A lithium battery safety optimization fast charging control method

Abstract

The present invention relates to the technical field of lithium-ion batteries, and in particular to a method for controlling safe and optimized fast charging of lithium batteries, comprising the following steps: S1. constructing an electrothermal aging coupling model for lithium-ion batteries, determining the control equations of the model, and establishing an offline charging environment required for reinforcement learning based on the model; S2. initializing a policy network and two value networks, defining the reinforcement learning state space and action space, and setting a charging reward function and a cost function; S3. updating the state space variables according to the battery model in step S1, and adding the current state, charging action, reward value, cost value, and next-moment state to an experience pool. The present invention can realize fast charging with electrothermal perception and health awareness, reconstructs the expression form of multi-objective optimization problems to enhance the safety of the charging strategy, and at the same time reduces the computational complexity of online charging by adopting an offline training and online deployment mode, making it more efficient in actual charging applications.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium-ion batteries, and in particular to a safe, optimized, and fast-charging control method for lithium batteries. Background Art

[0002] Rechargeable lithium-ion batteries have brought convenience to everyday life, but the key to enabling this rechargeability is charging technology. Improving lithium-ion battery charging speeds has become a hot topic, as efforts to alleviate range anxiety and improve electric vehicle user satisfaction are underway. While rapid charging significantly reduces charging time, it violates safety limits and can lead to battery overheating, shortened lifespan, and even safety issues. Therefore, charging lithium-ion batteries is a challenging task, and efficient charging strategies are crucial for optimizing lithium-ion battery operation and extending their lifespan.

[0003] Traditional charging strategies include the widely used constant current constant voltage (CC-CV) charging protocol and its improved versions. However, these charging methods lack feedback on battery status and safety parameters, resulting in suboptimal charging performance. Existing technologies propose model-based feedback charging strategies. Battery models are primarily divided into electrochemical models and equivalent circuit models. Based on this, charging optimization algorithms are designed to solve the multi-objective optimization problem of minimizing charging time while ensuring compliance with charging safety constraints. To address the high computational complexity, high model dependency, and poor online traceability of traditional optimization algorithms, deep reinforcement learning algorithms are introduced.

[0004] Chinese patent publication number CN112018465A uses the DDPG algorithm to train a neural network used to determine battery charging current, ensuring that the battery completes the charging task while meeting physical charging constraints. However, this method relies on expert experience to design a reward model, which cannot guarantee the optimal charging strategy. Furthermore, the physical charging constraints are set as soft constraints, posing safety risks during the charging process. Therefore, we propose a safe and optimized fast charging control method for lithium batteries to address these issues. Summary of the Invention

[0005] The present invention aims to overcome the deficiencies of the prior art and proposes a lithium battery safe optimized fast charging control method, comprising the following steps:

[0006] S1. Build a coupled electrothermal aging model for lithium-ion batteries, determine the model's control equations, and establish the offline charging environment required for reinforcement learning based on the model.

[0007] S2. Initialize a policy network and two value networks, define the reinforcement learning state space and action space, and set the charging reward function and cost function;

[0008] S3. Update the state space variables based on the battery model in step S1, and add the current state, charging action, reward value, cost value, and next-moment state to the experience pool. Iteratively train the charging policy network, value assessment network, and cost assessment network constructed using the distributed soft actor-critic algorithm and conservative policy optimization techniques until convergence, obtaining the converged policy network model.

[0009] S4. Collect the battery's current terminal voltage, surface temperature, and ambient temperature data in real time, use the state observer to estimate the unmeasurable battery charge state and core temperature, input the processed state information into the charging strategy network trained in step S3, and determine the optimal charging current trajectory.

[0010] Preferably, the electrothermal aging coupling model of the lithium-ion battery in step S1 includes an equivalent circuit model for capturing the behavior of the battery at different temperatures, a thermodynamic model for characterizing the internal and surface temperatures of the battery, and an aging model for monitoring the health status of the battery.

