Lithium ion battery thermal management method based on physical perception and entropy collaborative multi-agent

Abstract

A lithium ion battery thermal management method based on physical perception and entropy collaborative multi-agent, comprising: establishing a lithium ion battery electro-thermal coupling model containing data-driven compensation, and constructing a full state space system equation describing the dynamic characteristics of the battery; a TCN-Transformer hybrid neural network is constructed by fusing a physical constraint mechanism, and the network is used for online identification of the parameters in the electro-thermal coupling model to obtain real-time parameters; based on the full state space system equation and the real-time parameters, a minimum error entropy self-adaptive extended Kalman filter optimized by an entropy collaborative multi-agent flexible Actor-Critic algorithm is constructed to jointly estimate the state of charge and the battery temperature of the battery; based on the joint estimation result, a TD3 deep reinforcement learning control algorithm embedded with a differentiable safety layer is constructed, and the algorithm is used to realize direct control of the adaptive thermal management of the lithium ion battery under the guidance of rules. The present application realizes accurate, effective and safe lithium ion battery thermal management control.

Description

A Lithium-ion Battery Thermal Management Method Based on Physical Perception and Entropy Collaborative Multi-Agent System Technical Field

[0001] This invention relates to the field of new energy vehicle battery thermal management system technology, specifically to a lithium-ion battery thermal management method based on physical perception and entropy collaborative multi-agent. Background Technology

[0002] Lithium-ion batteries, with their advantages of high energy density, high specific power, lightweight, and long cycle life, have become the mainstream choice for power batteries in new energy vehicles. However, existing lithium-ion battery thermal management technologies face the following bottlenecks when dealing with high-dynamic, strongly coupled operating conditions:

[0003] (1) The traditional second-order RC equivalent circuit model (ECM) has a large modeling error under dynamic operating conditions;

[0004] (2) Parameter identification relying on pure data lacks physical consistency constraints and is prone to outputting invalid parameters;

[0005] (3) Classical Kalman filtering is poorly robust to non-Gaussian noise in real-world data;

[0006] (4) When pursuing optimal performance, the control strategy based on reinforcement learning (RL) lacks differentiability constraints on the safety boundary, which poses a safety hazard.

[0007] The above problems prevent existing technologies from achieving accurate, effective, and safe thermal management control of lithium-ion batteries. Summary of the Invention

[0008] The purpose of this invention is to provide a lithium-ion battery thermal management method based on physical perception and entropy collaborative multi-agent, so as to achieve accurate, effective and safe lithium-ion battery thermal management control.

[0009] A lithium-ion battery thermal management method based on physical perception and entropy collaborative multi-agent system includes:

[0010] Step S1: Establish a lithium-ion battery electro-thermal coupling model including data-driven compensation, and construct a full-state-space system equation describing the dynamic characteristics of the battery.

[0011] Step S2: Construct a TCN-Transformer hybrid neural network that integrates physical constraint mechanisms. Use this hybrid neural network to identify the parameters in the electro-thermal coupling model online to obtain real-time parameters.

[0012] Step S3: Based on the full-state space system equations constructed in step S1 and the real-time parameters identified in step S2, a minimum error entropy adaptive extended Kalman filter optimized by the entropy-cooperative multi-agent flexible Actor-Critic algorithm is constructed to jointly estimate the state of charge and battery temperature of the battery.

[0013] Step S4: Based on the joint estimation results output in step S3, construct the TD3 deep reinforcement learning control algorithm with an embedded differentiable safety layer. This algorithm enables rule-guided adaptive thermal management direct control of lithium-ion batteries.

[0014] The lithium-ion battery thermal management method based on physical perception and entropy cooperative multi-agent provided by the present invention has the following beneficial effects:

[0015] 1. This invention innovatively constructs a physically constrained TCN-Transformer hybrid neural network for online identification of model parameters. By introducing a physical constraint layer and smoothness loss into the deep learning framework, it effectively solves the problems of traditional parameter identification relying on pure data-driven methods lacking physical consistency constraints and easily outputting invalid parameters. This invention ensures that the electrical model parameters (such as ohmic internal resistance, approximate capacitance, etc.) simulated under high dynamic conditions have strict physical consistency and numerical stability, significantly improving the modeling accuracy of the electro-thermal coupling model and laying the foundation for subsequent maximum state estimation.

