Battery management system based on adaptive digital twinning

Abstract

The invention relates to the technical field of BMS and the like, and provides a battery management system based on adaptive digital twinning, a physical layer of the battery management system comprises a battery pack, a sensor network and an edge computing node, and is responsible for data acquisition and preprocessing; the digital twin layer comprises a self-adaptive multi-scale model and a real-time data engine, battery behaviors are dynamically simulated by coupling electrochemical, thermal and aging models, a future state trajectory prediction result is output, and the real-time data engine fuses sensor data, historical data and simulation data to drive model updating; the intelligent decision-making layer comprises a reinforcement learning controller and a fault prediction module which are deployed in a local server, the reinforcement learning controller dynamically optimizes a charging and discharging strategy according to a prediction result of the digital twinborn layer and issues and executes the charging and discharging strategy, and the fault prediction module analyzes multi-source time sequence data based on an LSTM network so as to early warn thermal runaway and short circuit risks in advance. According to the invention, long-term accurate mapping and adaptive adjustment between the battery physical entity and the digital model can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of BMS, digital twin and reinforcement learning, and particularly relates to a battery management system based on adaptive digital twin. BACKGROUND

[0002] In the traditional battery management system (BMS), SOC estimation adopts an equivalent circuit model (ECM), which is difficult to adapt to dynamic characteristics such as battery aging and temperature changes, and the error will increase with the use time. With the development of digital twin (DT) technology, the application of digital twin (DT) in the battery management system (BMS) has attracted more and more attention. However, existing DT technology usually focuses on constructing pre-calibrated DT for state estimation and prediction. These DT technologies lack the ability to dynamically adapt to battery aging changes and changing operating environments, thus limiting their effectiveness in intelligent decision-making and making it difficult to improve the performance of BMS. Moreover, existing DT technology mostly uses offline modeling, which cannot respond to sudden changes in battery state (such as lithium precipitation caused by fast charging) in real time. In state estimation and prediction, the production parameters, historical operating condition data and real-time sensor data are not effectively fused, resulting in limited prediction accuracy.

[0003] Therefore, there is an urgent need in the art to develop a battery management system based on adaptive digital twin to ensure accurate long-term mapping between the physical entity of the battery and the digital twin model, and to adaptively adjust the digital twin model as the BMS develops, thereby maintaining SOC balance in different situations. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a battery management system based on adaptive digital twin to ensure accurate long-term mapping between the physical entity of the battery and the digital twin model, and to adaptively adjust the digital twin model as the BMS develops, thereby maintaining SOC balance in different situations.

[0005] The battery management system based on adaptive digital twin provided by the present application comprises: a physical layer comprising a battery pack, a sensor network for collecting battery voltage, temperature, current and impedance data of the battery pack, and an edge computing node for data preprocessing; a digital twin layer comprising an adaptive multi-scale model and a real-time data engine; the adaptive multi-scale model is a multi-physical field model coupling an electrochemical model, a thermal model and an aging model, used for dynamically simulating battery behavior and outputting prediction results containing future state trajectories of the battery; the real-time data engine is used to fuse real-time data from the sensors of the physical layer, historical operating condition data stored in the edge computing node and battery production parameters, and simulation data generated by the adaptive multi-scale model, to drive the adaptive multi-scale model to update dynamically; The intelligent decision layer comprises a reinforcement learning controller and a fault prediction module deployed on a local server; the reinforcement learning controller is in communication connection with the digital twin layer and is configured to dynamically optimize the charging and discharging strategy according to the prediction result output by the digital twin layer, and send the optimized charging and discharging strategy to the physical layer for execution; the fault prediction module is in communication connection with the digital twin layer and the physical layer, and is configured to analyze the time series data from the digital twin layer and the physical layer based on an LSTM network to give an early warning of thermal runaway and short circuit risk.