[0011] Preferably, the control equation derivation process of the lithium-ion battery electrothermal aging coupling model in step S1 is as follows:

[0012] The physical behavior of the battery is described by a temperature-dependent second-order RC equivalent circuit model. According to Kirchhoff's law, the governing equation can be derived as:

[0013]

[0014] V t (t) = V ocυ (SoC)-V1(t)-V2(t)-R0(t,T)I(t)

[0015] Where, SoC, I, C n 、V t They represent the battery's state of charge, charging current, nominal capacity, and terminal voltage, respectively; d and t represent the differential operator and time, respectively; the open circuit voltage V ocυ is a polynomial function of SoC; V1 and V2 are the voltages of the two resistor-capacitor pairs, respectively; the ohmic internal resistance R0, polarization resistances R1 and R2, and polarization capacitances C1 and C2 are expressed as functions of SoC and temperature T; under the assumption of longitudinal uniformity, a two-state thermal model is used to capture the lumped thermal characteristics of the cylindrical battery, and the governing equation for the heat transfer process is:

[0016]

[0017] Where, T s 、T c and T fare the battery surface temperature, core temperature and ambient temperature respectively; R s and R c are heat convection resistance and heat conduction resistance respectively; C s and C c are the heat capacities of the battery surface and the core, respectively; the heat generation rate Q is given by the Bernardi heat generation model:

[0018] Q(t)=(V ocυ (SoC)-V t (t))I(t)+I(t)T m (t)E n (SoC, t)

[0019]

[0020] Where, E n is the entropy change in the electrochemical reaction; T m is the average temperature of the battery; in addition, the semi-empirical model of the Arrhenius equation is used to evaluate the battery capacity decay rate, and the charge throughput is used to quantify the capacity decay at different charge rates:

[0021]

[0022] Where Ah is the total discharge ampere-hour throughput; R g ,z,B,E a , c are the universal gas constant, power law factor, pre-exponential factor, activation energy and charge rate respectively; ΔSoH k is the battery health status degradation value corresponding to the kth sampling moment; Δt is the sampling interval.

[0023] Preferably, the implementation process of step S2 is as follows:

[0024] The state space and action space are defined as:

[0025] s t ={SoC(t),V t (t), T c (t)}

[0026] a t ={I(t)|I(t)∈[0,I max ]}

[0027] Where s t and a t Represent the state observed and the action taken at time step t; I maxis the maximum charging current constrained by the constraint; the charging strategy is formulated through a multi-objective optimization problem to minimize the charging time required to reach the desired SoC while ensuring that the charging constraints are met; the reward function is set to be only related to the charging speed:

[0028] r t =|SoC tar -SoC(t)|

[0029] Where r t represents the reward obtained at time step t; SoC tar is the expected SoC, indicating that the charging process is completed; the charging constraints are designed as cost functions respectively:

[0030]

[0031] C soh (t) = |ΔSoH t |

[0032] Where V t，max 、V t，min 、T c，max 、T c，min are the upper and lower cut-off voltages and the highest and lowest core temperature thresholds respectively; C υolt and C temp represents the penalty term for battery overvoltage and overtemperature, and C soh represents the aging of the battery; the total cost function at each step is described as:

[0033] c t =C υolt (t)+C temp (t)+C soh (t).

[0034] Preferably, the cost evaluation network implementation process in step S3 is as follows:

[0035] Rewrite the original constrained optimization problem, and the Lagrangian dual form of the problem is:

[0036]

[0037] Where A is the Lagrange multiplier; π is the strategy adopted; (s, a) is the state-action pair; and are the estimated values of the true reward and cost function respectively; d is the total constraint threshold; ρ π(s) represents the state distribution caused by the policy π; the underestimation of the cost function during the charging process is used to train other state-action pairs and accumulates into a large deviation in the temporal difference learning, resulting in charging behavior that violates the constraints; a conservative policy optimization technique is used to construct a cost evaluation network to solve the underestimation problem, and the upper confidence bound of the true cost value is used instead of the estimate, that is, the "mean + weighted standard deviation" of the set is calculated, which is expressed as:

[0038]

[0039] Where, is the upper confidence limit estimate of the cost; E represents the use of E bootstrap networks with the same network structure and independent training process Cost value estimation is performed; k is a conservative parameter; the cost evaluation network parameters are collectively referred to as ψ, and the parameters of each cost sub-network are updated using the time difference error method.

