charging process electrical parameter data includes the voltage, current, and temperature time sequence data measured in real time by the charging pile for the connected vehicle during the charging process. The battery state joint estimation algorithm engine module is connected to the real-time monitoring data acquisition module to receive high-precision charging process electrical parameter data. It has a built-in joint estimation algorithm based on the fusion of electrochemical model and data-driven approach. Independent of the battery management system of the connected vehicle, it estimates and outputs the core state parameters of the power battery of the connected vehicle in real time. The core state parameters of the power battery include real-time true state of charge, battery health status, and optimal charging curve prediction based on the current state. The intelligent scheduling decision module connects to both the real-time monitoring data acquisition module and the battery state joint estimation algorithm engine module. It receives the real-time status of the charging pile group, requests from vehicles waiting to be charged, and predictions of core battery state parameters and charging curves. Based on a multi-objective optimization algorithm, this module uses core battery state parameters and charging curve predictions as key decision-making criteria to dynamically generate scheduling schemes. The scheduling schemes include: matching the most suitable charging pile to vehicles waiting to be charged, recommending the optimal charging power and time, and dynamically redistributing power to already charged vehicles. The charging control and order management module connects the intelligent scheduling decision module and each charging pile. It is used to execute scheduling plans, generate charging orders, and send control instructions containing power and time parameters to the designated charging pile to start or adjust the charging process. At the same time, it manages the entire process of order payment, settlement and exception handling. The data management and analysis module stores all historical data, including charging pile operation data, vehicle battery status estimation data, scheduling decision logs, and order data. It also analyzes and optimizes the model parameters of the joint battery status estimation algorithm engine and the scheduling strategy of the intelligent scheduling decision module.

[0019] In a specific embodiment of the present invention, this architecture completely changes the passive mode of traditional charging pile groups operating independently or simply relying on BMS commands. By introducing an independent battery state estimation engine, the system obtains the gold standard data required for scheduling decisions, thereby fundamentally avoiding scheduling errors such as long charging sessions being interrupted or unreasonable power allocation caused by inaccurate BMS data. It realizes a paradigm upgrade from vehicle-pile communication matching to precise resource scheduling based on the actual needs of the battery, significantly improving parking space turnover rate, overall operational efficiency, and user satisfaction.

[0020] Specifically, the battery state joint estimation algorithm engine module includes a data preprocessing and feature extraction unit, an online model parameter identification and correction unit, a multi-state joint estimation unit, and a charging curve dynamic prediction unit. The data preprocessing and feature extraction unit is used to receive and process real-time charging process time-series data measured by the charging pile. The time-series data includes terminal voltage. Charging current and battery surface temperature Meanwhile, the unit filters and denoises the raw data, removes outliers, and extracts feature vectors for state estimation. The feature vectors include incremental capacity IC curve features, voltage relaxation curve features, and statistical features at multiple time scales. The online model parameter identification and correction unit is connected to the data preprocessing and feature extraction unit. It is used to identify the key parameters of the battery equivalent circuit model online and in real time based on the processed data. This unit includes a state observer and a parameter observer. Among them, state observer Used to track the internal state of a system as it changes over time, the state variables typically include the state of charge, terminal voltage, and polarization voltage; parameter observer. This is used for online identification and updating of key parameters in the battery equivalent circuit model. Key parameters include ohmic internal resistance, polarization resistance, and polarization capacitance. The collaborative process of the state observer and parameter observer filter is as follows: The parameter observer first predicts the parameters at the current time based on the parameter estimate value at the previous time step. The state observer uses the parameter value predicted in the previous step, combined with the system input, to predict the system state at the current time step. The state observer updates the state prediction using the actual measurement value to obtain the optimal state estimate and calculates the state prediction error. The parameter observer updates the parameter prediction using the state prediction value provided by the state observer and the actual measurement value to obtain the optimal parameter estimate. The multi-state joint estimation unit is connected to both the data preprocessing and feature extraction unit and the online model parameter identification and correction unit. It is used to receive feature vectors and updated battery model parameters. This unit integrates multiple heterogeneous joint estimation algorithms and outputs preliminary estimates of the battery's state of charge, health state and internal temperature state respectively. It also uses a fusion weighting algorithm based on improved fuzzy entropy to dynamically assign and fuse multiple preliminary estimates and output a high-precision final joint estimation result. The charging curve dynamic prediction unit is connected to the multi-state joint estimation unit. It takes the SOC and SOH in the final joint estimation result as the core input, combines the battery's historical charging data, and uses a data-driven model to predict the optimal charging current-voltage curve from the initial SOC to the target SOC in the current state and the accurate estimated charging time. The estimated charging time is used as the core decision basis for the charging pile group scheduling system.

