A multi-objective adaptive control method for a new energy vehicle battery pack thermal management system

Abstract

The application discloses a new energy automobile battery pack thermal management system multi-objective adaptive control method, comprising: constructing a thermal resistance-thermal capacity dynamic network model of a battery module and a cooling flow channel, and building a reinforcement learning virtual training environment; constructing a multi-objective composite reward function; constructing a neural network model based on a deep deterministic policy gradient, using the multi-objective composite reward function to guide the gradient update of the policy network parameters until the neural network model converges; solidifying the trained and converged policy network parameters, deploying them into a vehicle thermal management controller, and outputting the optimal coolant flow rate under the current working condition; and mapping the optimal coolant flow rate to an electronic water pump speed instruction. The application can solve the problems that the prior art is difficult to meet the millisecond-level real-time response demand under extreme working conditions due to large calculation load, is limited by relying only on linear deduction of the temperature distribution of the first and last modules, and cannot dynamically adjust the strategy in real time.

Description

Technical Field

[0001] This invention relates to the field of new energy vehicle technology, and more specifically to a multi-objective adaptive control method for a new energy vehicle battery pack thermal management system. Background Technology

[0002] As the energy core of new energy vehicles, the battery pack's operating status directly affects the vehicle's range, power response, and safety performance. Lithium-ion batteries dominate the market due to their high specific energy and long lifespan, but they exhibit extremely high sensitivity to operating temperature. Existing research indicates that maintaining the battery within a suitable temperature range of 25-40℃ and strictly controlling the temperature difference between modules within 5℃ is crucial for ensuring stable performance. Under extreme conditions of high temperature or high load, the reversible and irreversible heat generation effects within the battery are superimposed. If heat dissipation is not timely, it will not only accelerate capacity degradation and increase internal resistance, but in severe cases, it may even induce thermal runaway accidents. Therefore, developing a battery thermal management and control strategy that balances high efficiency, precision, and low energy consumption is particularly urgent.

[0003] Existing battery pack thermal management control methods mainly include PID control, model predictive control (MPC), and heuristic optimization strategies represented by genetic algorithms (GA). However, existing battery pack thermal management control methods also have problems such as difficulty in meeting the millisecond-level real-time response requirements under extreme operating conditions due to high computational load, limitations caused by relying solely on linear extrapolation of temperature distribution from the first and last modules, and the inability to provide real-time feedback for dynamic adjustment strategies.

[0004] The Deep Deterministic Policy Gradient (DDPG) algorithm, by integrating the perceptual capabilities of deep learning with the decision-making power of reinforcement learning, exhibits the unique advantage of "offline training and online inference." Unlike genetic algorithms that rely on repeated online iterations, the DRL controller internalizes complex nonlinear control laws into the neural network through trial and error with the environment. During actual vehicle operation, it can output the optimal policy with only one forward propagation, achieving millisecond-level ultra-fast response. Summary of the Invention

[0005] In view of this, the present invention provides a multi-objective adaptive control method for the thermal management system of a new energy vehicle battery pack, in order to solve the problems of existing technologies, such as the inability to meet the millisecond-level real-time response requirements under extreme conditions due to the large computational load, the limitations caused by relying solely on the linear extrapolation of temperature distribution from the first and last modules, and the inability to provide real-time feedback on dynamic adjustment strategies.

[0006] A multi-objective adaptive control method for a thermal management system of a new energy vehicle battery pack includes:

[0007] Step S1: A dynamic network model of thermal resistance and thermal capacity of battery module and cooling channel is constructed using the lumped parameter method. Based on this model, a reinforcement learning virtual training environment including state space, action space and environmental feedback mechanism is built. In this model, the internal resistance of battery module is corrected by the battery health status, thereby correcting the heat generation equation of battery module. Based on the nonlinear changes of coolant density, viscosity and thermal conductivity with temperature, the Reynolds number, Prandtl number and convective heat transfer coefficient are updated in real time.