[0016] 2. This invention constructs a minimum error entropy adaptive extended Kalman filter optimized by an entropy-coordinated multi-agent flexible Actor-Critic algorithm, realizing a unique dual entropy optimization mechanism. On the one hand, the maximum policy entropy mechanism of reinforcement learning ensures that the agent fully explores the noise covariance parameter, avoiding local optimization. On the other hand, the minimum error entropy effectively suppresses the influence of non-Gaussian noise and measurement outliers, breaking through the dependence of traditional Kalman filtering on the Gaussian noise assumption, significantly enhancing the robustness of the joint estimation of battery state of charge (SOC) and battery temperature (SOT), and significantly improving the estimation accuracy under dynamic conditions compared with traditional methods.

[0017] 3. This invention constructs a TD3 deep reinforcement learning control algorithm with an embedded differentiable safety layer, innovatively solving the gradient interruption problem faced by traditional reinforcement learning when dealing with safety constraints. By introducing a differentiable safety layer, the mapping relationship between the system state and the safety boundary can be actively learned while maintaining continuous gradient propagation, thereby achieving smooth, unadjusted active derating control when operating close to the safety boundary. This invention achieves the optimal balance between safety and control efficiency while ensuring absolute battery thermal safety, avoiding the lag of traditional rule-based control and the dangerous exploratory behavior of traditional reinforcement learning, ultimately achieving accurate, effective, and safe thermal management control for lithium-ion batteries. Attached Figure Description

[0018] Figure 1 is a flowchart illustrating the lithium-ion battery thermal management method based on physical perception and entropy collaborative multi-agent provided by the present invention.

[0019] Figure 2 is a comparison of the estimation of ohmic internal resistance under dynamic load conditions by the method proposed in this invention with the traditional CNN-LSTM algorithm and the actual value.

[0020] Figure 3 is a comparison of the SOT estimation by the proposed method under non-Gaussian noise interference with the standard EKF algorithm and the true value;

[0021] Figure 4 is a comparison of the thermal management control effect of the proposed method on safety constraints with the standard TD3 algorithm and safety boundary. Detailed Implementation

[0022] To facilitate understanding of the present invention, a more complete description will be given below with reference to various embodiments. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0024] Please refer to Figure 1. An embodiment of the present invention provides a lithium-ion battery thermal management method based on physical perception and entropy collaborative multi-agent, including steps S1-S4:

[0025] Step S1: Establish a lithium-ion battery electro-thermal coupling model including data-driven compensation, and construct a full-state-space system equation describing the dynamic characteristics of the battery.

[0026] Specifically, step S1 includes steps S101-S104:

[0027] Step S101: Based on the principle of second-order RC equivalent circuit, a terminal voltage output model of lithium-ion battery is constructed, and Gaussian process regression is introduced as a data-driven compensation term to correct the unmodeled dynamic residuals that cannot be described by the physical model. The terminal voltage output model constitutes the output equation of the full state space system.

[0028] The expression for the terminal voltage output model of a lithium-ion battery is as follows:

[0029]

[0030] in, for The terminal voltage of the battery at any given time; To and State of charge at time t and Core temperature of moment The relevant open-circuit voltage; for Load current at any given moment; The internal resistance in ohms varies with temperature; and These are two parallel RC loops in the second-order RC equivalent circuit model. Polarization voltage at time; This is the Gaussian process regression term used to compensate for the voltage residual. The input feature vector;

[0031] covariance kernel function A quadratic exponential kernel function is used. Used to describe the correlation between input features The expression is:

[0032]

[0033] in, For signal variance, The total number of dimensions of the input features. For feature dimensions The characteristic length, and They are vectors and In the Dimensional component values.

[0034] Step S102: Construct an internal state space model of the battery's electrical and thermal characteristics. The internal state space model includes an electrical state model, a battery electro-thermal coupling model, and a thermal state space model.

[0035] The battery state of charge (SOC) is updated using the ampere-hour integral method, and the expression for the electrical state model is as follows:

[0036]

[0037] in, for State of charge at time t, For Coulomb efficiency, For time step, This refers to the battery's rated capacity.