[0021] In a specific embodiment of the invention, the data preprocessing unit first cleans and extracts effective features from the original charging process data. The online model parameter identification unit, through a dual-filter architecture, tracks the time-varying parameters inside the battery in real time, ensuring that the model always fits the current aging and temperature state of the battery. The multi-state joint estimation unit integrates the preliminary results of multiple heterogeneous algorithms and intelligently identifies and integrates the optimal estimate through an improved fuzzy entropy weighting method, ensuring the robustness and accuracy of the results. The charging curve dynamic prediction unit, based on high-precision state estimation, predicts the optimal charging path and time in the future. This modular design decomposes the complex battery state estimation problem into manageable and optimizable sub-tasks, constructing an adaptive, high-precision battery digital twin that does not rely on a BMS. This not only provides the scheduling system with a reliable key decision input of the estimated charging time, but it can also serve as an independent value-added service, enhancing the system's technical barriers and commercial value, and is the cornerstone for solving the core pain points of current charging scheduling.

[0022] Specifically, the online model parameter identification and correction unit adopts a parameter identifier based on the extended Kalman filter (EKF). Its state-space model includes the following state equations: The observation equation is ,in This represents the battery model parameter vector at time k. Including ohmic internal resistance Polarization resistance Polarized capacitors The open-circuit voltage (OCV) model parameters are given in units of ohms (Ω), farads (F), and volts (V). The terminal voltage observation at time k is expressed in volts (V). For the nonlinear observation function based on the Thevenin model, The SOC reference value at time k is initially provided by the ampere-hour integration method. and These are process noise and observation noise, respectively. The EKF filter iteratively calculates the Kalman gain and updates the parameter vector in real time. The estimated value is used to adapt to battery aging and temperature changes.

[0023] In a specific embodiment of the present invention, the parameter identification problem of the battery equivalent circuit model is transformed into a state estimation problem. It takes the model parameter vector as the state to be estimated and the terminal voltage measurement value as the observation, establishes a nonlinear state-space model, and after linearization processing, extends the Kalman filter algorithm to calculate the Kalman gain in real time through predictive recursive iteration, thereby optimally correcting the estimated value of the parameter vector so that it dynamically follows the actual changes of battery ohmic internal resistance, polarization parameters, etc. with aging and temperature.

[0024] Specifically, the multi-state joint estimation unit integrates multiple heterogeneous joint estimation algorithms, including: A joint SOC-SOH estimator based on Extended Kalman Filter (EKF); A joint SOC-SOH estimator based on unscented Kalman filter (UKF); A state estimator using a Long Short-Term Memory (LSTM) network.

[0025] In specific embodiments of the present invention, a single algorithm path is abandoned. The EKF-based estimator utilizes the model's advantages to handle nonlinear systems, the UKF-based estimator does not require linearization and is better at estimating strongly nonlinear systems, and the LSTM-based data-driven estimator excels at learning complex patterns from historical data. The three algorithms perform preliminary estimations of battery state from three different dimensions: model-driven, statistical learning, and data-driven, respectively, forming a complementary relationship. The integration of heterogeneous algorithms significantly improves the robustness and fault tolerance of the system. When the performance of one algorithm degrades due to model mismatch, abnormal noise, or data quality issues, other algorithms can provide supplementation or correction. This architecture avoids the risk of the entire system collapsing due to the failure of a single algorithm, ensuring that the system can always output relatively stable and reliable state estimation results in highly dynamic and uncertain real-world vehicle charging environments, thereby enhancing the system's engineering practicality and reliability.