[0040] Preferably, the charging strategy network and the value assessment network in step S3 are implemented as follows:

[0041] Traditional reinforcement learning methods overestimate the reward value function during charging training, which will cause the charging speed to decrease. However, the value distribution learning method can significantly improve the overestimation problem of the reward value function, thereby reducing the instability of policy performance during the learning process.

[0042] The training phase is based on an actor-critic network framework, which includes a charging policy network and a value assessment network. The charging policy network takes as input the lithium battery state space variables and outputs the action value that controls charging. The value assessment network is used to evaluate the quality of the current charging policy. The parameters of the charging policy network are collectively referred to as φ, and the parameters of the value assessment network are collectively referred to as θ.

[0043] The distributed soft actor-critic algorithm assumes that the distribution of random rewards follows a Gaussian distribution, expressed as:

[0044]

[0045] Where, represents the value distribution of rewards; Q θ (s, a) and σ θ (s, a) represents the mean and standard deviation of the value distribution; is the standard operator of Gaussian distribution; the deviation size is dynamically adjusted by learning the variance of the distribution to achieve the effect of suppressing over-estimation; the value assessment network gradient consists of two parts: mean-related gradient and variance-related gradient, expressed as:

[0046]

[0047] Where, represents the gradient of the i-th evaluation network; is the gradient operator for the parameter θ; the clipping function C p (·) is used to alleviate gradient explosion or disappearance; b is the clipping boundary to alleviate the sensitivity of reward scale; and are the target minimum distribution network and the target minimum value network respectively; r and γ are the immediate reward and discount coefficient respectively; s′ and a′ are the state and action at the next moment; Z is the soft state-action reward; α is the weighting factor of the entropy term; the target network parameters of the charging strategy and value evaluation are collectively referred to as and Used to improve stability during training; use time difference error to update the value evaluation network parameter θ;

[0048] The safe charging problem is described as a solution to the optimal policy π * The strategy maximizes the long-term reward while limiting the cumulative cost, and the update target J of the policy network π Expressed as:

[0049]

[0050] In the formula, the symbol Indicates mean operation; the symbol ~ indicates random sampling based on probability distribution; is the playback buffer of sampled experience; the weighting factor α decays during training; π φ (a|s) represents the probability of action a occurring in state s; the charging strategy network is updated using the temporal difference error to maximize the reward target.

[0051] Preferably, the charging strategy deployment process in step S4 is as follows:

[0052] First, the state observer and the trained charging strategy network π φ Deploy it to the charging controller; set the charging target and constraint thresholds; in each sampling cycle, collect the current terminal voltage, surface temperature, and ambient temperature data of the battery, use the state observer to estimate the battery SoC and core temperature, input the state space variables into the deployed charging strategy network and obtain the output action, denormalize the action value to obtain the expected charging current, and input it into the battery environment until the charging process is complete.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] The present invention makes intelligent decisions on charging behavior based on key states, ensuring long battery life while balancing battery charging speed and physical constraint satisfaction. By rewriting the expression of multi-objective optimization problems and applying distributed reinforcement learning algorithms, the phenomenon of unstable strategy performance during the learning process is reduced, and the strategy performance after convergence is improved. The offline training and online deployment model is adopted to improve the online traceability of charging. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a diagram of an electrothermal coupling model of a cylindrical lithium battery in a lithium battery safety optimization and fast charging control method of the present invention;

[0056] Figure 2 A diagram showing the training process of a lithium battery charging strategy based on a distributed reinforcement learning algorithm in a lithium battery safety optimization and fast charging control method of the present invention;

[0057] Figure 3 This is a comparison diagram of the charging strategy proposed in the lithium battery safety optimization fast charging control method of the present invention and the traditional CCCV charging algorithm. DETAILED DESCRIPTION

[0058] The following description is intended to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are only examples, and those skilled in the art will appreciate other obvious variations.