[0038] To describe the internal temperature distribution of a battery, a battery electro-thermal coupling model is constructed. The key to this model is to achieve strong coupling between the electrical state and the thermal state by calculating the total heat generation rate inside the battery. The expression of the battery electro-thermal coupling model is as follows:

[0039]

[0040] in, for The total heat generation rate of the battery at any given time. It is the sum of the ohmic resistance and the polarization resistance. It is the entropy heat coefficient;

[0041] Based on the thermodynamic principle of energy conservation, the expression for the thermal state-space model is:

[0042]

[0043]

[0044] in, For the core heat capacity of the battery, The total heat generation rate inside the battery. Battery core temperature rate of change over time The surface temperature of the battery. The thermal resistance from the battery core to the surface. For the surface heat capacity of the battery, The rate of change of battery surface temperature over time. For ambient temperature, The thermal resistance for convective heat transfer from the battery surface to the environment. Cooling power applied to the thermal management system.

[0045] Step S103: Integrate the above electrical state model, battery electro-thermal coupling model, and thermal state space model to define the full state vector of the system.

[0046] Among them, the full state vector The expression is:

[0047]

[0048]

[0049]

[0050] in, For electrical state vectors, This is the thermal state vector;

[0051] In step S104, the expression for the full state-space system equations is:

[0052]

[0053] in, and They are respectively Time and The joint state vector at time t, It is a nonlinear state transition function. for The system input vector at time t, For those obtained through online identification The model parameter vector at time step, This is the system process noise vector. For the system observation vector, For the observation function, For measuring noise.

[0054] In step S104, by combining the output equation in step S101 and the full state vector in step S103, the full state space system equation describing the dynamic characteristics of the battery is finally constructed.

[0055] Step S2: Construct a TCN-Transformer hybrid neural network that integrates physical constraint mechanisms. Use this hybrid neural network to identify the parameters in the electro-thermal coupling model online to obtain real-time parameters.

[0056] Specifically, step S2 includes steps S201-S202:

[0057] Step S201: Construct a TCN layer (Temporal Convolutional Network) as the front-end module of the hybrid neural network to extract local temporal dependencies in lithium-ion battery operating data. The TCN layer extracts features from the input sequence through convolution operations, and finally outputs sequence features containing local dynamic characteristics. ;

[0058] Wherein, step S201 satisfies the following formula:

[0059]

[0060] in, For a moment The convolutional output features, Given an input time series vector, Indicates to Execution expansion rate dilated convolution operation, For convolution kernel function, The first convolution kernel Each weighting coefficient The kernel size is the convolution kernel size. For a moment The input element.

[0061] Step S202, output sequence features The input is fed into a Transformer encoder, which uses a multi-head self-attention mechanism to capture long-distance dependencies in the sequence. Physical constraints are embedded in the final output layer. The Softplus activation function is used to construct a non-negative constraint mapping so that the output corresponds to the time-varying parameter vector in the electro-thermal coupling model, thereby realizing online identification of the parameters in the electro-thermal coupling model and obtaining real-time parameters.

[0062] Among them, the multi-head self-attention mechanism extracts temporal features from different representation subspaces by using multiple self-attention heads in parallel, thereby realizing the mapping and aggregation from local dynamic features to global temporal features.

[0063] To ensure that the parameters output by the neural network satisfy the prior knowledge of the electro-thermal coupling physical model, physical constraints are embedded in the final output layer of the network. Specifically, a non-negative constraint mapping is constructed using the Softplus activation function, so that the output corresponds to the time-varying parameter vector in the electro-thermal coupling model.

[0064] This step enables an interpretable mapping and output from the black-box feature representation of a neural network to the parameters of a white-box model with explicit physical meaning.

[0065] In this embodiment, a total loss function is constructed that integrates data fitting error, parameter smoothness constraints, and model regularization. This is to drive the hybrid neural network to learn and approximate the real physical laws of batteries during training. The total loss function of the hybrid neural network is... for:

[0066]

[0067]

[0068] in, , and These are the weight coefficients for the data fitting loss, smoothing loss, and regularization term, respectively. This represents the measured vector of the battery terminal voltage. Based on the current estimated parameters The calculated model-predicted voltage vector, This is an L2 regularization term used to prevent network overfitting and improve the model's generalization performance. This is a smoothness loss, used to ensure that parameter changes are smooth; For those obtained through online identification The model parameter vector at time step, The total length of the observation sequence. This is the square operation for the L2 norm.

[0069] By optimizing the total loss function While accurately fitting the data, the network outputs physical parameters that conform to time-varying smooth prior knowledge, thus achieving collaborative training of data-driven and physical model constraints.