[0008] Step S2: Select the key node temperature, state of charge, health status, load current and ambient temperature of the battery pack as state variables, and select the coolant flow rate as the control variable. Based on the energy consumption of the cooling system, the temperature control accuracy of the battery pack and the maximum temperature difference of the battery module, construct a multi-objective composite reward function.

[0009] Step S3: Construct a neural network model based on deep deterministic policy gradient. This neural network model is a dual-network structure containing a policy network and a value network. It introduces an experience replay mechanism and a soft update strategy for the target network. In the reinforcement learning virtual training environment constructed in step S1, through the continuous interaction between the neural network model and the reinforcement learning virtual training environment, the gradient update of the policy network parameters is guided by a multi-objective composite reward function until the neural network model converges.

[0010] Step S4: Solidify the parameters of the policy network that has been trained and converged in step S3, and deploy them to the vehicle thermal management controller. The vehicle thermal management controller collects vehicle sensor data in real time as state input, and directly outputs the optimal coolant flow rate under the current operating conditions through forward propagation calculation of the policy network.

[0011] Step S5: The optimal coolant flow rate output in step S4 is mapped to the electric water pump speed command through the electric water pump characteristic curve, and sent to the actuator via the CAN bus to realize the adjustment of the battery pack thermal management system.

[0012] The multi-objective adaptive control method for the thermal management system of a new energy vehicle battery pack provided by the present invention has the following beneficial effects:

[0013] (1) This invention abandons the traditional online iterative optimization mode of genetic algorithms and adopts an offline training and online inference mechanism. During the online operation of the actual vehicle, the vehicle thermal management controller only needs to go through one forward propagation of the neural network to output control commands, reducing the calculation time from seconds to milliseconds, which can effectively cope with the extreme conditions of transient thermal shock such as rapid acceleration and fast charging of the vehicle.

[0014] (2) This invention uses the lumped parameter method to construct a dynamic network model of thermal resistance and thermal capacity for the battery module and cooling channel, overcoming the limitations of the prior art which relies solely on the linear extrapolation of temperature distribution from the first and last modules. By correcting the internal resistance of the battery module through the battery health status, the heat generation equation of the battery module is then corrected. Furthermore, based on the nonlinear changes in coolant density, viscosity, and thermal conductivity with temperature, the Reynolds number, Prandtl number, and convective heat transfer coefficient are updated in real time. The dynamic network model of thermal resistance and thermal capacity can accurately capture the thermal inertia of the fluid and the heat generation changes throughout its entire life cycle, providing a highly physically realistic training environment for the neural network model and significantly improving the credibility of the control strategy in real vehicle applications.

[0015] (3) The present invention constructs a neural network model based on deep deterministic policy gradient, which can use the adaptive capability of reinforcement learning to make up for the defect of parameter mismatch in physical model after battery aging. The neural network model can dynamically adjust the strategy according to real-time feedback, which can not only adapt to different ambient temperatures, but also automatically adapt to the increase in internal resistance and heat generation caused by battery SOH decay, without the need for manual recalibration of parameters, and realize optimal thermal management throughout the entire life cycle.

[0016] (4) The present invention realizes end-to-end mapping from sensor data to actuator commands (pump speed), without the need for complex intermediate feature extraction or cumbersome constraint solving, reducing the occupation of vehicle controller hardware resources and making it easy to deploy and promote in engineering. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the multi-objective adaptive control method for the thermal management system of a new energy vehicle battery pack provided in an embodiment of the present invention.

[0018] Figure 2 A comparison chart of the total energy consumption of cooling systems using different methods;

[0019] Figure 3 A comparison chart of the temperature control accuracy of battery packs using different methods;

[0020] Figure 4 This is a comparison chart of the maximum temperature difference of battery modules using different methods. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain embodiments of the present invention, and should not be construed as limiting the present invention.