[0059] like Figure 1-Figure 3 The method for controlling lithium battery fast charging with safety optimization is shown. First, the expression of the multi-objective charging optimization equation is rewritten to optimize the solution process of the safety constraint problem. A charging strategy network is trained to maximize the reward function and minimize the cost function. The trained charging strategy neural network is then used to intelligently control the charging current based on the current battery status information. This method meets the battery charging safety constraints while also satisfying the fast charging requirements and protecting the battery health. Specifically, the following steps are included:

[0060] Step 1: Construct a coupled electrothermal aging model for lithium-ion batteries, determine its control equations, and establish the offline charging environment required for reinforcement learning based on the model.

[0061] (1.1) Battery model

[0062] The present invention adopts Figure 1 The electrothermal aging coupled model shown is used to describe the characteristics of lithium-ion batteries, including an equivalent circuit model for capturing the behavior of the battery at different temperatures, a thermodynamic model for characterizing the internal and surface temperatures of the battery, and an aging model for monitoring the battery health status.

[0063] The physical behavior of the battery is described by a temperature-dependent second-order RC equivalent circuit model. According to Kirchhoff's law, the governing equation can be derived as:

[0064]

[0065] V t (t) = V ocυ (SoC)-V1(t)-V2(t)-R0(t,T)I(t)

[0066] Where, SoC, I, C n 、V t They represent the battery's state of charge, charging current, nominal capacity, and terminal voltage, respectively; d and t represent the differential operator and time, respectively; the open circuit voltage V ocυ is a polynomial function of SoC; V1 and V2 are the voltages of the two resistor-capacitor pairs, respectively; the ohmic internal resistance R0, polarization resistances R1 and R2, and polarization capacitances C1 and C2 are expressed as functions of SoC and temperature. Under the assumption of longitudinal uniformity, a two-state thermal model is used to capture the lumped thermal characteristics of the cylindrical battery. The governing equation for the heat transfer process is:

[0067]

[0068] Where, T s 、T c and T f are the battery surface temperature, core temperature and ambient temperature respectively; R s and R c are heat convection resistance and heat conduction resistance respectively; C s and C c are the heat capacities of the battery surface and the core, respectively; the heat rate S and Q are given by the Bernardi heat generation model:

[0069] Q(t)=(V ocυ (SoC)-V t (t))I(t)+I(t)T m (t)E n (SoC, t)

[0070]

[0071] Where, E n is the entropy change in the electrochemical reaction; T m is the average temperature of the battery. In addition, the semi-empirical model of the Arrhenius equation is used to evaluate the battery capacity decay rate, and the charge throughput is used to quantify the capacity decay at different charge rates:

[0072]

[0073] Where Ah is the total discharge ampere-hour throughput; R g ,z,E,E a , c are the universal gas constant, power law factor, pre-exponential factor, activation energy and charge rate respectively; ΔSoH k is the battery health status degradation value corresponding to the kth sampling moment; Δt is the sampling interval.

[0074] (1.2) Charging environment

[0075] In order to better represent the safe charging problem, the constraint Markov decision process is used to model the reinforcement learning environment and the tuple is defined as in and Denote the state space and action space, respectively; p represents the probability of transitioning from the current state-action pair to the next state; r, γ, c, and d represent the immediate reward, discount rate, cost, and constraint threshold, respectively. At each sampling time, the reinforcement learning agent selects an action based on the policy network's policy according to the current state. The environment then returns the reward, cost, and the next state selected based on the probabilities. Step 2: Initialize a policy network and two value networks, define the reinforcement learning state space and action space, and set the charging reward function and cost function.

[0076] (2.1) State space and action space

[0077] s t ={SoC(t),V t (t), T c (t)}

[0078] a t ={I(t)|I(t)∈[0,I max ]}

[0079] Where, and s t and a t Represent the state observed and the action taken at time step t; I max This is the maximum charging current with hard constraints.

