Power battery active thermal management system control method based on reinforcement learning

Abstract

According to the power battery active thermal management system control method based on reinforcement learning, a power battery air cooling system comprises a battery pack and a temperature control assembly composed of a draught fan and an air conditioner, the battery pack is composed of battery monomers distributed in a rectangular array, and the battery monomers are cylindrical batteries and are the same in axial direction; according to the method, a multi-scene dynamic working condition data set is generated by utilizing a simulation environment, the temperature of the battery module and the energy consumption of the system are accurately controlled on line through a data driving algorithm, and intelligent optimization strategies such as a genetic algorithm are introduced on the basis. The online self-adaptive updating of the inlet air speed and the refrigeration power is realized; meanwhile, a decision boundary is optimized, and an online switching module of a fan, an air conditioner and a combination mode thereof is designed to automatically adapt to different working condition requirements; compared with an existing control strategy, the energy efficiency performance, the real-time performance and the temperature uniformity of a control system on complex dynamic working conditions are remarkably improved, adaptability is higher, and robustness is higher.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power battery air-cooled thermal management, and particularly relates to a power battery active thermal management system control method based on reinforcement learning. BACKGROUND

[0002] The existing thermal management system methods mainly focus on two fields: structural design optimization and control strategy development. Passive heat dissipation is low-cost but limited in effect, liquid cooling is high in efficiency but complex in system and high in cost, and the air-cooled scheme can improve the heat dissipation performance to a certain extent through optimization of series-parallel flow channels, inlet and outlet arrangement and multi-inlet design, but becomes ineffective in high-temperature environments. For the air-cooled ABTMS, external energy (such as the battery pack of an electric vehicle) is used to drive the fan and air conditioner to circulate and pre-cool the airflow, thereby improving the cooling performance. However, without a suitable control strategy, the ABTMS may not be able to provide effective cooling to ensure the battery life and safety, despite the increased energy consumption. Therefore, it is crucial to study the comprehensive control of the vehicle air conditioner and fan to enhance the applicability and practicality of the ABTMS and provide valuable guidance for the actual deployment of electric vehicles.

[0003] The existing air-cooled ABTMS has the following main deficiencies in control strategy: first, the data dependency is high, real vehicle experiment is costly and difficult to cover all extreme conditions, and the real vehicle data set hinders the training and verification of data-driven models; second, the strategy adaptability is poor, existing methods are mostly based on fixed rules or model derivation, and there are optimization control problems such as multi-objective, complex model and unknown mechanism, making it difficult to effectively cope with complex and variable conditions in actual operation, such as US06, CLTC and WLTC, especially in multi-input, multi-output and strongly coupled systems, the control effect is unstable; third, the multi-objective trade-off is insufficient, the optimization ability is limited by the mutual restraint of battery pack temperature uniformity and system energy consumption optimization, existing methods mostly focus on a single target, lack of dynamic balance, and are prone to local optimum; fourth, the real-time performance is insufficient, especially in fast-changing conditions, it is difficult to balance temperature control and energy consumption optimization.

[0004] Although the fuel cell thermal management system data-driven iterative learning controller design method (CN116009390A) uses a data-driven iterative learning control method, it still has the following deficiencies: first, the controller needs to repeatedly learn and converge in the iteration domain, and the real-time performance is limited; the target only focuses on temperature tracking accuracy, without considering energy consumption optimization. The control optimization in the thermal management system of an electric vehicle and the operation method of the thermal management system (CN119365362A) belongs to offline pre-training or short-term prediction, and the lower-level control output is fixed on the model parameters and optimization algorithm at the training time, lacking online boundary adjustment. SUMMARY

[0005] In view of the deficiencies in the prior art, the application provides a control method for a power battery active thermal management system based on reinforcement learning, aiming to improve the thermal management efficiency.