[0022] Please see Figure 1The present invention provides a multi-objective adaptive control method for a thermal management system of a new energy vehicle battery pack, comprising steps S1 to S5:

[0023] Step S1: A dynamic network model of thermal resistance and thermal capacity of the battery module and cooling channel is constructed using the lumped parameter method. Based on this model, a reinforcement learning virtual training environment including state space, action space and environmental feedback mechanism is built. In this model, the internal resistance of the battery module is corrected by the battery health status, thereby correcting the heat generation equation of the battery module. Based on the nonlinear changes of coolant density, viscosity and thermal conductivity with temperature, the Reynolds number, Prandtl number and convective heat transfer coefficient are updated in real time.

[0024] The thermal resistance-thermal capacity dynamic network model is constructed through the following process:

[0025] The heat generation rate of the battery module consists of Joule heat and reaction entropy heat. To simulate the real characteristics of the battery throughout its entire life cycle, this embodiment introduces a healthy state. Dynamic correction is applied to the internal resistance. First, the heat generation equation of the battery module is established, with the following expression:

[0026]

[0027]

[0028] in, Indicates the first Each battery module Heat generation at all times express Load current at any given time The dynamic internal resistance of the battery module. Indicates the first The temperature of each battery module In a charged state, In good health; The entropy heat coefficient is shown in Table 1; For the first Each battery module at temperature The reference internal resistance under the current state of charge is taken as 0.8mΩ in this embodiment; The aging sensitivity coefficient is set to 0.6 in this embodiment. The calculation formula shows that as The internal resistance decreases, and the internal resistance increases non-linearly, thus accurately reflecting the increase in heat generation after aging.

[0029] Table 1 Entropy-Heat Coefficient Lookup Table

[0030]

[0031] To overcome the deficiency of constant coolant properties in traditional models, a fitting function for the coolant's physical parameters as a function of temperature was established. In this embodiment, the coolant is a 50% (v / v) ethylene glycol aqueous solution. The liquid cooling system mainly consists of convective heat transfer and heat conduction. The convective heat transfer equation for the battery module is established as follows:

[0032]

[0033]

[0034]

[0035]

[0036]

[0037]

[0038]

[0039]

[0040] in, For the first Convection heat transfer of each battery module The convective heat transfer coefficient; In this embodiment, the effective heat exchange area is 0.012283 m². 2 ; For the first The temperature of the coolant within the range of each battery module Let Reynolds number be 1. For Prandtl numbers, The thermal conductivity of the coolant. The thermal conductivity is the reference value. ; The diameter of the coolant pipe is 22 mm in this embodiment; For coolant density, The reference value is the coolant density. , This refers to the coolant flow rate. The dynamic viscosity of the coolant. This is the reference value for the dynamic viscosity of the coolant. This is the temperature calibration value. , The specific heat capacity of the coolant. This is the reference value for the specific heat capacity of the coolant. .

[0041] The heat transfer equation between battery modules is established, and the expression is:

[0042]

[0043] in, Indicates the first Thermal conductivity of each battery module The thermal conductivity coefficient between adjacent battery modules is taken as 1000 W / (m²) in this embodiment. ℃); The contact area between battery modules is 0.035811 mm in this embodiment. 2 ; The characteristic length of thermal conduction between battery modules is 0.01 m in this embodiment; Indicates the first The temperature of each battery module Indicates the first The temperature of each battery module.

[0044] Based on the principle of energy conservation, a set of coupled differential equations between battery nodes and fluid nodes is constructed to describe the dynamic evolution of the system. Specifically, the energy balance equation of the battery node is established:

[0045]

[0046] in, The specific heat capacity of a single battery module is taken as 1128.45 J / (kg). ℃); The mass of a single battery module is 3.2 kg in this embodiment; Indicates the first The rate of temperature change of each battery module over time.

[0047] Establish the energy balance equation for the fluid node:

[0048]

[0049] in, For the mass of coolant within a single battery module, Indicates the first Temperature change rate of coolant within the range of each battery module This refers to the coolant mass flow rate. For the first The temperature of the coolant within the range of each battery module.