[0080] (2.2) Charging multi-objective optimization problem

[0081] The charging strategy is formulated as a multi-objective optimization problem to minimize the charging time required to achieve the desired SoC while ensuring that the charging constraints are met. It can be expressed as follows:

[0082]

[0083] stV t，min <<V t(t)<<V t，max

[0084] T cmin <<T c (t)<<T c，max

[0085] In the formula, the symbol ∈ represents the “belongs to” relationship; ∫· represents the integral operator; t f Indicates the end time of the charging process and charging starts at time t0 by default; st is used to introduce constraints.

[0086] 1) Reward Function

[0087] Set the reward to only be related to charging behaviors that encourage faster charging:

[0088] r t =|SoC tar -SoC(t)|

[0089] Where r t represents the reward obtained at time step t; SoC tar The expected SoC indicates that the charging process is complete;

[0090] 2) Cost Function

[0091] During battery charging, the battery terminal voltage and core temperature need to be controlled below a reasonable threshold. At the same time, there are also requirements for the battery capacity decay rate. Based on this, the charging constraints are set as follows:

[0092]

[0093] C soh (t) = |ΔSoH t |

[0094] Where V t，max 、V t，min 、T c，max 、T c，min are the upper and lower cut-off voltages and the highest and lowest core temperature thresholds respectively; C υolt and C temp Used to punish battery overvoltage and overtemperature, while C soh represents the aging of the battery. The total cost function is described as:

[0095] c t =C υolt (t)+C temp (t)+C soh (t)

[0096] Step 3: Update the state space variables based on the battery model in step 1, and add the current state, charging action, reward value, cost value, and next moment state to the experience pool. Iteratively train the charging strategy network, value evaluation network, and cost evaluation network constructed based on the distributed soft actor-critic algorithm and conservative strategy optimization techniques until convergence, and obtain the converged strategy network model.

[0097] (3.1) Construction of cost assessment network

[0098] First, the expression of the original multi-objective optimization problem is rewritten. The Lagrangian dual form of the problem is:

[0099]

[0100] Where λ is the Lagrange multiplier; π is the strategy adopted; (s, a) is the state-action pair; and are the estimates of reward and cost functions respectively; d is the constraint threshold; ρ π (s) represents the state distribution caused by the policy π.

[0101] The underestimated cost function during charging is used to train other state-action pairs and accumulates into large deviations in temporal difference learning, resulting in charging behaviors that violate constraints. A cost evaluation network is constructed using conservative policy optimization techniques to address the underestimated cost. The upper confidence bound of the true cost value is used instead of the estimate, that is, the "mean + weighted standard deviation" of the set is calculated, expressed as:

[0102]

[0103] Where, is the upper confidence limit estimate of the cost; E represents the use of E bootstrap networks with the same network structure and independent training process Cost estimation is performed; k is a conservative parameter. The cost evaluation network parameters are collectively referred to as ψ, and the parameters of each cost sub-network are updated using the temporal difference error method.

[0104] (3.2) Construction of Charging Strategy Network and Value Assessment Network

[0105] The overestimation of the reward value function in charging training of traditional reinforcement learning methods will cause a decrease in charging speed, while the value distribution learning method can significantly improve the overestimation problem of the reward value function, thereby reducing the instability of strategy performance during the learning process.

[0106] The interactive training process of the environment and algorithm for lithium battery safety optimization fast charging strategy based on distributed reinforcement learning algorithm is as follows Figure 2As shown in the figure, the training phase is based on an actor-critic network framework, which includes a charging policy network and a value assessment network. The charging policy network takes as input the lithium battery state space variables and outputs the action value for controlling charging. The value assessment network is used to evaluate the quality of the current charging policy. The parameters of the charging policy network are collectively referred to as φ, and the parameters of the value assessment network are collectively referred to as θ.