[0006] The control method for the power battery active thermal management system based on reinforcement learning, the power battery air cooling system includes a battery pack and a temperature control component composed of a fan and an air conditioner, the battery pack is composed of battery monomers arranged in a rectangular array, the battery monomers are all cylindrical batteries and have the same axial direction, the outlet of the temperature control component is close to and faces the side surface of the battery pack, and the data-driven thermal management method includes the following steps. Step 1: Obtain a training sample data set; Step 2: Establish a random forest regression model based on the training sample data set; Step 3: Take the random forest regression model as a predictor, adjust the state of the predictor to be consistent with the initial state of the temperature control component, and set the target function vector of the predictor :

[0007] Among them, 、 and are the gains of each item, and the behavior of the control algorithm can be adjusted by adjusting these parameters; represents the average temperature, represents the expected average temperature, is the number of times the temperature difference exceeds the upper limit, represents the average power consumption; Step 4: Adopt the NSGA-II strategy to adaptively optimize the predictor to obtain an optimal predictive controller; Step 5: By introducing a mode switching module, the optimal predictive controller dynamically adjusts the control strategy according to the real-time temperature condition.

[0008] Further, step 1 specifically includes the following steps: Step 1.1: Establish a simulation environment for the power battery air cooling system; Step 1.2: Randomly initialize the simulation environment to obtain the initial temperature and initial SoC of the battery module; Step 1.3: Use different random seeds to generate the fan output power, inlet air speed and refrigeration power of the temperature control component, and input the simulation environment; Step 1.4: Repeat step 1.3, record the battery module temperature, SoC and system power consumption output by the simulation environment, and obtain simulation data under a random working condition; Step 1.5: Repeat steps 1.2 to 1.4 to obtain a set amount of simulation data and form a sample data set; Step 1.6: Select 70% of the sample dataset as the training sample dataset and 30% as the test sample dataset.

[0009] Further, step 2 specifically includes the following steps: Step 2.1: Perform self-sampling on the training sample dataset to obtain a new dataset consistent with the original training set size; Step 2.2: Construct a decision tree based on the new dataset; Step 2.3: Repeat steps 2.1 to 2.2 until the process is recursively repeated to construct a complete decision tree (wherein the process is performed in a recursive manner within each tree, i.e., at each node, the data is divided according to the optimal feature, and then the division process is repeated for the divided subsets. Until the current branch node meets the stopping condition (the number of node samples is less than the set threshold or all sample target variable values are the same), the construction of the decision tree is completed); Step 2.4: Repeat steps 2.1 to 2.3 until a specified number of decision tree sets are generated, and all decision tree sets are aggregated to form a random forest regression model; Step 2.5: The final output of the random forest regression model is the arithmetic mean of the regression results of each decision tree, i.e., the final prediction value; Step 2.6: Verify and optimize the random forest regression model using the test sample dataset.

[0010] Further, step 4 specifically includes the following steps: Step 4.1: At sampling time k, randomly generate a set of control signal trajectory sets according to the prediction step and the population size; Step 4.2: Input each control signal trajectory into the predictor for multi-step recursive prediction, and calculate the objective function vector ; Step 4.3: Calculate the fitness function of each control signal trajectory using the objective function vector r, and search iteratively through the genetic algorithm until the convergence condition is met, thereby obtaining the control signal trajectory with the optimal index value; Step 4.4: Output the actual control step item in the optimal control signal trajectory to the temperature control component in the power battery air-cooled system; Step 4.5: The temperature control component in the power battery air-cooled system responds to the optimal control signal to obtain the actual output, and synchronizes this response to the predictor for feedback correction, so that the predictor remains consistent with the state of the control object after completing one control, and corrects the accumulated prediction error; Step 4.6: Repeat steps 4.1 to 4.5 to continuously adjust backward until the simulation stops.

[0011] Further, the multi-step recursive prediction in step 4.1 is specifically recursive prediction, that is, when the step length is a set prediction step length, the recursive prediction uses the same predictor in each step, and uses the output of the previous step predictor as the input of the current step predictor.

[0012] Further, the control signal trajectory set in step 4.1 is: wherein, is the i-th control signal trajectory in the control signal trajectory set, is the control input at time k, and the control input is the inlet air speed and the refrigeration power of the simulation environment.

[0013] Further, the control step length is 1.

[0014] Further, the module switching rule in the mode switching module is: High-temperature control mode: when the average temperature of the battery module is higher than the upper limit of switching, the thermal management system is in the high-temperature control zone, and the inlet air speed and the refrigeration power are both in the range of 0~100%: Low-temperature control mode: when the average temperature is between the upper limit and the lower limit of switching, the thermal management system is switched to the low-temperature control mode, and the inlet air speed and the refrigeration power are both in the range of 0~50%: Shutdown mode: when the average temperature is lower than the lower limit of switching, the thermal management control system is turned off, and the inlet air speed and the refrigeration power are both set to zero.