[0050] In order to calculate and The value of can be defined and calculated based on the geometric parameters of the cooling pipe as follows, for the cross-sectional area of ​​the cooling pipe. :

[0051]

[0052] In an embodiment, The calculated result is 3.8 × 10 -4 m 2 The length of the cooling pipe corresponding to a single battery module; The calculation formula is:

[0053]

[0054] In an embodiment, The calculated result is 0.178m; for the coolant mass within a single battery module range. The calculation formula is as follows:

[0055]

[0056] The calculation formula is:

[0057] .

[0058] To accommodate discrete-time step training of reinforcement learning algorithms, the Euler method is used to discretize the energy balance equations for battery nodes and fluid nodes, resulting in the state update equations for the thermal resistance-thermal capacity dynamic network model:

[0059]

[0060]

[0061] in, Indicates the first Each battery module Temperature at any moment For simulating step size, Indicates the first Each battery module Temperature at any moment for Time of the first The temperature of the coolant within the range of each battery module for Time of the first The temperature of the coolant within the range of each battery module For the first The temperature of the coolant within the range of each battery module.

[0062] By using the above discretization equations, at each simulation step... Within 0.1s (in this embodiment), the temperature distribution of all nodes at the next moment is calculated based on the current action and state. This set of dynamic equations constitutes a high-fidelity virtual environment for the interaction of deep reinforcement learning agents in steps S2 and S3, enabling the model to perceive the nonlinear effect of actions on the system temperature.

[0063] Step S2: Select the key node temperature, state of charge, health status, load current and ambient temperature of the battery pack as state variables, and select the coolant flow rate as the control variable. Based on the energy consumption of the cooling system, the temperature control accuracy of the battery pack and the maximum temperature difference of the battery module, construct a multi-objective composite reward function.

[0064] For the state space, physical quantities that comprehensively reflect the system's thermodynamic state and aging degree are selected to construct state variables. To eliminate the influence of dimensions, each state quantity is normalized. The definition is as follows:

[0065]

[0066] in, This indicates the highest temperature of all battery modules in the battery pack at the current moment. This indicates the average temperature of the battery pack at the current moment. This represents the maximum temperature difference between battery modules. This represents the highest temperature change rate of the battery module. For ambient temperature, This indicates transpose.

[0067] For the action space, coolant flow rate is selected as the control variable. To adapt to the neural network output, it is normalized. Coolant flow rate The functional relationship between the action and the action is:

[0068]

[0069] in, Indicates the coolant flow rate action. This indicates the maximum flow rate that the electronic water pump can provide; in this embodiment, it is taken as 0.5 m / s.

[0070] To address the multi-objective optimization problem involving energy consumption, temperature control accuracy, and temperature uniformity, a multi-objective composite reward function is constructed. for:

[0071]

[0072] in, This represents the reward function for temperature control accuracy. Represents the uniformity penalty function. Indicates energy consumption penalty. , , These are the weighting coefficients.

[0073] Reward function for temperature control accuracy A piecewise Gaussian function is used to guide the temperature to remain within the optimal range (25-35°C). The calculation expression is as follows:

[0074]

[0075] For the uniform temperature penalty function No points are deducted when the temperature difference is less than 3℃; a linear penalty is applied when the temperature difference exceeds 3℃. The calculation expression is as follows:

[0076]

[0077] in, This indicates taking the maximum value.

[0078] Energy consumption penalty The calculation is based on the cubic relationship between pump power and flow velocity, and the expression is:

[0079] .

[0080] Step S3: Construct a neural network model based on deep deterministic policy gradient. This neural network model is a dual-network structure containing a policy network and a value network. It introduces an experience replay mechanism and a soft update strategy for the target network. In the reinforcement learning virtual training environment constructed in step S1, through continuous interaction between the neural network model and the reinforcement learning virtual training environment, the gradient update of the policy network parameters is guided by a multi-objective composite reward function until the neural network model converges.