[0026] Specifically, the fusion weight calculation steps of the fusion weighting algorithm based on improved fuzzy entropy are as follows: Calculate the residual of the terminal voltage estimation for each estimator at time k. The unit is volts (V). Build with The membership function is used as the input, and the fuzzy entropy corresponding to each estimator is calculated. This is used to measure the uncertainty of the estimator's output; Set adjustment factor and create a judgment vector. ,when When this happens, the corresponding output of the estimator is considered bad data and suppressed in the weight calculation; Calculate each estimator for the state Fusion weights of (SOC or SOH) , The final fusion estimate is calculated using the following formula: ,in This is the initial estimate for the m-th estimator; The charging curve dynamic prediction unit employs a long short-term memory network model based on an attention mechanism. The input sequence of this model consists of the joint estimated state sequence of historical and current time steps, the charging current sequence, the voltage sequence, and the battery model code. The output of the model is the optimal charging current sequence for multiple future time steps. Expected charging time The calculation formula is: ,in The target state of charge is 100% or a user-defined value. This represents the final fused SOC estimate at the current moment. The actual usable capacity of the current battery is obtained by multiplying the final fused SOH estimate by the rated capacity, in ampere-hours (Ah). This is the average value of the optimal charging current sequence predicted by the attention long short-term memory network model, in amperes.

[0027] In a specific embodiment of the present invention, the fusion weighted algorithm first calculates the voltage residual output of each estimator, quantifies its uncertainty through fuzzy entropy, and sets a threshold to suppress bad data. The weights are dynamically allocated according to the uncertainty; the lower the uncertainty, the higher the weight. Finally, intelligent weighted fusion is achieved. The charging curve prediction adopts the Attention-LSTM model, whose attention mechanism can dynamically focus on the most critical moment in the historical sequence for predicting the future charging current, thereby more accurately generating the optimal current sequence that balances speed and battery life, and calculating the accurate charging time accordingly.

[0028] Specifically, the intelligent scheduling decision module includes a data fusion and knowledge construction unit, a multi-objective optimization decision unit, and an adaptive strategy execution and evaluation unit that are connected in sequence to form a closed loop; The data fusion and knowledge construction unit is used to receive and fuse multi-source heterogeneous real-time data and historical data, and to construct and dynamically update the scheduling domain knowledge graph. The multi-objective optimization decision-making unit is connected to the data fusion and knowledge construction unit. Based on the semantic information and real-time status of the knowledge graph, it calls the hybrid algorithm library to perform multi-objective collaborative optimization calculations and generate the optimal or near-optimal scheduling strategy set. The adaptive strategy execution and evaluation unit is connected to the multi-objective optimization decision-making unit. It is used to distribute the scheduling strategy to the execution layer and perform online reinforcement learning and iterative optimization of the decision-making model and knowledge graph based on the reward signals and key performance indicators (KPIs) from environmental feedback.

[0029] In a specific embodiment of the present invention, the module mimics the closed loop of an intelligent agent that learns through perception, thinking, and action. The data fusion and knowledge construction unit integrates multi-source information to form a systematic domain knowledge graph, thereby completing situational awareness. The multi-objective optimization decision-making unit, based on the knowledge graph, calls a hybrid algorithm library to solve for the optimal solution set that satisfies multiple constraints such as cost, efficiency, and user experience, thereby completing strategy thinking. The adaptive strategy execution and evaluation unit executes the strategy and collects feedback. Through reinforcement learning, the decision-making model and knowledge graph are continuously optimized to achieve action and evolution.

[0030] Specifically, the data fusion and knowledge construction unit includes a multi-model data governance subunit, a dynamic knowledge graph construction subunit, and an interpretable feature extraction subunit; The multi-model data governance subunit is used to clean, standardize, and correlate and integrate the input power grid flow, equipment status, traffic conditions, and order demand data to establish a data pool with a unified spatiotemporal benchmark. The dynamic knowledge graph construction subunit, based on expert rules and data-driven approaches, represents entities and their relationships in a graph-based manner and utilizes a graph neural network (GNN) model to dynamically mine potential associations and state evolution patterns. Its graph update is represented as an incremental learning process, where entities are substations, charging piles, production lines, and vehicles, and relationships are connections, memberships, or loads. The interpretable feature extraction subunit employs deep learning model technology based on attention mechanism to quantify the contribution of each input feature to subsequent decision-making and generate a decision basis report.

[0031] In a specific embodiment of the present invention, the multi-model data governance subunit aligns heterogeneous data such as power grid, transportation, equipment, and orders in the spatiotemporal dimension to form a high-quality data pool. The dynamic knowledge graph construction subunit performs graph-based modeling of entities such as charging piles, vehicles, and power grid nodes, as well as their electrical connections, affiliations, load effects, and other relationships, and uses graph neural networks to mine deep-level association patterns.