[0015] Advantages of the present application: the method generates a multi-scenario dynamic working condition data set using a simulation environment, and uses a data-driven algorithm to accurately control the battery module temperature and system energy consumption online, and on this basis, introduces an intelligent optimization strategy such as a genetic algorithm to realize online adaptive updating of the inlet air speed and the refrigeration power; at the same time, the decision boundary is optimized, and an online switching module of the fan, the air conditioner and their combination mode is designed to automatically adapt to different working condition requirements; compared with existing control strategies, the present application significantly improves the energy efficiency performance, real-time performance and temperature uniformity of the control system under complex dynamic working conditions, and has stronger adaptability and higher robustness. Three evaluation indexes, average maximum temperature difference, average temperature deviation and average power, are proposed, a data set covering different states is generated using a random seed, a rolling optimization is improved based on a random forest algorithm through a genetic algorithm NSGA-II, and a mode switching module is added to optimize the control trajectory boundary, solving the problems of large calculation amount of evaluation control target effect, insufficient diversity of data set, poor adaptability, multi-objective high-dimensional optimization and local optimal problem of online dynamic control in existing data-driven methods. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1Flow chart for the present application; Figure 2 Schematic diagram for prediction and control time domain in the present application; Figure 3 Improved random forest prediction control algorithm; Figure 4 Comparison of average temperature and maximum temperature difference control effects of different algorithms Figure 5 Comparison of control effects of different algorithms; Figure 6 Comparison of average temperature and maximum temperature difference control effects before and after adding the mode switching module; Figure 7 Comparison of control effects before and after adding the mode switching module; Figure 8 Comparison of prediction control average temperature and maximum temperature difference control effects under different ambient temperatures. DETAILED DESCRIPTION

[0017] The present application will be described in detail below with reference to the accompanying drawings. The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the drawings, wherein the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the drawings are exemplary and are only used to explain the present application, and should not be interpreted as limiting the present application. The left, middle, right, up, down and other orientation terms in the embodiments of the present application are only relative concepts or are referenced to the normal use state of the product, and should not be considered as limiting.

[0018] The control method of the active thermal management system of the power battery based on reinforcement learning, the power battery air cooling system includes a battery pack and a temperature control component composed of a fan and an air conditioner, the battery pack is composed of 32 battery monomers distributed in a rectangular array, the battery monomers are all cylindrical batteries and have the same axial direction, the outlet of the temperature control component is close to and faces the side surface of the battery pack, the height of the flow channel is 46.91mm, and the data-driven thermal management method includes the following steps: Step 1: obtaining a training sample data set; specifically including the following steps: Step 1.1: establishing a simulation environment of the power battery air cooling system; Step 1.2: randomly initializing the simulation environment to obtain the initial temperature of the battery module and the initial SoC, the initial temperature of the battery module ranges from 293.15K to 308.15K, and the initial SoC ranges from 50% to 80%, with the average maximum temperature difference, the average temperature offset and the average power as evaluation indexes; thereby, the extended dynamic working condition data set is constructed by using the simulation platform with the real working condition data support, different initial temperatures, SOC and control variable combinations are covered, and sufficient training and prediction samples are provided for the data-driven algorithm; Step 1.3: Generate the fan output power of the temperature control component with different random seeds , the inlet air speed , and the refrigeration power , and input them into the simulation environment; For the jth random seed , let , , The independent variable at time t is:

[0019]

[0020] wherein represents a uniformly distributed random number under the jth random seed; are the upper and lower limits of the motor power, inlet air speed, and refrigeration power, respectively; Step 1.4: Repeat Step 1.3 and record the battery module temperature, SoC, and system power consumption output by the simulation environment under a random working condition to obtain simulation data; Step 1.5: Repeat Steps 1.2 to 1.4 to obtain a set amount of simulation data and form a sample data set; To ensure the diversity, sufficiency, and adaptability of the data, this section sets the length of each random working condition to 3000s and repeats the data generation 6 times, finally obtaining a sample data set with a total length of 18000s. The random forest regression model will be established based on this in the future; Step 1.6: Select 70% of the sample data set as the training sample data set and 30% as the test sample data set; Step 2: Establish a random forest regression model based on the training sample data set; specifically including the following steps: Step 2.1: Since the training sample data set has a large difference in the order of magnitude of each feature, normalization processing is required for the training sample data set; the training sample data set is subjected to Bootstrap sampling, i.e., randomly sampling a single sample with replacement from n simulation data, and repeating this step n times, thereby obtaining a new data set :