[0081] In this embodiment, a dual neural network structure is constructed. The first is the policy network (Critic), whose function is to fit a deterministic policy function. The policy network satisfies the following equation:

[0082]

[0083] in, Represents the policy network, Indicates the policy network parameters;

[0084] The input layer of the policy network contains 8 neurons, which receive 8-dimensional state variables. The hidden layer of the policy network is designed as two fully connected layers with 256 and 128 nodes respectively, and ReLU is used as the activation function between the layers. The output layer of the policy network contains one neuron and uses the Sigmoid activation function to limit the output value to the (0-1) range to correspond to the normalized coolant flow rate action.

[0085] Value networks (Actors) are used to fit action value functions. To evaluate the quality of the action. Represents a value network. This represents the parameters of the value network. The input layer of the value network contains 9 neurons, which are respectively connected to... and The input is a combination of two layers; the hidden layer of the value network adopts a two-layer fully connected structure with 256 and 128 nodes, and both use the ReLU activation function; the output layer of the value network contains one neuron and directly outputs the corresponding Q value in a linear output manner.

[0086] The training process and parameters will be updated further. Initialization and data sampling will begin first. The policy network parameters will be randomly initialized. and value network parameters Simultaneously, a target network with the same structure is established. The neural network model interacts in the reinforcement learning virtual training environment constructed in step S1, selecting actions according to the current policy. After the model executes an action, the environment provides feedback on the state for the next time step. and rewards The system will transition the state quadruple. , , , Store it in the experience replay pool K.

[0087] Next, the parameters of the value network are updated. N sample data are randomly sampled from the experience replay pool K, and the target value is calculated. This target value is determined by immediate rewards. The value network parameters are composed of the discounted value at the next time step. Then, by minimizing the mean squared error loss function between the predicted and target values, the gradient descent method is used to update the value network parameters. .

[0088] Finally, the policy network and target network are updated softly. Based on the updated value network, the gradient is calculated using the deterministic policy gradient formula, aiming to maximize the value network's score for the current policy output action, thereby updating the policy network parameters. After the main network is updated, the target network parameters are softly updated using the moving average method.

[0089] Step S4: Solidify the trained and converged strategy network parameters from step S3 and deploy them to the vehicle thermal management controller. The vehicle thermal management controller collects vehicle sensor data in real time as state input and directly outputs the optimal coolant flow rate under the current operating conditions through forward propagation calculation of the strategy network.

[0090] Specifically, step S4 includes:

[0091] After the neural network model is trained, the weight matrix and bias vector of the policy network are extracted and embedded into the vehicle thermal management controller. During the online operation phase of the actual vehicle, since it is impossible to directly measure all the physical states inside the battery, this step adopts a two-layer architecture of state observation-policy inference to achieve control. First, data acquisition and state estimation are performed. The vehicle thermal management controller acquires the raw sensor signals of the bottom layer with a period of 100ms, including sensor data of total voltage, load current, and temperature points. Using the state estimation module built into the BMS, the current SOC and SOH are estimated in real time using the ampere-hour integral method combined with the extended Kalman filter algorithm. At the same time, based on the temperature sampling points, online observation or extreme value mapping is performed in combination with the thermal resistance-thermal capacity dynamic network model established in step S1 to estimate the... , and ;

[0092] Subsequently, feature vector assembly and policy inference are performed. The estimated values ​​are combined with directly measured load current and ambient temperature, and normalized to construct state variables that meet the input requirements of the policy network. , will state variables The input to the solidified neural network model performs forward propagation computation. The computation process sequentially passes through the linear transformation and ReLU activation of the hidden layers, and finally through the sigmoid activation function of the output layer to output the optimal action under the current working condition in real time. This process involves only matrix multiplication and addition operations, with extremely low computational complexity, which can meet the real-time requirements of the vehicle controller.

[0093] Step S5: The optimal coolant flow rate output in step S4 is mapped to the electric water pump speed command through the electric water pump characteristic curve, and sent to the actuator via the CAN bus to realize the adjustment of the battery pack thermal management system.