[0070] Step S3: Based on the full-state space system equations constructed in step S1 and the real-time parameters identified in step S2, a minimum error entropy adaptive extended Kalman filter optimized by the entropy-cooperative multi-agent flexible Actor-Critic algorithm is constructed to jointly estimate the battery's state of charge and battery temperature.

[0071] Specifically, step S3 includes steps S301-S308:

[0072] Step S301, based on the nonlinear state transition function constructed in step S1 ,use Posterior state estimate at time 1 and system input vector ,calculate Predicted value of prior state at time 1 :

[0073]

[0074] Simultaneously, calculate the Jacobian matrix of the nonlinear state transition function at the current estimation point. :

[0075]

[0076] in, For the system input vector, Indicates partial derivative;

[0077] Step S302: In the entropy-cooperative multi-agent flexible Actor-Critic algorithm, the first agent, Agent-Q, is introduced to adaptively model the process noise. Agent-Q outputs a dynamically adjusted process noise covariance matrix based on the current system state, input, and historical information. , Specifically, it is based on vectors The dynamic process noise covariance matrix is ​​constructed using the main diagonal elements, combined with the Jacobian matrix. and posterior estimation error covariance matrix at time 1 The prior covariance prediction equation involving agent decision-making is derived, and its expression is:

[0078]

[0079] in, for The posterior estimation error covariance matrix at time 1. Indicates transpose. This is the adaptive action vector output by Agent-Q in response to process noise characteristics;

[0080] This step, by introducing reinforcement learning actions, enables adaptive adjustment of uncertainty during model prediction. Agent-Q's proactive decision-making achieves dynamic online compensation for system model uncertainty.

[0081] Step S303, calculate the innovation vector, the expression is:

[0082]

[0083] in, for The information vector at time;

[0084] Step S304: Introducing the second agent, Agent-R, from the entropy-cooperative multi-agent flexible Actor-Critic algorithm to adaptively model the measurement noise. Agent-R outputs a dynamically adjusted measurement noise covariance matrix based on the current system state, measurement information, and historical data. Calculate the Jacobian matrix of the observation function at the current prior estimate point. Then the Kalman gain was calculated. :

[0085]

[0086] in, For adaptive correction matrix, This is the dynamic measurement noise variance action vector output by Agent-R;

[0087] This gain incorporates Agent-R's dynamic assessment of measurement uncertainty, enhancing the system's robustness to anomalous observations.

[0088] Step S305: Update the posterior state based on Agent-R and minimum error entropy, where the Kalman gain is used to update the system state, and the posterior state estimation result is output.

[0089]

[0090] in, for The posterior state estimate at time t;

[0091] Simultaneously, update the posterior error covariance matrix:

[0092]

[0093] in, Let be the posterior error covariance matrix. It is an identity matrix.

[0094] Through the above steps, Agent-R's active decision-making, combined with the Kalman filter framework, enables accurate and robust estimation of battery state under complex operating conditions and uncertain measurement environments.

[0095] Step S306: Construct a shared reward function to simultaneously guide Agent-Q and Agent-R in balancing state estimation accuracy and policy exploration capability during the optimization process. The expression for the reward function is:

[0096]

[0097] in, for The collaborative reward value at any given moment. and These are the weighting coefficients for the mean squared error term and the entropy term, respectively. The information potential of the residual;

[0098] This reward function encourages the agent to actively maintain a certain degree of policy randomness while minimizing the estimation error in order to explore better noise covariance adjustment strategies.

[0099] Step S307: Based on the maximum entropy reinforcement learning framework, update the policy networks of Agent-Q and Agent-R simultaneously using the above reward function. The optimization objective is to maximize the weighted sum of the accumulated expected reward and the policy entropy, expressed as:

[0100]

[0101] in, The target value of the strategy to be maximized is... For mathematical expectation operators, For parameters The fitted soft Q-value function, For temperature coefficient, For parameter-based The action probability distribution output by the policy network. For the control state vector, , for The surface temperature of the battery at any given time. for The ambient temperature at that moment;

[0102] By introducing an entropy regularization term into the expression for the policy objective value The agent can be encouraged to maintain the randomness of its strategy while optimizing the cumulative reward, thereby maintaining sufficient exploration capabilities under complex and time-varying battery conditions and avoiding policy updates from getting stuck in local optima.