[0107] The distributed soft actor-critic algorithm assumes that the distribution of rewards is Gaussian, expressed as represents the value distribution of rewards; Q θ (s, a) and σ θ (s, a) represents the mean and standard deviation of the value distribution; is the standard operator of Gaussian distribution; learn the mean and variance of the distribution, and dynamically adjust the deviation size through the variance of the distribution to achieve the effect of suppressing over-estimation; the value evaluation network gradient is represented by the mean-related gradient and the variance-related gradient:

[0108]

[0109]

[0110] Where, represents the gradient of the i-th evaluation network; is the gradient operator for the parameter θ; the clipping function C p (·) is used to alleviate gradient explosion or disappearance; b is the clipping boundary to alleviate the sensitivity of reward scale; and are the target minimum distribution network and the target minimum value network respectively; r and are the reward and discount coefficient at γ respectively; s′ and a′ are the state and action at the next moment; Z is the soft state-action reward; α is the weighting factor of the entropy term; the target network parameters of the charging strategy and value evaluation are collectively referred to as and Used to improve stability during training; the time difference error is used to update the value evaluation network parameters θ.

[0111] The safe charging problem is described as a solution to the optimal policy π * The strategy maximizes the long-term reward while limiting the cumulative cost. The objective function is expressed as:

[0112]

[0113] Where, the temperature coefficient α gradually decays during the training process; It means that two value distribution networks are used; the charging strategy network is updated using the time difference error to maximize the reward target.

[0114] Step 4: Collect the battery’s current terminal voltage, surface temperature, and ambient temperature data in real time, use the state observer to estimate the unmeasurable battery state of charge and core temperature, and input the processed state information into the charging strategy network trained in step S3 to determine the optimal charging current trajectory.

[0115] The deployment process of the charging strategy is as follows:

[0116] First, the state observer and the trained charging strategy network π φ Deploy it to the charging controller; set the charging target and constraint thresholds; in each sampling cycle, collect the current terminal voltage, surface temperature, and ambient temperature data of the battery, use the state observer to estimate the battery SoC and core temperature, input the state space variables into the deployed charging strategy network and obtain the output action, denormalize the action value to obtain the expected charging current, and input it into the battery environment until the charging process is complete.

[0117] In this embodiment, a battery fast charging control method considering charging safety and health degradation is verified, and a comparative experiment is conducted with the CCCV strategy (including 2C-CCCV, 4C-CCCV, and 6C-CCCV) widely used in industry. The comparison results are as follows: Figure 3 As shown in Table 1, Figure 3 (a) is a comparison chart of charging current rates; Figure 3 (b) is a comparison diagram of battery state of charge; Figure 3 (c) is the battery terminal voltage comparison diagram; Figure 3 (d) is a comparison chart of battery core temperature.

[0118] Table 1 Comparison of capacity decay under long-term charging cycles

[0119] Charging algorithm DSAC 2C-CCCV 4C-CCCV 6C-CCCV Maximum capacity 98.21％ 98.92％ 97.64％ 95.23％

[0120] The results show that the CCCV strategy faces an irreconcilable contradiction between charging speed and constraint satisfaction. Although the 2C-CCCV strategy achieves thermal safety, the charging speed is significantly slower than other strategies. In contrast, when the current rate is increased to 4C, the risk of overheating is introduced. The charging control strategy based on the distributed reinforcement learning algorithm shows a strong ability to balance charging speed and constraints. It limits the core temperature of the lithium battery to below the set 45°C, and the charging time is close to that of the 4C-CCCV strategy, but has thermal perception capabilities. The maximum capacity (after 200 cycles) is greater than 97.64% of the 4C-CCCV strategy and 95.23% of the 6C-CCCV strategy, indicating a longer healthy lifespan.

[0121] The results show that compared with the traditional CC-CV algorithm, the present invention can effectively ensure charging safety and health degradation while shortening the charging time, thereby reducing the user's charging anxiety. The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions only describe the principles of the present invention. The present invention will have various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements all fall within the scope of the invention claimed for protection. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.