[0094] Specifically, step S5 includes:

[0095] Based on the pressure-flow characteristic curve of the electronic water pump and the flow resistance characteristics of the pipeline, the coolant flow rate is pre-calibrated. With the speed of the electric water pump The fitting function relationship is:

[0096]

[0097] in, , , These are calibration coefficients. It should be noted that during calibration, , The physical dimensions need to be determined in conjunction with related quantities, so that The calculation results The calculation results, and Its physical dimensions are all in units of rotational speed, i.e., rpm. Therefore... The unit is revolutions. minutes per square meter The unit is revolutions per meter (rpm). The unit is revolutions per minute.

[0098] The vehicle thermal management controller will output the optimal action in step S4. Multiply To obtain the target flow rate Then Substituting the values ​​into the above fitted function, the electric water pump speed command is calculated. .

[0099] Finally, closed-loop control is executed. The vehicle thermal management controller transmits the calculated engine speed command via the CAN bus. The data is sent to the electronic water pump drive unit. The water pump drive unit adjusts the motor duty cycle (PWM) to make the actual speed follow the target speed, thereby changing the coolant flow rate. The change in coolant flow rate directly affects the temperature field distribution inside the battery pack. The new temperature data will be collected by the sensors in the next control cycle and used for state estimation, thus forming a complete "sensing-decision-execution" closed-loop control system to achieve precise management of battery temperature.

[0100] The following simulation test was conducted on this invention. A real vehicle was driven at its maximum speed (140 km / h) in an environment with a temperature of 40°C. The energy consumption, temperature control accuracy, and maximum temperature difference of the battery module under the PID control strategy, the genetic algorithm (GA) control strategy, and the method of this invention were compared and analyzed. Figure 2 As can be seen, under the same operating conditions, the cooling system of this invention has the lowest energy consumption. From... Figure 3 As can be seen, the genetic algorithm (GA) control strategy reaches the target temperature faster and has higher control accuracy than the PID control strategy of this invention, but this invention improves control accuracy compared to the genetic algorithm (GA) control strategy. Figure 4 As can be seen, the battery module of the present invention has the smallest temperature difference, only about 0.42℃.

[0101] In summary, the multi-objective adaptive control method for the thermal management system of new energy vehicle battery packs according to the above embodiments has the following beneficial effects:

[0102] (1) This invention abandons the traditional online iterative optimization mode of genetic algorithms and adopts an offline training and online inference mechanism. During the online operation of the actual vehicle, the vehicle thermal management controller only needs to go through one forward propagation of the neural network to output control commands, reducing the calculation time from seconds to milliseconds, which can effectively cope with the extreme conditions of transient thermal shock such as rapid acceleration and fast charging of the vehicle.

[0103] (2) This invention uses the lumped parameter method to construct a dynamic network model of thermal resistance and thermal capacity for the battery module and cooling channel, overcoming the limitations of the prior art which relies solely on the linear extrapolation of temperature distribution from the first and last modules. By correcting the internal resistance of the battery module through the battery health status, the heat generation equation of the battery module is then corrected. Furthermore, based on the nonlinear changes in coolant density, viscosity, and thermal conductivity with temperature, the Reynolds number, Prandtl number, and convective heat transfer coefficient are updated in real time. The dynamic network model of thermal resistance and thermal capacity can accurately capture the thermal inertia of the fluid and the heat generation changes throughout its entire life cycle, providing a highly physically realistic training environment for the neural network model and significantly improving the credibility of the control strategy in real vehicle applications.

[0104] (3) The present invention constructs a neural network model based on deep deterministic policy gradient, which can use the adaptive capability of reinforcement learning to make up for the defect of parameter mismatch in physical model after battery aging. The neural network model can dynamically adjust the strategy according to real-time feedback, which can not only adapt to different ambient temperatures, but also automatically adapt to the increase in internal resistance and heat generation caused by battery SOH decay, without the need for manual recalibration of parameters, and realize optimal thermal management throughout the entire life cycle.