[0032] Specifically, the multi-objective optimization decision-making unit includes a hybrid algorithm library, a multi-objective solution and trade-off sub-unit, and a real-time dynamic rescheduling trigger; This hybrid algorithm library integrates heuristic and metaheuristic algorithms, reinforcement learning algorithms, and prediction and sequence modeling algorithms. The heuristic and metaheuristic algorithms are used to quickly obtain high-quality feasible solutions. The heuristic and metaheuristic algorithms form a genetic algorithm (GA), where the fitness function can be defined as a weighted sum of multiple objectives. ,in These represent the objective function values ​​for cost, time, and energy consumption, respectively. For dynamically adjustable weights, reinforcement learning algorithms are used to learn long-term optimal policies in dynamic environments, including Deep Q-Networks (DQN). The core Q-learning update formula is: ,in Let $\begin{case}{a}{a}$ be the expected value of the cumulative reward per unit of state action value function. The state at time t, For the action to be taken, For immediate rewards, the unit's scalar value is related to the target. For learning rate, The dimensionless discount factor is used to weigh current and future rewards. Prediction and sequence modeling algorithms are used to predict load, demand, and ETA, including time series-based neural networks. Their output formula can be expressed as... ,in The predicted value for time t. Given historical sequence data, W and b are the weight matrix and bias vector, respectively. For activation functions; The multi-objective solution and trade-off sub-unit uses the Pareto optimal front search algorithm to generate a non-dominated solution set and provides a visual trade-off analysis through a human-computer interaction interface for decision-makers to select the final solution; The real-time dynamic rescheduling trigger is based on an event-driven mechanism. When a key state variable is detected to deviate from a preset threshold, it automatically triggers a local or global recalculation of the decision-making process. The driving mechanisms include new order insertion, equipment failure, and traffic congestion.

[0033] In specific embodiments of this invention, the hybrid algorithm library selects the optimal algorithm for different characteristics of the scheduling problem. Metaheuristic algorithms such as genetic algorithms excel at global search in the solution space, quickly obtaining high-quality feasible solutions. Reinforcement learning algorithms learn long-term optimal strategies under dynamic uncertainty conditions through continuous interaction with the environment, while sequence prediction algorithms accurately predict future loads and demands. Pareto front solutions visualize the multi-objective optimization results for decision-makers to weigh, and event-driven rescheduling ensures the system can respond to emergencies in real time.

[0034] Specifically, the adaptive policy execution and evaluation unit includes a policy distribution and verification subunit, a multi-dimensional feedback monitoring subunit, and an online learning and model update subunit; The strategy distribution and verification subunit converts the optimized scheduling instructions into control instructions and issues them after pre-verification of security and feasibility based on a digital twin or simulation environment. The multi-dimensional feedback monitoring subunit collects system feedback data in real time after the strategy is executed. The system feedback data includes physical status, business indicators and economic indicators. The physical status is voltage and location, the business indicators are fulfillment rate and waiting time, and the economic indicators are cost and revenue. The online learning and model update subunit collects feedback data and continuously updates the parameters of the value network or policy network in the decision-making model through a reinforcement learning framework. At the same time, successful disposal cases and revised policies are used as new knowledge samples and integrated into the dynamic knowledge graph through case reasoning methods to achieve the self-evolution of the knowledge base.

[0035] In a specific embodiment of the present invention, simulation verification is performed using digital twins before strategy distribution to ensure the safety and feasibility of instructions. After execution, the multi-dimensional feedback monitoring subunit comprehensively collects effect data from the physical, business, and economic layers. The online learning subunit uses this feedback as reward signals for reinforcement learning, driving the update of neural network parameters in the decision-making model. At the same time, successful handling cases are abstracted into knowledge and integrated into a dynamic knowledge graph through case reasoning, continuously enriching the system's experience and realizing a complete value loop from decision execution to decision execution learning optimization. Digital twin pre-verification significantly reduces the risk of safety accidents or equipment damage caused by instruction errors. Online learning based on multi-dimensional feedback makes the system no longer a fixed black box after deployment. Its scheduling strategy can be continuously iterated and optimized with the feedback of actual operational data, becoming more intelligent with use. The evolution of the knowledge graph allows the system to accumulate and reuse experience in handling various abnormal working conditions, significantly improving its ability to cope with complex scenarios.