[0103] In step S308, after Agent-Q and Agent-R are updated, the state of charge and battery temperature of the battery are jointly estimated.

[0104] Step S4: Based on the joint estimation results output in step S3, construct the TD3 deep reinforcement learning control algorithm with an embedded differentiable safety layer. This algorithm enables rule-guided adaptive thermal management direct control of lithium-ion batteries.

[0105] Specifically, step S4 includes steps S401-S404:

[0106] Step S401: Define the state space of the thermal management control system. With action space ,in, , , for The cooling power at any given moment is measured in real time. and output Construct a closed-loop thermal management and control environment to form a complete Markov decision-making process;

[0107] Step S402: At the output of the Actor network of the TD3 algorithm, a differentiable safety layer based on Lagrange multipliers is embedded. Through this differentiable safety layer, the original action output by the hybrid neural network is mapped to the physical feasible domain that satisfies the safety constraints.

[0108] In the differentiable safety layer, first define the safety constraints for the battery core temperature:

[0109]

[0110] in, For safety constraint functions, Predicted by the electro-thermal coupling model The core temperature of the battery at any given time. The upper limit of safe temperature;

[0111] Secondly, a quadratic programming projection problem is constructed to find the feasible action that has the shortest Euclidean distance to the original action and satisfies the above safety constraints. By introducing the Lagrange multiplier method to solve this constrained optimization problem, an efficient analytical projection solution can be derived, expressed as:

[0112]

[0113]

[0114] in, For the revised safety procedures, The original action output by the Actor network. For optimal Lagrange multipliers, This is the gradient operator for cooling power.

[0115] This differentiable safety layer ensures that the cooling power output by the control strategy will never physically cause the battery core temperature to exceed the safety limit. Simultaneously, the entire projection process is differentiable with respect to the parameters of the policy network, allowing gradients to pass smoothly through the safety layer during backpropagation, thus enabling safe-constrained training of the policy network.

[0116] Step S403: Construct a comprehensive reward function based on the differentiable security layer, and use a dual Critic network structure to evaluate the value of state-action pairs, while introducing a target policy smoothing mechanism during the evaluation process;

[0117] The comprehensive reward function satisfies the following formula:

[0118]

[0119] in, For control strategy in The overall rewards earned at any time For the optimal target operating temperature of the battery, for The energy consumption generated by constantly performing cooling actions. It is a linear rectified function. , , These are the weighting coefficients for the temperature tracking item, energy consumption item, and safety constraint item, respectively.

[0120] By designing a comprehensive reward function that includes a temperature tracking error term, a cooling energy consumption penalty term, and a safety constraint violation penalty term, the policy network can be guided to minimize cooling system energy consumption and strictly avoid temperature exceedances while maintaining the core temperature close to the optimal operating temperature.

[0121] In this embodiment, two identical but parameter-independent Critic networks are constructed to form a dual-Critic network structure, which calculates the value of a given state-action pair. When calculating the objective Q-value, the minimum of the two is used to improve policy stability and mitigate overestimation.

[0122] Furthermore, to improve the stability and robustness of the algorithm, this invention introduces a target policy smoothing mechanism to smooth the target Q value. The target policy smoothing mechanism satisfies the following equation:

[0123]

[0124] in, For the smoothed target action, This is the action truncation function. For deterministic actions that are output by the target execution network based on the state at the next time step and corrected by a differentiable safety layer, Added Gaussian noise, The cutoff threshold for Gaussian noise. , These are the minimum and maximum values ​​of the cooling power operating range, respectively.

[0125] Step S404 involves updating the evaluation network parameters by minimizing the soft Q-value loss of the dual Critic network structure, and updating the execution network parameters by maximizing the policy gradient to complete the algorithm iteration and finally train the lithium battery thermal management control strategy.

[0126] The performance of the proposed method for battery parameter calibration and SOT estimation under complex operating conditions was verified and compared with traditional deep learning and real values. Based on the MATLAB simulation platform, Figure 2 shows that the proposed method effectively mitigates the parameter oscillation phenomenon of the traditional CNN-LSTM algorithm under dynamic load conditions, significantly improving the dynamic tracking stability of the ohmic internal resistance gain. Figure 3 shows that the proposed method overcomes the interference of non-Gaussian noise (spiking), and its SOT estimation accuracy is improved by approximately 15% compared to the standard EKF algorithm, demonstrating strong robustness. Figure 4 compares the thermal management control of the TD3 deep reinforcement learning control algorithm with embedded differentiable safety layers using the proposed method with that of the standard TD3 algorithm. Figure 4 shows that compared to the temperature exceedance caused by the standard TD3 algorithm, the proposed method achieves adaptive smooth control near the safety threshold, resulting in higher system safety.