[0105] (4) The present invention realizes end-to-end mapping from sensor data to actuator commands (pump speed), without the need for complex intermediate feature extraction or cumbersome constraint solving, reducing the occupation of vehicle controller hardware resources and making it easy to deploy and promote in engineering.

[0106] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A multi-objective adaptive control method for a thermal management system of a new energy vehicle battery pack, characterized in that, include: Step S1: A dynamic network model of thermal resistance and thermal capacity of battery module and cooling channel is constructed using the lumped parameter method. Based on this model, a reinforcement learning virtual training environment including state space, action space and environmental feedback mechanism is built. In this model, the internal resistance of battery module is corrected by the battery health status, thereby correcting the heat generation equation of battery module. Based on the nonlinear changes of coolant density, viscosity and thermal conductivity with temperature, the Reynolds number, Prandtl number and convective heat transfer coefficient are updated in real time. Step S2: Select the key node temperature, state of charge, health status, load current and ambient temperature of the battery pack as state variables, and select the coolant flow rate as the control variable. Based on the energy consumption of the cooling system, the temperature control accuracy of the battery pack and the maximum temperature difference of the battery module, construct a multi-objective composite reward function. Step S3: Construct a neural network model based on deep deterministic policy gradient. This neural network model is a dual-network structure containing a policy network and a value network. It introduces an experience replay mechanism and a soft update strategy for the target network. In the reinforcement learning virtual training environment constructed in step S1, through the continuous interaction between the neural network model and the reinforcement learning virtual training environment, the gradient update of the policy network parameters is guided by a multi-objective composite reward function until the neural network model converges. Step S4: Solidify the parameters of the policy network that has been trained and converged in step S3, and deploy them to the vehicle thermal management controller. The vehicle thermal management controller collects vehicle sensor data in real time as state input, and directly outputs the optimal coolant flow rate under the current operating conditions through forward propagation calculation of the policy network. Step S5: The optimal coolant flow rate output in step S4 is mapped to the electric water pump speed command through the characteristic curve of the electric water pump, and sent to the actuator through the CAN bus to realize the adjustment of the battery pack thermal management system. In step S1, the thermal resistance-thermal capacity dynamic network model is constructed through the following process: The heat generation equation for the battery module is established as follows: in, Indicates the first Each battery module Heat generation at all times express Load current at any given time The dynamic internal resistance of the battery module. Indicates the first The temperature of each battery module In a charged state, For a healthy state, The entropy heat coefficient, For the first Each battery module at temperature and the reference internal resistance under the current state of charge, This refers to the aging sensitivity coefficient. The convective heat transfer equation for the battery module is established as follows: in, For the first Convection heat transfer of each battery module The convective heat transfer coefficient is... For effective heat exchange area, For the first The temperature of the coolant within the range of each battery module Let Reynolds number be 1. For Prandtl numbers, The thermal conductivity of the coolant. The thermal conductivity is the reference value. The diameter of the coolant pipe. For coolant density, The reference value is the coolant density. This refers to the coolant flow rate. The dynamic viscosity of the coolant. This is the reference value for the dynamic viscosity of the coolant. This is the temperature calibration value. The specific heat capacity of the coolant. This is the reference value for the specific heat capacity of the coolant; The heat transfer equation between battery modules is established, and the expression is: in, Indicates the first Thermal conductivity of each battery module The thermal conductivity coefficient between adjacent battery modules. This refers to the contact area between battery modules. The characteristic length of thermal conductivity between battery modules. Indicates the first The temperature of each battery module Indicates the first The temperature of each battery module; Establish the energy balance equation for the battery node: in, For the specific heat capacity of a single battery module, The mass of a single battery module; Indicates the first The rate of temperature change of each battery module over time; Establish the energy balance equation for the fluid node: in, For the mass of coolant within a single battery module, Indicates the first Temperature change rate of coolant within the range of each battery module This refers to the coolant mass flow rate. For the first The temperature of the coolant within the range of each battery module; Discretizing the energy balance equations for battery nodes and fluid nodes yields the state update equations for the thermal resistance-thermal capacity dynamic network model: in, Indicates the first Each battery module Temperature at any moment For simulating step size, Indicates the first Each battery module Temperature at any moment for Time of the first The temperature of the coolant within the range of each battery module for Time of the first The temperature of the coolant within the range of each battery module For the first The temperature of the coolant within the range of each battery module.