[0006] Compared with the prior art, the present application has the following beneficial effects: The present application provides a battery management system based on adaptive digital twin, which comprises: a physical layer comprising a battery pack, a sensor network for collecting battery voltage, temperature, current and impedance data of the battery pack, and an edge computing node for data preprocessing; a digital twin layer comprising an adaptive multi-scale model and a real-time data engine; the adaptive multi-scale model is a multi-physical field model coupling an electrochemical model, a thermal model and an aging model, which is used to dynamically simulate battery behavior and output prediction results containing future state trajectory of the battery; the real-time data engine is used to fuse real-time data from the sensor, historical working condition data and battery production parameters stored in the edge computing node, and simulation data generated by the adaptive multi-scale model, to drive the adaptive multi-scale model to update dynamically; an intelligent decision layer comprising a reinforcement learning controller and a fault prediction module deployed on a local server; the reinforcement learning controller is in communication connection with the digital twin layer and is configured to dynamically optimize the charging and discharging strategy according to the prediction result output by the digital twin layer, and send the optimized charging and discharging strategy to the physical layer for execution; the fault prediction module is in communication connection with the digital twin layer and the physical layer, and is configured to analyze the time series data from the digital twin layer and the physical layer based on an LSTM network to give an early warning of thermal runaway and short circuit risk. The present application can ensure accurate long-term mapping between the battery physical entity and the digital twin model, adaptively adjust the digital twin model with the development of the BMS, and achieve SOC balance under different conditions. BRIEF DESCRIPTION OF DRAWINGS

[0007] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings, which are not necessarily drawn to scale, like reference numerals describe similar components throughout the several views. It should be understood that the Figures are merely meant to be illustrative and that: Fig. 1 is a schematic diagram of an architecture of a battery management system based on adaptive digital twinning according to an embodiment of the present application; Fig. 2 is a flowchart of a scenario-aware particle swarm optimization algorithm according to an embodiment of the present application. DETAILED DESCRIPTION

[0008] In order to enable persons skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below 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, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by persons skilled in the art without creative labor should be within the scope of protection of the present application.

[0009] Referring to Figs. 1-2 , the present application provides a battery management system based on adaptive digital twinning, comprising: a physical layer comprising a battery pack, a sensor network for collecting battery voltage, temperature, current and impedance data of the battery pack, and an edge computing node for data preprocessing; a digital twinning layer comprising an adaptive multi-scale model and a real-time data engine; the adaptive multi-scale model is a multi-physical field model coupling electrochemical model, thermal model and aging model, used for dynamically simulating battery behavior and outputting prediction results containing future state trajectory of the battery; the real-time data engine is used to fuse real-time data from the sensor of the physical layer, historical working condition data and battery production parameters stored in the edge computing node, and simulation data generated by the adaptive multi-scale model, to drive the adaptive multi-scale model to update dynamically; an intelligent decision-making layer comprising a reinforcement learning controller and a fault prediction module deployed on a local server; the reinforcement learning controller is in communication connection with the digital twinning layer and is configured to dynamically optimize the charging and discharging strategy according to the prediction results output by the digital twinning layer, and send the optimized charging and discharging strategy to the physical layer for execution; the fault prediction module is in communication connection with the digital twinning layer and the physical layer, and is configured to analyze the time series data from the digital twinning layer and the physical layer based on an LSTM network, to early warn thermal runaway and short circuit risk.

[0010] In this embodiment, the adaptive multi-scale model coupled with the electrochemical model, thermal model and aging model can solve the problem that the traditional equivalent circuit model is difficult to adapt to the dynamic characteristics of battery aging and temperature changes, and realize high-fidelity dynamic simulation of the complex physical and chemical behavior of the battery, thereby improving the accuracy of state estimation and prediction from the root of the model. The real-time data engine is configured to fuse real-time sensor data from the physical layer, historical operating data and battery production parameters stored in the edge computing node, and simulation data generated by the adaptive multi-scale model, thereby solving the problem of ineffective fusion of production parameters, historical operating data and real-time sensor data, and realizing deep fusion and driving of multi-source heterogeneous data in the whole life cycle. The adaptive multi-scale model is driven for dynamic updating, thereby solving the problem of lack of dynamic adaptation to battery aging changes and changing operating environment, and the problem that offline modeling cannot respond to sudden changes in battery state in real time, realizing online self-correction and learning of the digital twin model, and ensuring long-term accurate mapping between the digital twin model and the physical entity. The reinforcement learning controller is configured to dynamically optimize the charging and discharging strategy according to the prediction results including future state trajectories output by the digital twin layer, thereby solving the problem that the effectiveness of existing digital twin technology in intelligent decision-making is limited, and realizing adaptive and optimal control based on forward-looking prediction, thereby improving system performance and life under the premise of ensuring safety. The fault prediction module is configured to analyze the time series data from the digital twin layer and the physical layer based on the LSTM network, and to provide early warning of thermal runaway and short circuit risk in advance, thereby solving the problem that the existing BMS cannot respond to sudden changes in battery state (such as lithium precipitation caused by fast charging) in real time, realizing early and proactive warning of potential serious faults, and enhancing the safety and reliability of the system.