[0036] Specifically, the collaborative working mechanism of the intelligent scheduling decision module is as follows: From feature extraction and knowledge reasoning to optimization and solution, each step retains and outputs readable intermediate results and logical chains, forming a complete decision tracing report; The upper-level multi-objective optimization decision-making unit handles strategic-level resource allocation and long-term planning, while the lower-level adaptive strategy execution and evaluation unit handles tactical-level real-time adjustments and anomaly recovery. The two layers collaborate through a shared knowledge graph. The module adopts a microservice architecture. Its core units and algorithm library are defined through standardized interfaces. By replacing the domain-specific knowledge graph architecture and objective function, it can be quickly adapted to different application scenarios such as power dispatching, transportation and logistics, intelligent manufacturing, and charging pile management.

[0037] In a specific embodiment of the present invention, the system ensures that every step from data to decision is traceable through an interpretable pipeline, forming a logical chain. It adopts a global-local two-layer optimization, with the upper layer performing long-term resource planning across sites and the lower layer handling real-time scheduling within a single site. The two collaborate through a shared knowledge graph state, microservice architecture, and standardized interface design, which decouples the core algorithm unit from domain-specific configurations.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concept, should be covered within the scope of protection of the present invention.

Claims

1. A smart management and scheduling system for charging pile groups, characterized in that, include: The system includes a real-time monitoring data acquisition module, a battery state joint estimation algorithm engine module, an intelligent scheduling decision module, a charging control and order management module, and a data management and analysis module. The real-time monitoring data acquisition module is connected to each charging pile in the charging pile group and is used to collect and upload charging pile operation status data and high-precision charging process electrical parameter data in real time. The battery state joint estimation algorithm engine module receives the high-precision charging process electrical parameter data. It has a built-in joint estimation algorithm based on the fusion of electrochemical model and data-driven approach to estimate and output the core state parameters of the power battery of the currently connected vehicle in real time. The intelligent scheduling decision module is used to receive the real-time status of the charging pile group, the requests of vehicles waiting to be charged, and the core state parameters of the battery and the prediction of the charging curve. Based on a multi-objective optimization algorithm, the module dynamically generates a scheduling scheme with the core state parameters of the battery and the prediction of the charging curve as the key decision basis. The charging control and order management module connects the intelligent scheduling decision module and each charging pile. It is used to execute scheduling plans, generate charging orders, and send control commands to designated charging piles to start or adjust the charging process. At the same time, it manages the entire process of order payment, settlement and exception handling. The data management and analysis module is used to store all historical data and to optimize the model parameters of the battery state joint estimation algorithm engine and the scheduling strategy of the intelligent scheduling decision module through analysis.

2. The intelligent management and scheduling system for charging pile groups according to claim 1, characterized in that, The battery state joint estimation algorithm engine module specifically includes a data preprocessing and feature extraction unit, a model parameter online identification and correction unit, a multi-state joint estimation unit, and a charging curve dynamic prediction unit. The data preprocessing and feature extraction unit is used to receive and process real-time charging process time-series data measured by the charging pile, the time-series data including terminal voltage. Charging current and battery surface temperature Meanwhile, the unit filters and denoises the raw data, removes outliers, and extracts feature vectors for state estimation. The feature vectors include incremental capacity IC curve features, voltage relaxation curve features, and statistical features at multiple time scales. The online model parameter identification and correction unit is connected to the data preprocessing and feature extraction unit, and is used to identify key parameters of the battery equivalent circuit model online and in real time based on the processed data. This unit includes a state observer and a parameter observer. The state observer The parameter observer is used to track the internal state of a system as it changes over time. These state variables typically include state of charge, terminal voltage, and polarization voltage. This is used for online identification and updating of key parameters in the battery equivalent circuit model. The key parameters include ohmic internal resistance, polarization resistance, and polarization capacitance. The collaborative process of the state observer and parameter observer filter is as follows: the parameter observer first predicts the parameters at the current time based on the parameter estimate value at the previous time step; the state observer uses the parameter value predicted in the previous step, combined with the system input, to predict the system state at the current time step; the state observer updates the state prediction using the actual measurement value to obtain the optimal state estimate and calculates the state prediction error; the parameter observer updates the parameter prediction using the state prediction value provided by the state observer and the actual measurement value to obtain the optimal parameter estimate. The multi-state joint estimation unit is connected to both the data preprocessing and feature extraction unit and the online model parameter identification and correction unit. It is used to receive the feature vector and the updated battery model parameters. This unit integrates multiple heterogeneous joint estimation algorithms, outputs preliminary estimates of the battery's state of charge, health state and internal temperature state, and uses a fusion weighting algorithm based on improved fuzzy entropy to dynamically allocate and fuse multiple preliminary estimates, outputting a high-precision final joint estimation result. The dynamic prediction unit for the charging curve is connected to the multi-state joint estimation unit. It takes the SOC and SOH in the final joint estimation result as the core input, combines the battery's historical charging data, and uses a data-driven model to predict the optimal charging current-voltage curve from the initial SOC to the target SOC and the accurate estimated charging time in the current state. The estimated charging time is used as the core decision-making basis of the charging pile group scheduling system.