2. The multi-objective adaptive control method for the thermal management system of a new energy vehicle battery pack according to claim 1, characterized in that, In step S2, the state variable The definition is as follows: in, This indicates the highest temperature of all battery modules in the battery pack at the current moment. This indicates the average temperature of the battery pack at the current moment. This represents the maximum temperature difference between battery modules. This represents the highest temperature change rate of the battery module. For ambient temperature, This indicates transpose.

3. The multi-objective adaptive control method for the thermal management system of a new energy vehicle battery pack according to claim 2, characterized in that, In step S2, the multi-objective composite reward function is constructed. for: in, This represents the reward function for temperature control accuracy. Represents the uniformity penalty function. Indicates energy consumption penalty. , , These are the weighting coefficients. This indicates taking the maximum value. This indicates the coolant flow rate action, satisfying: , This indicates the maximum flow rate that the electric water pump can provide.

4. The multi-objective adaptive control method for the thermal management system of a new energy vehicle battery pack according to claim 3, characterized in that, In step S3, the policy network satisfies the following equation: in, Represents the policy network, Indicates the policy network parameters; The input layer of the policy network contains 8 neurons, which receive 8-dimensional state variables. The hidden layers of the policy network are designed as two fully connected layers with 256 and 128 nodes respectively, and ReLU is used as the activation function between the layers; the output layer of the policy network contains one neuron and uses the Sigmoid activation function. Value networks are used to fit action value functions. To evaluate the quality of the action. Represents a value network. This represents the parameters of the value network. The input layer of the value network contains 9 neurons, which are respectively connected to... and The input is a combination of two layers; the hidden layer of the value network adopts a two-layer fully connected structure with 256 and 128 nodes, and both use the ReLU activation function; the output layer of the value network contains one neuron and directly outputs the corresponding Q value in a linear output manner.

5. The multi-objective adaptive control method for the thermal management system of a new energy vehicle battery pack according to claim 4, characterized in that, Step S4 specifically includes: After the neural network model is trained, the weight matrix and bias vector of the policy network are extracted and embedded into the vehicle thermal management controller. During the online operation phase of the actual vehicle, data acquisition and state estimation are performed first. The vehicle thermal management controller collects the raw sensor signals from the underlying layer, including sensor data of total voltage, load current, and temperature points. Using the state estimation module built into the BMS, the current SOC and SOH are estimated in real time using the ampere-hour integral method combined with the extended Kalman filter algorithm. At the same time, based on the temperature sampling points, online observation or extreme value mapping is performed using the thermal resistance-thermal capacity dynamic network model established in step S1 to estimate the... , and ; The estimated values ​​were then combined with directly measured load current and ambient temperature, and normalized to construct state variables that meet the input requirements of the policy network. , will state variables The input solidified neural network model performs forward propagation calculations and outputs the optimal action under the current working condition in real time. .

6. The multi-objective adaptive control method for the thermal management system of a new energy vehicle battery pack according to claim 5, characterized in that, Step S5 specifically includes: Based on the pressure-flow characteristic curve of the electronic water pump and the flow resistance characteristics of the pipeline, the coolant flow rate is pre-calibrated. With the speed of the electric water pump The fitting function relationship is: in, , , These are calibration coefficients; The vehicle thermal management controller will output the optimal action in step S4. Multiply To obtain the target flow rate Then Substituting the above fitting function into the formula, the electric water pump speed command is calculated and sent to the actuator via the CAN bus to regulate the battery pack thermal management system.

Images