[0127] The lithium-ion battery thermal management method based on physical perception and entropy collaborative multi-agent described above has the following beneficial effects:

[0128] 1. This invention innovatively constructs a physically constrained TCN-Transformer hybrid neural network for online identification of model parameters. By introducing a physical constraint layer and smoothness loss into the deep learning framework, it effectively solves the problems of traditional parameter identification relying on pure data-driven methods lacking physical consistency constraints and easily outputting invalid parameters. This invention ensures that the electrical model parameters (such as ohmic internal resistance, approximate capacitance, etc.) simulated under high dynamic conditions have strict physical consistency and numerical stability, significantly improving the modeling accuracy of the electro-thermal coupling model and laying the foundation for subsequent maximum state estimation.

[0129] 2. This invention constructs a minimum error entropy adaptive extended Kalman filter optimized by an entropy-coordinated multi-agent flexible Actor-Critic algorithm, realizing a unique dual entropy optimization mechanism. On the one hand, the maximum policy entropy mechanism of reinforcement learning ensures that the agent fully explores the noise covariance parameter, avoiding local optimization. On the other hand, the minimum error entropy effectively suppresses the influence of non-Gaussian noise and measurement outliers, breaking through the dependence of traditional Kalman filtering on the Gaussian noise assumption, significantly enhancing the robustness of the joint estimation of battery state of charge (SOC) and battery temperature (SOT), and significantly improving the estimation accuracy under dynamic conditions compared with traditional methods.

[0130] 3. This invention constructs a TD3 deep reinforcement learning control algorithm with an embedded differentiable safety layer, innovatively solving the gradient interruption problem faced by traditional reinforcement learning when dealing with safety constraints. By introducing a differentiable safety layer, the mapping relationship between the system state and the safety boundary can be actively learned while maintaining continuous gradient propagation, thereby achieving smooth, unadjusted active derating control when operating close to the safety boundary. This invention achieves the optimal balance between safety and control efficiency while ensuring absolute battery thermal safety, avoiding the lag of traditional rule-based control and the dangerous exploratory behavior of traditional reinforcement learning, ultimately achieving accurate, effective, and safe thermal management control for lithium-ion batteries.