[0011] It should be noted that the system can be applied to the fast charging scenario of an electric vehicle; in this scenario, the reinforcement learning controller is specifically configured to dynamically optimize a non-standard fast charging curve that first rises and then falls according to the maximum allowed charging current and the temperature trajectory of each cell predicted by the digital twin layer within the next five minutes; when the digital twin layer predicts that the temperature of any cell will exceed the safety warning threshold within the next three minutes, the reinforcement learning controller reduces the charging current and increases the cooling system power in advance to shorten the total charging time under the premise of ensuring safety.

[0012] It should be noted that the system can be applied to the energy storage station scenario on the grid side or the user side; in this scenario, the aging model of the digital twin layer is constructed and updated based on the historical operating data of the battery in the whole life cycle, and is combined with the future environmental temperature data provided by the weather forecast; the reinforcement learning controller dynamically adjusts the daily charging and discharging plan and power curve of the energy storage system according to the life prediction provided by the aging model and the future environmental temperature data, so as to slow down the aging rate of the battery pack.

[0013] Preferably, the adaptive multi-scale model dynamically adjusts its model parameters through an online parameter identification algorithm to track the characteristic drift of the battery due to aging and environmental changes; the online parameter identification algorithm is a scenario-aware particle swarm optimization algorithm, which is configured to define the current scenario by monitoring the state of charge, average temperature, health status and operating conditions of the battery, and call a pre-established mapping relationship library of scenarios and optimal parameters; the scenario-aware particle swarm optimization algorithm uses prior knowledge in the mapping relationship library to intelligently initialize the particle swarm representing the candidate parameter set at the beginning of optimization, to quickly and accurately identify key model parameters including ohmic resistance, polarization parameters and capacity.

[0014] In this embodiment, the scenario-aware particle swarm optimization algorithm is used as the online parameter identification algorithm, and is configured to define the current scenario by monitoring the state of charge, average temperature, health status and operating conditions of the battery, call a pre-established mapping relationship library of scenarios and optimal parameters, and use prior knowledge to intelligently initialize the particle swarm, which can quickly and accurately identify key model parameters such as ohmic resistance, polarization parameters and capacity, and effectively solve the problem of model mismatch when the traditional parameter identification method is used in battery aging and environmental changes. The operating conditions can include at least one of the charge and discharge current rate, the environmental temperature and the working mode.

[0015] Preferably, the operation process of the scenario-aware particle swarm optimization algorithm is configured to perform the following steps in sequence: continuously monitoring the battery state, starting the optimization process when the preset trigger condition is met; calculating the similarity of the current scenario with the historical scenarios in the mapping relationship library, and assigning the corresponding historical optimal parameter set to the particle swarm as the initial position according to the similarity result; iteratively updating the initialized particle swarm, and after the iteration meets the termination condition, updating the obtained global optimal parameter set to the adaptive multi-scale model, and learning and updating the corresponding relationship between the global optimal parameter set and the current scenario to the mapping relationship library.