[0021] wherein is the ith sample data obtained by sampling with replacement from the training sample data set . Step 2.2: Constructing the decision tree based on the new data set; randomly selecting f features (f is less than the total number of features) at each node, and selecting an optimal feature according to the decision index among the extracted features to divide the left and right subtrees of the decision tree. For a certain candidate split point, the data set is divided into left subset and right subset according to the feature, so that the minimum, the total mean square error (MSE) of the node is defined as: wherein, is the average value of the target variable in the data set ; is the true target value of the i-th sample; is the number of samples in the data set ; and and are the number of samples in the left and right subsets; Step 2.3: repeating steps 2.1 to 2.2 until the process is recursively repeated to construct a complete decision tree, wherein the process is performed in a recursive manner within each tree, i.e. at each node, the data is divided according to the optimal feature, and then the subset after division is continuously repeated; until the current branch node meets the stopping condition of the node sample number being less than the set threshold or all samples having the same target variable value, the construction of the decision tree is completed; Step 2.4: repeating steps 2.1 to 2.3 until a specified number of decision tree sets are generated, and all the decision tree sets are collected to form a random forest regression model, and the random forest regression model parameters are set as shown in Table 1; For regression problems, the final output of the regression model is the arithmetic mean of the regression results of each decision tree;

[0022] wherein, is the prediction value of the i-th tree for the sample; N is the total number of generated trees; is the final prediction value of the random forest model for the sample; Step 2.5: the final output of the random forest regression model is the arithmetic mean of the regression results of each decision tree, i.e. the final prediction value; Step 2.6: verifying the random forest regression model using the test sample data set, and adjusting the parameters if the effect is not good, and retraining and optimizing; Step 2.6: verifying the random forest regression model using the test sample data set, and adjusting the parameters if the effect is not good, and retraining and optimizing; ​The results of error analysis and comparison of multiple linear regression and random forest regression based on the original values of the test set are shown in Table 2; it can be concluded that the random forest regression is superior to the multiple linear regression in terms of root mean square error (RMSE) and mean relative error (MRE) of temperature, SoC and power consumption, in other words, the random forest regression has better regression performance on the problem studied in this method and meets the accuracy requirements;

[0023] Step 3: Take the random forest regression model as a predictor, adjust the state of the predictor to be consistent with the initial state of the temperature control component, and set the target function vector of the predictor :

[0024] wherein, , and are the gains of each item, and the behavior of the control algorithm can be adjusted by adjusting these parameters; represents the average temperature, represents the expected average temperature, is the number of times the temperature difference exceeds the upper limit, represents the average power consumption; Step 4: Adopt the NSGA-II strategy to adaptively optimize the predictor to obtain an optimal predictive controller; specifically including the following steps: Step 4.1: At sampling time k, a set of control signal trajectory sets are randomly generated according to the (set) prediction step and the population number; multi-step recursive prediction is a recursive prediction, that is, when the step is the set prediction step, the recursive prediction uses the same predictor at each step and uses the output of the previous step predictor as the input of the current step predictor; the control signal trajectory set is: wherein, is the i-th control signal trajectory in the control signal trajectory set, is the control input at time k, and the control input is the inlet air speed and refrigeration power of the simulation environment; Step 4.2: input each control signal trajectory into the predictor for multi-step recursive prediction, and calculate the target function vector through the predicted response; wherein the predicted response is a series of predicted values of the battery module temperature, SoC and system power consumption of the simulation environment within the prediction step of the output of the predictor; Step 4.3: Calculate the fitness function R of each control signal trajectory using the objective function vector r, and search iteratively through the genetic algorithm until the convergence condition is met, so as to obtain the control signal trajectory with the optimal index value; that is, when the difference between the maximum value of the crowding distance of all N individuals on the Pareto front of the Gth generation and the (G-1) th generation is less than the threshold value ε, the control signal trajectory with the optimal index value is obtained:

[0025] Wherein: R is the fitness function, that is, the multi-dimensional objective function value vector, and are the objective function vectors of the fan and the refrigeration system, is the adjacent fitness value after sorting on the target j, and M is the number of target functions of the objective function vector; Step 4.4: output the actual control step in the optimal control signal trajectory to the temperature control component in the power battery air-cooled system; wherein the actual control step is the actual control step in the rolling time domain, and the actual control step is less than or equal to the prediction step ; Wherein, is the optimal control sequence after optimization, the optimal control signal trajectory actually input to the temperature control component; Step 4.5: the temperature control component in the power battery air-cooled system responds to the optimal control signal to obtain the actual output , and synchronizes this response to the predictor for feedback correction, so that the predictor is consistent with the state of the control object after completing one control, corrects the cumulative prediction error, and is updated to:

[0026] Wherein, is the state estimation value for recursion in the predictor; Step 4.6: repeat steps 4.1 to 4.5, and continuously adjust backward until the simulation stops, that is, through the feedback correction of step 4.5, the actual output ​​A synchronous back-predictor is used to correct the prediction starting point at the next time; in combination with the step 4.6 rolling optimization, the real-time correction of the model cumulative error and the continuous self-adaptation of the system state change are realized by the controller; wherein, the backward adjustment refers to that, as the control time domain rolling window advances, the predictor moves backward the prediction starting time with a fixed control step, and the control signal trajectory is regenerated on the basis of the new state; the condition for stopping the simulation is that: when the prediction starting point of the rolling time domain has advanced to the termination time of the simulation working condition, i.e. the sampling time k is greater than or equal to the simulation working condition length-prediction step, it is determined that the simulation is ended; As shown in Figure 2 , the selection of the prediction time domain length (prediction step) and the control time domain length (actual control step) will greatly affect the calculation efficiency and control effect of the algorithm; generally speaking, setting a relatively large prediction step can effectively reduce system fluctuations and better cope with the delay system, but it will also bring greater calculation amount; for the control step, in order to ensure the control effect, the prediction step needs to increase with the increase of the control step, therefore, in order to control the prediction step within a reasonable range, the control step is usually set to 1; Thus, in the random forest predictive control framework, the intelligent optimization strategy of the genetic algorithm is combined to solve the high-dimensional nonlinear control problem, avoid the traditional algorithm from falling into local optimum, and realize the multi-objective dynamic balance of temperature and energy consumption by predicting and updating the inlet wind speed and refrigeration power; Step 5: By introducing a mode switching module, the optimal predictive controller dynamically adjusts the control strategy according to the real-time temperature condition; the module switching rule in the mode switching module is: High-temperature control mode: when the average temperature of the battery module is higher than the upper limit of switching, i.e. , the thermal management system is in the high-temperature control area, and the value range of the inlet wind speed and the refrigeration power at this time is from 0 to 100%; Low-temperature control mode: when the average temperature is between the upper limit and the lower limit of switching, i.e. , the thermal management system is switched to low-temperature control, and at this time, the value range of the inlet wind speed and the refrigeration power is reduced to 0-50%; Shutdown mode: when the average temperature is lower than the lower limit of switching, i.e. , the thermal management control system is turned off, and at this time, the inlet wind speed and the refrigeration power are both set to zero; Wherein, represents the average temperature, represents the mode switching temperature offset parameter, represents the desired average temperature; Thus, the decision boundary is optimized, the fan, air conditioner and combined mode are automatically switched online through online working condition, the control flexibility is significantly improved, and the system robustness is enhanced.

[0027] By adjusting the fan and air conditioning control strategy in real time, the cooling system energy consumption is effectively reduced and the battery thermal management performance is improved, and the charging and discharging working condition changes are quickly adapted. Through the extended US06 cycle verification, the average temperature offset is reduced by 28.0%, the average temperature difference is reduced by 61.6%, the average power consumption is reduced by 22.0%, and the control response time is shortened to 170s, which is obviously superior to the existing method. It shows that the present application further improves the thermal management efficiency, enhances the system adaptability and real-time performance in terms of control mode, methodology and performance optimization, which is beneficial to prolong the battery life and improve the running safety of the vehicle, and has stronger innovation and practicality. Embodiment 1

[0028] This section will discuss the improved random forest predictive control algorithm (forest predictive control algorithm with a mode-switching mechanism, RFPC-MSM) as Figure 3 , and the control performance in the thermal management system compared with traditional control algorithms (including empirical-based and model-based).