Claims

1. A lithium battery safety optimization fast charging control method, characterized in that: The following steps are involved: S1. Build a coupled electrothermal aging model for lithium-ion batteries, determine the model's control equations, and establish the offline charging environment required for reinforcement learning based on the model. S2. Initialize a policy network and two value networks, define the reinforcement learning state space and action space, and set the charging reward function and cost function; S3. Update the state space variables based on the battery model in step S1, and add the current state, charging action, reward value, cost value, and next-moment state to the experience pool. Iteratively train the charging policy network, value assessment network, and cost assessment network constructed using the distributed soft actor-critic algorithm and conservative policy optimization techniques until convergence, obtaining the converged policy network model. S4. Collect the battery's current terminal voltage, surface temperature, and ambient temperature data in real time, use the state observer to estimate the unmeasurable battery charge state and core temperature, input the processed state information into the charging strategy network trained in step S3, and determine the optimal charging current trajectory.

2. A lithium battery safety optimized fast charging control method according to claim 1, characterized in that: The lithium-ion battery electrothermal aging coupling model in step S1 includes an equivalent circuit model for capturing the behavior of the battery at different temperatures, a thermodynamic model for characterizing the internal and surface temperatures of the battery, and an aging model for monitoring the health status of the battery.

3. A lithium battery safety optimized fast charging control method according to claim 2, characterized in that: The derivation process of the control equation of the lithium-ion battery electrothermal aging coupling model in step S1 is as follows: The physical behavior of the battery is described by a temperature-dependent second-order RC equivalent circuit model. According to Kirchhoff's law, the governing equation can be derived as: V t (t)=V ocv (SoC)-V1(t)-V2(t)-R0(t,T)I(t) Where, SoC, I, C n 、V t They represent the battery's state of charge, charging current, nominal capacity, and terminal voltage, respectively; d and t represent the differential operator and time, respectively; the open circuit voltage V ocv is a polynomial function of SoC; V1 and V2 are the voltages of the two resistor-capacitor pairs, respectively; the ohmic internal resistance R0, polarization resistances R1 and R2, and polarization capacitances C1 and C2 are expressed as functions of SoC and temperature T; under the assumption of longitudinal uniformity, a two-state thermal model is used to capture the lumped thermal characteristics of the cylindrical battery, and the governing equation for the heat transfer process is: Where, T s 、T c and T f are the battery surface temperature, core temperature and ambient temperature respectively; R s and R c are heat convection resistance and heat conduction resistance respectively; C s and C c are the heat capacities of the battery surface and the core, respectively; the heat generation rate Q is given by the Bernardi heat generation model: Q(t)=(V ocv (SoC)-V t (t))I(t)+I(t)T m (t)E n (SoC,t) Where, E n is the entropy change in the electrochemical reaction; T m is the average temperature of the battery; in addition, the semi-empirical model of the Arrhenius equation is used to evaluate the battery capacity decay rate, and the charge throughput is used to quantify the capacity decay at different charge rates: Where Ah is the total discharge ampere-hour throughput; R g ,z,B,E a , c are the universal gas constant, power law factor, pre-exponential factor, activation energy and charge rate respectively; ΔSoH k is the battery health status degradation value corresponding to the kth sampling moment; Δt is the sampling interval.