[0016] In this embodiment, the running process of the scenario-aware particle swarm optimization algorithm is configured to sequentially perform the steps of continuously monitoring the defined scenario, starting the optimization when the condition is met, calculating the scenario similarity, assigning the historical optimal parameters as the initial position, iteratively updating, and updating the model and the mapping relationship library, which can realize complete closed-loop management of the parameter identification process, solve the problem of lack of systematic process guidance of traditional optimization algorithms, ensure the orderliness and reliability of algorithm execution, and avoid instability caused by random operation. In addition, intelligent initialization based on scenario similarity can improve search efficiency and enable the algorithm to quickly locate the high-quality solution area. At the same time, the optimization results are fed back to update the mapping relationship library, which can give the system continuous learning ability, so that the algorithm performance is continuously optimized as the running time increases. When calculating the similarity between the current scenario and the historical scenarios in the mapping relationship library, the similarity can be calculated by the Euclidean distance of the battery state parameters (including the state of charge, average temperature, health status, and operating condition).

[0017] Preferably, the iterative updating process of the particle swarm is configured to update the speed and position of each particle in each iteration according to a speed update formula containing the individual historical optimal position, the global historical optimal position of the population, and a scenario gravity item pointing to the optimal parameter set corresponding to the similar scenario.

[0018] In this embodiment, the iterative updating process of the particle swarm is configured to update the speed and position of each particle in each iteration according to a speed update formula containing the individual historical optimal position, the global historical optimal position of the population, and a scenario gravity item pointing to the optimal parameter set corresponding to the similar scenario, which can realize multiple guidance of the optimization search direction and solve the problems of traditional particle swarm algorithms, such as being easily trapped in local optimum and slow convergence speed. The scenario gravity item can be a vector whose direction points to the optimal parameter set corresponding to the most similar historical scenario to the current scenario, and whose size is proportional to the similarity.

[0019] Preferably, the preset trigger condition includes any one of a plurality of situations, and the plurality of situations include a change in the operating condition of the battery, reaching a preset periodic trigger time point, and a residual error between a voltage prediction value in the prediction result output by the adaptive multi-scale model and a physical layer sensor measured value continuously exceeding a set threshold.

[0020] In this embodiment, the preset trigger condition is configured to include any one of the following situations: battery operating condition switching, reaching a periodic trigger time point, and the residual error between the adaptive multi-scale model voltage prediction value and the physical layer sensor measured value continuously exceeding the set threshold. The intelligent triggering of parameter identification can be realized, the problem that the fixed period or manual triggering method cannot respond to system state changes in time can be solved, the timeliness of parameter identification can be ensured, the correction process can be started immediately when the model accuracy decreases, the adaptive triggering based on model accuracy can be realized, the intelligent level of the system can be improved, unnecessary waste of computing resources can be avoided, and the system operation efficiency can be optimized while ensuring the model accuracy. The operating condition switching can include that the charge and discharge current rate changes by more than 10% or the working mode is switched from charging to discharging.

[0021] Preferably, the system further comprises a high-precision state estimation module that fuses the voltage and current measurements of the sensors in the sensor network and the voltage prediction value in the prediction result output by the adaptive multi-scale model using an extended Kalman filter to estimate the state of charge of the battery in real time. In this embodiment, the high-precision state estimation module is configured to fuse the voltage and current measurements of the physical layer sensors and the voltage prediction value in the prediction result output by the adaptive multi-scale model using an extended Kalman filter, which can realize high-precision real-time estimation of the battery state of charge, solve the problems of error accumulation of the traditional ampere-hour integral method and limited estimation accuracy of the equivalent circuit model, fully utilize the accuracy of the measurement information and the priori of the model prediction, improve the estimation accuracy, improve the robustness of the estimation result, provide a reliable state information basis for precise battery management, and support the system to make more optimized control decisions.

[0022] Further, the scenario-aware particle swarm optimization algorithm ensures the long-term accurate mapping of the adaptive multi-scale model and the physical entity, provides an accurate model basis for the high-precision state estimation module, and at the same time provides a high-fidelity decision simulation environment for the reinforcement learning controller; the high-precision state estimation module provides reliable core state input for the enhanced state space of the reinforcement learning controller by providing accurate state of charge estimation; the reinforcement learning controller executes the optimization strategy based on the accurate state of charge and the high-fidelity future state trajectory, and the system operation data generated by the reinforcement learning controller in turn provides a data basis for the scenario-aware particle swarm optimization algorithm to continuously perform parameter identification.