[0029]

[0030] For empirical-based algorithms, this section selects the rule-based method (rule-based, RB), and for model-based algorithms, the MPC (model-based predictive control) algorithm is selected, and the MPC algorithm also adds a mode switching module. The control rules set for RB are: first set the inlet air speed and refrigeration power to the maximum value, until the average temperature of the battery module drops to the control off threshold, set the inlet air speed and refrigeration power to zero; when the battery module temperature rebounds to the control on threshold, set the inlet air speed and refrigeration power to the maximum value, and so on. The parameter settings of each algorithm are shown in Table 3, and the parameters not mentioned are consistent with the previous text.

[0031] As Figure 4 , the time required for MPC to reach the target temperature is the longest, which is 1200s, and the time required for RB and RFPC is 170s and 620s respectively. At the same time, RB performs poorly in controlling the maximum temperature difference of the battery module, and the maximum temperature difference has been exceeding the set maximum temperature difference since 90s. MPC performs better than RB in controlling the maximum temperature difference, and the total time length exceeding the allowed maximum temperature difference is 610s, but RFPC has better performance, and this data is only 100s.

[0032] As Figure 5Among the three algorithms, RB performs the worst in all aspects, MPC performs the second, and RFPC performs the best. Compared with RB, RFPC can reduce the average temperature deviation by 28.0%, reduce the average maximum temperature difference by 61.6%, and reduce the average power consumption by 22.0%. Compared with MPC, RFPC can reduce the average temperature deviation by 22.1%, reduce the average maximum temperature difference by 18.4%, and reduce the average power consumption by 6.8%. Embodiment 2

[0033] The algorithm configuration is the same as that in Embodiment 1, as shown in Figure 6 When the prediction step is 10s, the prediction control system without the mode switching module reaches 296K at the end of the test working condition, which is still lower than the target average temperature. In contrast, the system with the mode switching module reaches the target average temperature for the first time at 620s, and then always keeps near the target temperature, only with a small deviation (<1.5K) in the interval of 1550-2200s. After adding the mode switching module, the time of the maximum temperature difference of the battery module exceeding the allowable temperature difference is greatly shortened (from 550s to 100s).

[0034] As shown in Figure 7 , the process average temperature of the system with the mode switching module is closer to the target temperature in the whole test working condition, the average temperature deviation is reduced by 76.7% compared with the system without the mode switching module, the average maximum temperature difference is reduced by 25.8%, and the average power consumption of the control system is reduced by 40.5%. Embodiment 3

[0035] The influence of the environment temperature, as shown in Figure 8 , for the two working conditions in which the environment temperature is higher than the target temperature, the control difficulty of the average temperature and the maximum temperature difference increases when the environment temperature is higher. For the working condition in which the environment temperature is consistent with the target temperature, the method can well control the average temperature and the maximum temperature difference, and basically controls the average temperature near the target temperature except for a slight deviation near 1100s.

[0036] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above embodiments, and the above embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A control method of a power battery active thermal management system based on reinforcement learning, characterized in that: The power battery air cooling system comprises a battery pack and a temperature control assembly composed of a fan and an air conditioner, the battery pack is composed of battery monomers arranged in a rectangular array, the battery monomers are all cylindrical batteries and have the same axial direction, the outlet of the temperature control assembly is close to and faces the side of the battery pack, and the data-driven thermal management method comprises the following steps: Step 1: obtaining a training sample data set; Step 2: establishing a random forest regression model based on the training sample data set; Step 3: Random forest regression model as predictor, and adjust the state of the predictor to be consistent with the initial state of the temperature control component, and set the target function vector of the predictor : ； wherein, , and are the gains of each term, by adjusting these parameters the behavior of the control algorithm can be tuned; denotes the average temperature, denotes the desired average temperature, is the number of times the temperature difference exceeds the upper limit, denotes the average power consumption; Step 4: adaptively optimizing the predictor by using the NSGA-II strategy to obtain an optimal predictive controller; Step 5: by introducing a mode switching module, the optimal predictive controller dynamically adjusts the control strategy according to the real-time temperature condition.