4. A lithium battery safety optimized fast charging control method according to claim 3, characterized in that: The implementation process of step S2 is as follows: The state space and action space are defined as: s t ={SoC(t),V t (t),T c (t)} a t ={I(t)|I(t)∈[0,I max ]} Where s t and a t Represent the state observed and the action taken at time step t; I max is the maximum charging current constrained by the constraint; the charging strategy is formulated through a multi-objective optimization problem to minimize the charging time required to reach the desired SoC while ensuring that the charging constraints are met; the reward function is set to be only related to the charging speed: r t =|SoC tar -SoC(t)| Where r t represents the reward obtained at time step t; SoC tar is the expected SoC, indicating that the charging process is completed; the charging constraints are designed as cost functions respectively: C soh (t)=|ΔSoH t | Where V t，max 、V t，min 、T c，max 、T c，min are the upper and lower cut-off voltages and the highest and lowest core temperature thresholds respectively; C volt and C temp represents the penalty term for battery overvoltage and overtemperature, and C soh represents the aging of the battery; the total cost function at each step is described as: c t =C volt (t)+C temp (t)+C soh (t)。 5. The lithium battery safety optimized fast charging control method according to claim 4, characterized in that: The cost evaluation network implementation process in step S3 is as follows: Rewrite the original constrained optimization problem, and the Lagrangian dual form of the problem is: Where λ is the Lagrange multiplier; π is the strategy adopted; (s, a) is the state-action pair; and are the estimated values of the true reward and cost function respectively; d is the total constraint threshold; ρ π (s) represents the state distribution caused by the policy π; the cost function underestimated during the charging process is used to train other state-action pairs and accumulates into a large deviation in the temporal difference learning, resulting in charging behavior that violates the constraint; A conservative strategy optimization technique is used to construct a cost evaluation network to solve the underestimation problem. The upper confidence bound of the true cost value is used instead of the estimate, that is, the "mean + weighted standard deviation" of the set is calculated, which is expressed as: Where, is the upper confidence limit estimate of the cost; E represents the use of E bootstrap networks with the same network structure and independent training process Cost value estimation is performed; k is a conservative parameter; the cost evaluation network parameters are collectively referred to as ψ, and the parameters of each cost sub-network are updated using the time difference error method.

6. A lithium battery safety optimized fast charging control method according to claim 5, characterized in that: The implementation process of the charging strategy network and the value assessment network in step S3 is as follows: Traditional reinforcement learning methods overestimate the reward value function during charging training, which will cause the charging speed to decrease. However, the value distribution learning method can significantly improve the overestimation problem of the reward value function, thereby reducing the instability of policy performance during the learning process. The training phase is based on an actor-critic network framework, which includes a charging policy network and a value assessment network. The charging policy network takes as input the lithium battery state space variables and outputs the action value that controls charging. The value assessment network is used to evaluate the quality of the current charging policy. The parameters of the charging policy network are collectively referred to as φ, and the parameters of the value assessment network are collectively referred to as θ. The distributed soft actor-critic algorithm assumes that the distribution of random rewards follows a Gaussian distribution, expressed as: Where, represents the value distribution of rewards; Q θ (s, a) and σ θ (s, a) represents the mean and standard deviation of the value distribution; is the standard operator of Gaussian distribution; the deviation size is dynamically adjusted by learning the variance of the distribution to achieve the effect of suppressing over-estimation; the value assessment network gradient consists of two parts: mean-related gradient and variance-related gradient, expressed as: Where, represents the gradient of the i-th evaluation network; is the gradient operator for the parameter θ; the clipping function C p (·) is used to alleviate gradient explosion or disappearance; b is the clipping boundary to alleviate the sensitivity of reward scale; and are the target minimum distribution network and the target minimum value network respectively; r and are the reward and discount coefficient at γ respectively; s′ and a′ are the state and action at the next moment; Z is the soft state-action reward; α is the weighting factor of the entropy term; the target network parameters of the charging strategy and value evaluation are collectively referred to as and Used to improve stability during training; use time difference error to update the value evaluation network parameter θ; The safe charging problem is described as a solution to the optimal policy π * The strategy maximizes the long-term reward while limiting the cumulative cost, and the update target J of the policy network π Expressed as: In the formula, the symbol Indicates mean operation; the symbol ~ indicates random sampling based on probability distribution; A playback buffer for sampled experience; The weighting factor α decays during training; π φ (a|s) represents the probability of action a occurring in state s; the charging strategy network is updated using the temporal difference error to maximize the reward target.

7. A lithium battery safety optimized fast charging control method according to claim 6, characterized in that: The charging strategy deployment process in step S4 is as follows: First, the state observer and the trained charging strategy network π φ Deploy it to the charging controller; set the charging target and constraint thresholds; in each sampling cycle, collect the current terminal voltage, surface temperature, and ambient temperature data of the battery, use the state observer to estimate the battery SoC and core temperature, input the state space variables into the deployed charging strategy network and obtain the output action, denormalize the action value to obtain the expected charging current, and input it into the battery environment until the charging process is complete.

Images