[0023] In this embodiment, the overall scheme among the scenario-aware particle swarm optimization algorithm, the high-precision state estimation module and the reinforcement learning controller is constructed, wherein the scenario-aware particle swarm optimization algorithm ensures the long-term accurate mapping of the adaptive multi-scale model and the physical entity, provides an accurate model basis for the high-precision state estimation module, and provides a high-fidelity decision simulation environment for the reinforcement learning controller; the high-precision state estimation module provides reliable core state input for the enhanced state space of the reinforcement learning controller by providing accurate state of charge estimation; the reinforcement learning controller executes an optimization strategy based on the accurate state of charge and the high-fidelity future state trajectory, and the system operation data generated by the optimization strategy in turn provides a data basis for the scenario-aware particle swarm optimization algorithm to continuously perform parameter identification, so that a multi-module collaborative enhancement effect at the system level can be achieved. The scheme of this embodiment can solve the problem of isolated operation and lack of synergistic effect of each functional module of the traditional battery management system. In this embodiment, the deep fusion at the algorithm level can create a positive enhancement cycle - accurate models guarantee the accuracy of state estimation, accurate state input supports the reliability of intelligent decision-making, and new data generated by the optimization strategy further promotes the fine calibration of the model, forming a self-improving system ecology.

[0024] Preferably, the intelligent decision-making layer further comprises a battery equalization management module; the battery equalization management module is in communication connection with the digital twin layer and is configured to apply a model predictive control algorithm to predict the state changes of each single battery in the future for a plurality of steps based on the state of charge estimation values of each single battery in the battery pack provided by the digital twin layer, and calculate the optimal control signal for realizing the active equalization of the state of charge of the battery pack through rolling optimization.

[0025] In this embodiment, by configuring the battery equalization management module in the intelligent decision-making layer and making it in communication connection with the digital twin layer, applying the model predictive control algorithm, predicting the future state changes of each single battery based on the state of charge estimation values of each single battery in the battery pack provided by the digital twin layer, and calculating the optimal equalization control signal through rolling optimization, the intelligent active equalization of the state of charge of the battery pack can be realized, the overall service life of the battery pack can be effectively prolonged, and the problems of response lag and limited effect of the traditional equalization strategy can be solved. Among them, the feedforward control based on model prediction enables the system to foresee the future unbalanced trend and realize early intervention. Rolling optimization ensures the dynamic adaptability of the equalization strategy, which can adjust the control action according to the real-time state. In addition, the battery equalization management module can adjust the charge distribution of each single battery in the battery pack through the active equalization circuit according to the optimal control signal calculated by the model predictive control algorithm.

[0026] Preferably, the state space of the reinforcement learning controller is an augmented state space; the augmented state space contains the current instantaneous state physical quantities of the battery and fuses the future short-term and medium-term sustainable peak power, core temperature and state of charge change trajectories in the prediction results output by the digital twin layer; the action space of the reinforcement learning controller is at least one of the charge and discharge current set point, the discharge current limit and the cooling system power level; the reward function of the reinforcement learning controller is a multi-objective trade-off composite function, including at least the efficiency reward encouraging low internal resistance operation, the life reward punishing accelerated aging behavior, the safety reward severely punishing behavior close to or beyond the safety boundary, and the operation task reward ensuring the completion of basic charge and discharge functions.

[0027] In this embodiment, by configuring the state space of the reinforcement learning controller as an augmented state space containing the current instantaneous state physical quantities of the battery and fusing the future state trajectory output by the digital twin layer, configuring the action space as at least one of the charge and discharge current set point, the discharge current limit and the cooling system power level, and configuring the reward function as a multi-objective trade-off composite function, intelligent optimization of the charge and discharge strategy can be achieved, and the problem that traditional control methods are difficult to balance multiple objectives is solved. The design of the augmented state space gives the controller the ability to predict, enabling it to make current decisions based on future state trends. The multi-objective reward function ensures balanced optimization of the control strategy in multiple dimensions such as efficiency, life and safety. The future short-term and medium-term can refer to time ranges of 1-5 minutes and 5-30 minutes, respectively.