3. The intelligent management and scheduling system for charging pile groups according to claim 2, characterized in that, The online model parameter identification and correction unit specifically adopts a parameter identifier based on the Extended Kalman Filter (EKF). Its state-space model includes the following state equations: The observation equation is ,in This represents the battery model parameter vector at time k. Including ohmic internal resistance Polarization resistance Polarized capacitors The open-circuit voltage (OCV) model parameters are given in units of ohms (Ω), farads (F), and volts (V). The terminal voltage observation at time k is expressed in volts (V). For the nonlinear observation function based on the Thevenin model, The SOC reference value at time k is initially provided by the ampere-hour integration method. and These are process noise and observation noise, respectively. The EKF filter iteratively calculates the Kalman gain and updates the parameter vector in real time. The estimated value is used to adapt to battery aging and temperature changes.

4. The intelligent management and scheduling system for charging pile groups according to claim 2, characterized in that, The multi-state joint estimation unit integrates multiple heterogeneous joint estimation algorithms, including: A joint SOC-SOH estimator based on Extended Kalman Filter (EKF); A joint SOC-SOH estimator based on unscented Kalman filter (UKF); A state estimator using a Long Short-Term Memory (LSTM) network.

5. The intelligent management and scheduling system for charging pile groups according to claim 2, characterized in that, The fusion weighting algorithm based on improved fuzzy entropy has the following steps for calculating the fusion weights: Calculate the residual of the terminal voltage estimation for each estimator at time k. The unit is volts (V). Build with The membership function is used as the input, and the fuzzy entropy corresponding to each estimator is calculated. This is used to measure the uncertainty of the estimator's output; Set adjustment factor and create a judgment vector. ,when When this happens, the corresponding output of the estimator is considered bad data and suppressed in the weight calculation; Calculate each estimator for the state Fusion weights of (SOC or SOH) , The final fusion estimate is calculated using the following formula: ,in This is the initial estimate for the m-th estimator; The dynamic prediction unit for the charging curve employs a long short-term memory network model based on an attention mechanism. The input sequence of this model consists of a joint estimated state sequence from historical and current times, a charging current sequence, a voltage sequence, and a battery model code. The output of the model is the optimal charging current sequence for multiple future time steps. The estimated filling time The calculation formula is ,in The target state of charge is 100% or a user-defined value. This represents the final fused SOC estimate at the current moment. The actual usable capacity of the current battery is obtained by multiplying the final fused SOH estimate by the rated capacity, in ampere-hours (Ah). The value is the average of the optimal charging current sequence predicted by the attention long short-term memory network model, in amperes.

6. The intelligent management and scheduling system for charging pile groups according to claim 1, characterized in that, The intelligent scheduling decision module includes a data fusion and knowledge construction unit, a multi-objective optimization decision unit, and an adaptive strategy execution and evaluation unit that are connected in sequence and form a closed loop. The data fusion and knowledge construction unit is used to receive and fuse multi-source heterogeneous real-time data and historical data, and to construct and dynamically update the scheduling domain knowledge graph. The multi-objective optimization decision unit is connected to the data fusion and knowledge construction unit, and is used to call the hybrid algorithm library to perform multi-objective collaborative optimization calculations based on the semantic information and real-time status of the knowledge graph, and generate the optimal or near-optimal scheduling strategy set. The adaptive strategy execution and evaluation unit is connected to the multi-objective optimization decision-making unit. It is used to distribute the scheduling strategy to the execution layer and perform online reinforcement learning and iterative optimization of the decision model and knowledge graph based on the reward signals and key performance indicators (KPIs) from the environmental feedback.