Claims

1. A lithium-ion battery thermal management method based on physical perception and entropy collaborative multi-agent system, characterized in that, include: Step S1: Establish a lithium-ion battery electro-thermal coupling model including data-driven compensation, and construct a full-state-space system equation describing the dynamic characteristics of the battery. Step S2: Construct a TCN-Transformer hybrid neural network that integrates physical constraint mechanisms. This hybrid neural network is used to identify parameters in the electro-thermal coupling model online to obtain real-time parameters. Step S3: Based on the full-state-space system equations constructed in Step S1 and the real-time parameters identified in Step S2, construct a minimum error entropy adaptive extended Kalman filter optimized by the entropy-coordinated multi-agent flexible Actor-Critic algorithm. This jointly estimates the battery's state of charge and battery temperature. Step S4: Based on the joint estimation results output in Step S3, construct a TD3 deep reinforcement learning control algorithm with an embedded differentiable safety layer. This algorithm enables rule-guided adaptive thermal management direct control of the lithium-ion battery. Specifically, Step S3 includes: Step S301: Based on the nonlinear state transition function constructed in Step S1... ,use Posterior state estimate at time 1 and system input vector ,calculate Predicted value of prior state at time 1 ： in, For those obtained through online identification The model parameter vector at time step; simultaneously, the Jacobian matrix of the nonlinear state transition function at the current estimation point is calculated. ： in, For the system input vector, Indicates partial derivative, Given an input time series vector, The current estimated parameters are used; in step S302, the first agent, Agent-Q, in the entropy-cooperative multi-agent flexible Actor-Critic algorithm is introduced to adaptively model the process noise. Agent-Q outputs a dynamically adjusted process noise covariance matrix based on the current system state, input, and historical information. Combined with Jacobian matrix and posterior estimation error covariance matrix at time 1 The prior covariance prediction equation involving agent decision-making is derived, and its expression is: in, for The posterior estimation error covariance matrix at time 1. Indicates transpose. The Agent-Q outputs an adaptive action vector based on the process noise characteristics; Step S303, calculate the innovation vector, expressed as: in, for The new information vector at time, For the system observation vector, For the observation function, for The system input vector at time t, For those obtained through online identification The model parameter vector at time step; Step S304, introduce the second agent, Agent-R, in the entropy-cooperative multi-agent flexible Actor-Critic algorithm to adaptively model the measurement noise. Agent-R outputs a dynamically adjusted measurement noise covariance matrix based on the current system state, measurement information, and historical information. Calculate the Jacobian matrix of the observation function at the current prior estimate point. Then the Kalman gain was calculated. ： in, For adaptive correction matrix, The dynamic measurement noise variance action vector is output by Agent-R; Step S305, posterior state update is performed based on Agent-R and minimum error entropy, wherein the system state is updated using Kalman gain, and the posterior state estimation result is output: in, for The posterior state estimate at time step 1; simultaneously, the posterior error covariance matrix is ​​updated: in, Let be the posterior error covariance matrix. The identity matrix is ​​used; in step S306, a shared reward function is constructed to simultaneously guide Agent-Q and Agent-R in balancing state estimation accuracy and policy exploration capability during the optimization process. The expression of the reward function is: in, for The collaborative reward value at any given moment. and These are the weighting coefficients for the mean squared error term and the entropy term, respectively. The information potential of the residual; Step S307, based on the maximum entropy reinforcement learning framework, simultaneously update the policy networks of Agent-Q and Agent-R using the above reward function. The optimization objective is to maximize the weighted sum of the accumulated expected reward and the policy entropy, expressed as: in, The target value of the strategy to be maximized is... For mathematical expectation operators, For parameters The fitted soft Q-value function, For temperature coefficient, For parameter-based The action probability distribution output by the policy network. For the control state vector, ， for State of charge at time t, for The core temperature of every moment for The surface temperature of the battery at any given time. for Load current at any given time for The ambient temperature at that time; Step S308, after Agent-Q and Agent-R are updated, the state of charge of the battery and the battery temperature are jointly estimated.

2. The lithium-ion battery thermal management method based on physical perception and entropy collaborative multi-agent as described in claim 1, characterized in that, Step S1 specifically includes: Step S101, based on the principle of second-order RC equivalent circuit, constructing the terminal voltage output model of the lithium-ion battery, and introducing Gaussian process regression as a data-driven compensation term, wherein the terminal voltage output model constitutes the output equation of the full state space system; Step S102, constructing the internal state space model of the battery's electrical and thermal characteristics, wherein the internal state space model includes an electrical state model, a battery electro-thermal coupling model, and a thermal state space model; Step S103, integrating the above electrical state model, battery electro-thermal coupling model, and thermal state space model to define the full state vector of the system; Step S104, combining the output equation in Step S101 and the full state vector in Step S103, finally constructing the full state space system equation describing the dynamic characteristics of the battery.

3. The lithium-ion battery thermal management method based on physical perception and entropy collaborative multi-agent as described in claim 2, characterized in that, In step S101, the expression for the terminal voltage output model of the lithium-ion battery is: in, for The terminal voltage of the battery at any given time; To and State of charge at time t and Core temperature of moment The relevant open-circuit voltage; The internal resistance in ohms varies with temperature; and These are two parallel RC loops in the second-order RC equivalent circuit model. Polarization voltage at time; This is the Gaussian process regression term used to compensate for the voltage residual. The input feature vector; covariance kernel function A quadratic exponential kernel function is used. The expression is: in, For signal variance, The total number of dimensions of the input features. For feature dimensions The characteristic length, and They are vectors and In the The component values ​​of the dimension; in step S102, the expression of the electrical state model is: in, for State of charge at time t, For Coulomb efficiency, For time step, The rated capacity of the battery; the expression for the battery electro-thermal coupling model is: in, for The total heat generation rate of the battery at any given time. It is the sum of the ohmic resistance and the polarization resistance. Let be the entropy coefficient; the expression for the thermal state-space model is: in, For the core heat capacity of the battery, The total heat generation rate inside the battery. Battery core temperature rate of change over time The surface temperature of the battery. The thermal resistance from the battery core to the surface. For the surface heat capacity of the battery, The rate of change of battery surface temperature over time. For ambient temperature, The thermal resistance for convective heat transfer from the battery surface to the environment. Cooling power applied to the thermal management system.