2. The control method of the active thermal management system for power batteries based on reinforcement learning according to claim 1, characterized in that: Step 1 specifically comprises the following steps: Step 1.1: establishing a simulation environment of the power battery air cooling system; Step 1.2: randomly initializing the simulation environment to obtain the initial temperature and initial SoC of the battery module; Step 1.3: using different random seeds to generate the fan output power, inlet air speed and refrigeration power of the temperature control assembly, and inputting the simulation environment; Step 1.4: repeating step 1.3 to record the battery module temperature, SoC and system power consumption output by the simulation environment to obtain simulation data under a random working condition; Step 1.5: repeating steps 1.2 to 1.4 to obtain a set amount of simulation data and form a sample data set; Step 1.6: selecting 70% from the sample data set as the training sample data set and 30% as the test sample data set.

3. The control method of the active thermal management system for power batteries based on reinforcement learning according to claim 1, characterized in that: Step 2 specifically comprises the following steps: Step 2.1: using the bootstrap sampling method on the training sample data set to obtain a new data set consistent with the original training set; Step 2.2: constructing a decision tree based on the new data set; Step 2.3: repeating steps 2.1 to 2.2 until the process is recursively repeated to construct a complete decision tree; Step 2.4: repeating steps 2.1 to 2.3 until a specified number of decision tree sets are generated, and all the decision tree sets are combined to form a random forest regression model; Step 2.5: the final output of the random forest regression model is the arithmetic mean of the regression results of each decision tree, which is the final prediction value; Step 2.6: verifying and optimizing the random forest regression model by using the test sample data set.

4. The control method of the active thermal management system for power batteries based on reinforcement learning according to claim 1, characterized in that: Step 4 specifically comprises the following steps: Step 4.1: at sampling time k, a set of control signal trajectory sets are randomly generated according to the prediction step and the population size; Step 4.2: Each control signal trajectory is input into the predictor for multi-step recursive prediction and the objective function vector is computed by the predicted response ; Step 4.3: using the objective function vector r to calculate the fitness function of each control signal trajectory, and through iterative search by the genetic algorithm, until the convergence condition is met, thereby obtaining the control signal trajectory with the optimal index value; Step 4.4: outputting the actual control step item in the optimal control signal trajectory to the temperature control assembly in the power battery air cooling system; Step 4.5: the temperature control assembly in the power battery air cooling system responds to the optimal control signal to obtain the actual output, and synchronizes the response to the predictor for feedback correction, so that the predictor keeps consistent with the state of the control object after completing a control, and corrects the accumulated prediction error; Step 4.6: Repeat steps 4.1 to 4.5, adjusting backward constantly until the simulation stops.

5. The control method of the active thermal management system for power batteries based on reinforcement learning according to claim 4, characterized in that: The multi-step recursive prediction in step 4.1 is specifically recursive prediction, that is, the recursive predictor uses the same predictor in each step and uses the output of the previous step predictor as the input of the current step predictor when the step length is set to the set prediction step length.

6. The control method of the active thermal management system for power batteries based on reinforcement learning according to claim 4, characterized in that: The control signal trajectory set in step 4.1 is: ； wherein, is the i-th control signal trajectory in the control signal trajectory set, is the control input at time k, the control input is the inlet wind speed and the refrigeration power of the simulation environment.

7. The control method of the active thermal management system for power batteries based on reinforcement learning according to claim 4, characterized in that: The control step length is 1.

8. The control method of the active thermal management system for power batteries based on reinforcement learning according to claim 1, characterized in that: The module switching rule in the mode switching module is: High temperature control mode: when the average temperature of the battery module is higher than the upper limit of switching, the thermal management system is in the high temperature control zone, and the value range of the inlet air speed and the refrigeration power at this time is from 0~100%: Low temperature control mode: when the average temperature is between the upper limit and the lower limit of switching, the thermal management system is switched to the low temperature control, at this time, the value range of the inlet air speed and the refrigeration power is reduced to 0~50%: Shutdown mode: when the average temperature is lower than the lower limit of switching, the thermal management control system is turned off, and the inlet air speed and the refrigeration power are both set to zero.

Images