[0028] Preferably, the augmented state space of the reinforcement learning controller further contains the fault probability score and the predicted fault occurrence time output by the fault prediction module in real time as key safety state features; the safety reward term in the reward function is configured to be negatively related to the fault probability score and the predicted fault occurrence time, so as to drive the reinforcement learning controller to actively avoid risks, learn and execute preventive control strategies that can reduce the fault probability score and prolong the predicted fault occurrence time.

[0029] In this embodiment, by including the fault probability score and the predicted fault occurrence time output by the fault prediction module in real time in the augmented state space of the reinforcement learning controller, and configuring the safety reward term to be negatively related to these parameters, intelligent decision optimization under safety constraints can be achieved, and the problem that traditional reinforcement learning may take risky strategies in safety-critical scenarios is solved. The risk warning information is directly included in the decision-making process, which enables the controller to have risk perception ability. The negative correlation between the safety reward and the risk parameters can establish an effective safety constraint mechanism to drive the controller to actively avoid risks. The fault probability score and the predicted fault occurrence time can be calculated and output by the LSTM network of the fault prediction module based on multi-source time series data.

[0030] Preferably, the time series data analyzed by the fault prediction module includes: time series data from the physical layer, including direct measurements of voltage, current, temperature, and internal pressure of the battery pack; time series data from the digital twin layer, including state estimates of state of charge, state of health, peak power provided by the adaptive multi-scale model, and model internal states of polarization voltage, ohmic internal resistance; and residual sequences between measured values of sensors in the sensor network and corresponding predicted values of the adaptive multi-scale model, which are used as independent time series data sources to enhance the timeliness and accuracy of fault early warning.

[0031] In this embodiment, by configuring the time series data analyzed by the fault prediction module to include direct measurements from the physical layer, state estimates and model internal states from the digital twin layer, and residual sequences between measured values and predicted values, and explicitly using residual sequences as independent data sources to enhance early warning capabilities, precise early warning of multi-level fault feature fusion can be achieved, and the problem of insufficient reliability of single data source fault detection can be solved. Among them, multi-source data fusion can make full use of the accuracy of direct measurement, the deep information of model state and the sensitivity of residual sequence, and construct a comprehensive fault perception system. Residual sequences as effective carriers of early fault features can improve the timeliness and accuracy of fault detection.

[0032] It should be noted that the above embodiments are only preferred specific embodiments of the present application, and the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. The protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A battery management system based on adaptive digital twinning, characterized by, The application relates to a battery management system, comprising: a physical layer, including a battery pack, a sensor network for collecting battery voltage, temperature, current and impedance data of the battery pack, and an edge computing node for data preprocessing; a digital twin layer, including an adaptive multi-scale model and a real-time data engine; the adaptive multi-scale model is a multi-physical field model coupling an electrochemical model, a thermal model and an aging model, used for dynamically simulating battery behavior and outputting prediction results containing future state trajectories of the battery; the real-time data engine is used for fusing real-time data from the sensor network, historical working condition data and battery production parameters stored in the edge computing node and simulation data generated by the adaptive multi-scale model, so as to drive the adaptive multi-scale model to dynamically update; an intelligent decision-making layer, including a reinforcement learning controller and a fault prediction module deployed on a local server; the reinforcement learning controller is in communication connection with the digital twin layer and is configured to dynamically optimize a charging and discharging strategy according to the prediction results output by the digital twin layer and to send the optimized charging and discharging strategy to the physical layer for execution; the fault prediction module is in communication connection with the digital twin layer and the physical layer and is configured to analyze time series data from the digital twin layer and the physical layer based on an LSTM network to early warn thermal runaway and short circuit risks.