7. The intelligent management and scheduling system for charging pile groups according to claim 6, characterized in that, The data fusion and knowledge construction unit includes a multi-model data governance subunit, a dynamic knowledge graph construction subunit, and an interpretable feature extraction subunit; The multi-model data governance subunit is used to clean, standardize, and correlate and fuse the input power grid flow, equipment status, traffic conditions, and order demand data to establish a data pool with a unified spatiotemporal benchmark. The dynamic knowledge graph construction subunit is based on expert rules and data-driven approach. It represents entities and their relationships in a graph form and uses a graph neural network (GNN) model to dynamically mine potential associations and state evolution patterns. Its graph update is represented as an incremental learning process, where entities are substations, charging piles, production lines and vehicles, and relationships are connections, memberships or loads. The interpretable feature extraction subunit employs deep learning model technology based on attention mechanism to quantify the contribution of each input feature to subsequent decision-making and generate a decision basis report.

8. The intelligent management and scheduling system for charging pile groups according to claim 6, characterized in that, The multi-objective optimization decision-making unit includes a hybrid algorithm library, a multi-objective solution and trade-off subunit, and a real-time dynamic rescheduling trigger. The hybrid algorithm library integrates heuristic and metaheuristic algorithms, reinforcement learning algorithms, and prediction and sequence modeling algorithms. The heuristic and metaheuristic algorithms are used to quickly obtain high-quality feasible solutions. These heuristic and metaheuristic algorithms are genetic algorithms (GA), and the fitness function of a genetic algorithm can be defined as a weighted sum of multiple objectives. ,in These represent the objective function values ​​for cost, time, and energy consumption, respectively. The weights are dynamically adjustable. The reinforcement learning algorithm, including Deep Q-Network (DQN), is used to learn long-term optimal policies in dynamic environments. The core Q-learning update formula is: ,in Let $\state\action\value\value$ be the expected value of the cumulative reward per unit. The state at time t, For the action to be taken, For immediate rewards, the unit's scalar value is related to the target. For learning rate, The discount factor is dimensionless and used to weigh current and future rewards. The prediction and sequence modeling algorithms are used for predicting load, demand, and ETA, including time-series-based neural networks, and their output formula can be expressed as follows: ,in The predicted value for time t. Given historical sequence data, W and b are the weight matrix and bias vector, respectively. For activation functions; The multi-objective solution and trade-off subunit uses the Pareto optimal front search algorithm to generate a non-dominated solution set and provides a visual trade-off analysis through a human-computer interaction interface for decision-makers to select the final solution. The real-time dynamic rescheduling trigger is based on an event-driven mechanism. When a key state variable is detected to deviate from a preset threshold, it automatically triggers a local or global recalculation of the decision-making process. The driving mechanism includes new order insertion, equipment failure, and traffic congestion.

9. The intelligent management and scheduling system for charging pile groups according to claim 6, characterized in that, The adaptive strategy execution and evaluation unit includes a strategy distribution and verification subunit, a multi-dimensional feedback monitoring subunit, and an online learning and model update subunit. The strategy distribution and verification subunit converts the optimized scheduling instructions into control instructions, and then issues them after pre-verification of security and feasibility based on a digital twin or simulation environment. The multi-dimensional feedback monitoring subunit collects system feedback data in real time after the strategy is executed. The system feedback data includes physical status, business indicators and economic indicators. The physical status is voltage and location, the business indicators are fulfillment rate and waiting time, and the economic indicators are cost and revenue. The online learning and model update subunit collects feedback data and continuously updates the parameters of the value network or policy network in the decision model through a reinforcement learning framework. At the same time, successful disposal cases and corrected policies are used as new knowledge samples and integrated into the dynamic knowledge graph through case reasoning methods to achieve the self-evolution of the knowledge base.

10. The intelligent management and scheduling system for charging pile groups according to claim 6, characterized in that, The collaborative working mechanism of the intelligent scheduling decision module is as follows: From feature extraction and knowledge reasoning to optimization and solution, each step retains and outputs readable intermediate results and logical chains, forming a complete decision tracing report; The upper-level multi-objective optimization decision-making unit handles strategic-level resource allocation and long-term planning, while the lower-level adaptive strategy execution and evaluation unit handles tactical-level real-time adjustments and anomaly recovery. The two layers collaborate through a shared knowledge graph. The module adopts a microservice architecture. Its core units and algorithm library are defined through standardized interfaces. By replacing the domain-specific knowledge graph architecture and objective function, it can be quickly adapted to different application scenarios such as power dispatching, transportation and logistics, intelligent manufacturing, and charging pile management.

Images