4. The lithium-ion battery thermal management method based on physical perception and entropy collaborative multi-agent as described in claim 3, characterized in that, In step S103, the full state vector The expression is: in, For electrical state vectors, Let be the thermal state vector; in step S104, the expression for the full state-space system equations is: in, and They are respectively Time and The joint state vector at time t, It is a nonlinear state transition function. This is the system process noise vector. For measuring noise.

5. The lithium-ion battery thermal management method based on physical perception and entropy collaborative multi-agent as described in claim 4, characterized in that, Step S2 specifically includes: Step S201, constructing a TCN layer as the front-end module of a hybrid neural network to extract local temporal dependencies in lithium-ion battery operating data. The TCN layer extracts features from the input sequence through convolution operations, and finally outputs sequence features containing local dynamic characteristics. Step S202, output sequence features The input is fed into a Transformer encoder, which uses a multi-head self-attention mechanism to capture long-distance dependencies in the sequence. Physical constraints are embedded in the final output layer. The Softplus activation function is used to construct a non-negative constraint mapping so that the output corresponds to the time-varying parameter vector in the electro-thermal coupling model, thereby realizing online identification of the parameters in the electro-thermal coupling model and obtaining real-time parameters.

6. The lithium-ion battery thermal management method based on physical perception and entropy collaborative multi-agent as described in claim 5, characterized in that, Step S201 satisfies the following formula: in, For a moment The convolutional output features, Indicates to Execution expansion rate dilated convolution operation, For convolution kernel function, The first convolution kernel Each weighting coefficient The kernel size is the convolution kernel size. For a moment The input element.

7. The lithium-ion battery thermal management method based on physical perception and entropy collaborative multi-agent as described in claim 6, characterized in that, The total loss function of the hybrid neural network for: in, 、 and These are the weight coefficients for the data fitting loss, smoothing loss, and regularization term, respectively. This represents the measured vector of the battery terminal voltage. Based on the current estimated parameters The calculated model-predicted voltage vector, For L2 regularization terms, For smoothness loss, The total length of the observation sequence. This is the square operation for the L2 norm.

8. The lithium-ion battery thermal management method based on physical perception and entropy collaborative multi-agent as described in claim 7, characterized in that, Step S4 specifically includes: Step S401, defining the state space of the thermal management control system. With action space ,in, ， ， for The cooling power at any given moment is measured in real time. and output A closed-loop thermal management control environment is constructed, forming a complete Markov decision process; in step S402, a differentiable safety layer based on Lagrange multipliers is embedded at the output of the Actor network of the TD3 algorithm. This differentiable safety layer maps the original action output by the hybrid neural network to a physically feasible region that satisfies safety constraints. The differentiable safety layer satisfies the following equation: in, For safety constraint functions, Predicted by the electro-thermal coupling model The core temperature of the battery at any given time. The upper limit of safe temperature; in, For the revised safety procedures, The original action output by the Actor network. For optimal Lagrange multipliers, The gradient operator for cooling power is used; in step S403, a comprehensive reward function is constructed based on this differentiable safety layer, and a dual-Critic network structure is used to evaluate the value of state-action pairs. A target policy smoothing mechanism is introduced during the evaluation process. The comprehensive reward function satisfies the following equation: in, For control strategy in The overall rewards earned at any time For the optimal target operating temperature of the battery, for The energy consumption generated by constantly performing cooling actions. It is a linear rectified function. 、 、 These are the weighting coefficients for the temperature tracking term, energy consumption term, and safety constraint term, respectively; the target strategy smoothing mechanism satisfies the following equation: in, For the smoothed target action, This is the action truncation function. This refers to the deterministic action executed by the target based on the state at the next time step, after being corrected by a differentiable safety layer. Added Gaussian noise, The cutoff threshold for Gaussian noise. 、 These are the minimum and maximum values ​​of the cooling power action space, respectively; in step S404, the evaluation network parameters are updated by minimizing the soft Q-value loss of the dual Critic network structure, and the execution network parameters are updated by maximizing the policy gradient, thus completing the algorithm iteration and finally training the lithium battery thermal management control strategy.

Images