2. The system of claim 1, wherein, The adaptive multi-scale model dynamically adjusts model parameters thereof through an online parameter identification algorithm to track characteristic drift of the battery caused by aging and environmental changes; the online parameter identification algorithm is a scenario-aware particle swarm optimization algorithm, which is configured to define a current scenario by monitoring a state of charge, an average temperature, a health state and a working condition of the battery and to call a pre-established mapping relationship library of scenarios and optimal parameters; the scenario-aware particle swarm optimization algorithm uses prior knowledge in the mapping relationship library to intelligently initialize a particle swarm representing a candidate parameter set at the beginning of optimization, so as to quickly and accurately identify key model parameters including an ohmic resistance, polarization parameters and a capacity.

3. The system of claim 2, wherein, The operation process of the scenario-aware particle swarm optimization algorithm is configured to sequentially perform the following steps: continuously monitoring a battery state and starting an optimization process when a preset trigger condition is met; calculating a similarity of a current scenario and historical scenarios in the mapping relationship library and assigning corresponding historical optimal parameter sets to the particle swarm as initial positions according to the similarity results; iteratively updating the initialized particle swarm and, after iteration termination, updating a global optimal parameter set obtained to the adaptive multi-scale model and learning and updating a corresponding relationship between the global optimal parameter set and the current scenario to the mapping relationship library.

4. The system of claim 3, wherein, The iterative updating process of the particle swarm is configured to update the speed and position of each particle in each iteration according to a speed update formula containing an individual historical optimal position, a global historical optimal position of the population and a scenario gravity term pointing to an optimal parameter set corresponding to a similar scenario.

5. The system of claim 3, wherein, The preset triggering condition includes any one of a plurality of situations, and the plurality of situations include that a running condition of the battery changes, a preset periodic triggering time point is reached, and a residual error between a voltage prediction value in the prediction result output by the adaptive multi-scale model and a physically-layer sensor measured value continuously exceeds a set threshold.

6. The system of claim 2, wherein, The system further includes a high-precision state estimation module that fuses voltage and current measurement values of sensors in the sensor network and the voltage prediction value in the prediction result output by the adaptive multi-scale model to estimate the state of charge of the battery in real time.

7. The system of claim 1, wherein, The intelligent decision layer further includes a battery equalization management module; the battery equalization management module is in communication connection with the digital twin layer and is configured to apply a model predictive control algorithm to predict state changes of each monomer in the future for a plurality of steps based on state of charge estimation values of each monomer in the battery pack provided by the digital twin layer, and calculate optimal control signals for realizing active equalization of the state of charge of the battery pack through rolling optimization.

8. The system of claim 1, wherein, The state space of the reinforcement learning controller is an enhanced state space; the enhanced state space contains current instantaneous state physical quantities of the battery and fuses future short-term and medium-term sustainable peak power, core temperature and state of charge change trajectories in the prediction result output by the digital twin layer; the action space of the reinforcement learning controller is at least one of a charging and discharging current set point, a discharging current limit and a cooling system power level; the reward function of the reinforcement learning controller is a multi-objective weighted composite function, including at least an efficiency reward encouraging low internal resistance operation, a life reward punishing accelerated aging behavior, a safety reward severely punishing behavior close to or beyond the safety boundary, and an operation task reward ensuring completion of basic charging and discharging functions.

9. The system of claim 8, wherein, The enhanced state space of the reinforcement learning controller further contains a fault probability score and a predicted fault occurrence time output by the fault prediction module in real time as key safety state features; the safety reward item in the reward function is configured to be negatively correlated with the fault probability score and the predicted fault occurrence time, so as to drive the reinforcement learning controller to actively avoid risks, learn and execute preventive control strategies that can reduce the fault probability score and prolong the predicted fault occurrence time.

10. The system of claim 1, wherein, The time series data analyzed by the fault prediction module includes: time series data from the physical layer, including direct measurement values of voltage, current, temperature and battery pack internal pressure; time series data from the digital twin layer, including state estimation values of state of charge, health state and peak power provided by the adaptive multi-scale model, and model internal states of polarization voltage and ohmic internal resistance; and a residual error sequence between measured values of sensors in the sensor network and corresponding prediction values of the adaptive multi-scale model, which is used as an independent time series data source to enhance the early nature and accuracy of fault early warning.

Images