Electric vehicle whole vehicle thermal management method based on iTransformer time series prediction and TD3 reinforcement learning

Abstract

The application discloses an electric vehicle whole vehicle thermal management method based on iTransformer time series prediction and TD3 reinforcement learning, comprising the following steps: 1, fusing vehicle networking cloud end traffic flow data and vehicle driving data to pre-process and construct a feature matrix; 2, using an iTransformer network to predict vehicle speed, acceleration and divide power consumption levels; 3, building a reinforcement learning state containing road conditions, temperature, comfort, power consumption levels and an action set of an executing mechanism, and constructing a multi-dimensional segmented compound reward function; 4, building a double Critic network based on a TD3 algorithm, and obtaining an optimal Actor model through interactive sample collection and iterative training, which is used for real-time regulation and control of vehicle thermal management. The application can realize driving load pre-judgment by fusing cloud end road condition data, complete thermal management pre-regulation in advance, reduce instantaneous high power consumption of a heat dissipation system, balance power component temperature control, cabin comfort and air conditioner energy consumption at the same time, thereby improving the adaptive control level of whole vehicle thermal management and reducing the energy consumption of the whole vehicle.

Description

Technical Field

[0001] This invention belongs to the field of thermal management technology for new energy vehicles, specifically a thermal management method for electric vehicles based on iTransformer time-series prediction and TD3 reinforcement learning. Background Technology

[0002] With the rapid development of the new energy vehicle industry, integrated heat pump thermal management has become a core supporting system for pure electric vehicles. A single circuit simultaneously completes the temperature control of the power battery, the heat dissipation of the drive motor, and the heating and cooling of the cabin. The system has multiple components and multiple circuits that are highly coupled. During vehicle operation, external conditions such as road congestion, slopes, and sunlight continue to change. The control process needs to simultaneously consider the reliability of the power components, the thermal comfort of the ride, and the energy consumption of the air conditioning. Multiple objectives are mutually constrained, making control optimization quite difficult.

[0003] Currently, most mass-produced vehicles use calibrated MAP threshold control or single-loop PID feedback control. These solutions rely solely on onboard sensors to collect real-time temperature and power signals for passive adjustment. Constrained by the inherent hysteresis of the heat transfer system, they can only initiate cooling or heating actions after component temperatures deviate from the ideal range, failing to anticipate load changes caused by the road ahead. When vehicles enter high-load conditions such as highway driving, hill climbing, or traffic jams, the heat generated by the battery and motor rises rapidly. The system can only rely on increasing the speed of the compressor and cooling pump to enhance heat dissipation, significantly increasing parasitic power consumption in thermal management and reducing the vehicle's range. At the same time, linear control logic is difficult to adapt to the strong nonlinear coupling characteristics of the thermal system, easily leading to frequent temperature fluctuations and unstable cabin comfort under dynamic conditions.

[0004] To address the lag issues inherent in traditional control methods, the industry is gradually introducing time-series prediction and deep reinforcement learning to conduct research on intelligent control. At the level of driving condition prediction, most existing time-series prediction models rely solely on local historical vehicle speed and current data for modeling, rarely integrating external time-series information such as traffic flow and road conditions from the vehicle-to-everything (V2X) cloud platform. Lacking support from global road condition characteristics ahead, these models suffer from significant errors in predicting future speed and acceleration trends, making it difficult to support early pre-regulation by thermal management systems.

[0005] At the level of reinforcement learning control schemes, existing vehicle thermal management control mostly uses the DDPG algorithm to handle continuous control variables. This algorithm only sets up a single evaluation network, and the iterative process has the inherent defect of Q-value overestimation. The reward fluctuations during the training phase are large, the convergence stability is poor, and the control robustness is insufficient under the variable road conditions of real vehicles. At the same time, the design dimensions of existing reinforcement learning reward functions are single-dimensional. Most of them only construct reward and punishment logic for a single indicator such as temperature or energy consumption, without distinguishing different driving load levels to set segmented constraints. They cannot simultaneously balance the optimization objectives of multiple mutually restrictive factors such as power component temperature control, cabin thermal comfort, air conditioning energy consumption, and solar heat load, and the effect of multi-objective collaborative equilibrium control is limited. Summary of the Invention

[0006] This invention addresses the shortcomings of existing technologies by proposing a vehicle thermal management method for electric vehicles based on iTransformer time-series prediction and TD3 reinforcement learning. This method aims to integrate vehicle network cloud-based traffic data to predict future vehicle load and pre-regulate the thermal management system, thereby mitigating the surge in heat dissipation and power consumption caused by concentrated component heating under high-load conditions, effectively improving the vehicle's actual driving range. Furthermore, it coordinates and optimizes multiple constraints, including power component operational stability, cabin thermal comfort, and air conditioning energy consumption, to enhance the adaptive control capability of the vehicle's thermal management system in complex roads and changing environments, delaying battery and motor aging, and balancing driving safety and passenger experience. Ultimately, this enhances the engineering application value of intelligent thermal management solutions for new energy vehicles.

[0007] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: The present invention provides a method for thermal management of electric vehicles based on iTransformer temporal prediction and TD3 reinforcement learning, characterized by the following steps: Step 1: Obtain historical traffic flow data of the road segment where the target vehicle is located from the vehicle-to-everything (V2X) cloud platform, and simultaneously obtain the target vehicle's own historical driving data. Perform time alignment and data preprocessing on the historical traffic flow data and the historical driving data to construct a multi-dimensional input feature matrix. ;in, For time steps, The number of variables; Step 2: Construct the iTransformer temporal prediction network, including: embedding layer, multi-head self-attention sub-layer, projection layer, and perform... The process is performed to obtain the predicted vehicle speed sequence. and acceleration sequence ; Step 3: Based on and The predicted power consumption level of the drive motor is obtained. ; like ,and Then let =1; like ,and Then let =2; like ， and Then let =2; like ,and Then let =3; like Then let =3; in, Indicates the first-level speed threshold. This represents the second-order velocity threshold, and < , Indicates the first-order acceleration threshold. This represents the second-order acceleration threshold, and < ; Step 4: Constructing the state parameters for TD3 deep reinforcement learning ,in, Indicates light intensity. Indicates the temperature of the drive motor. Indicates battery temperature. This indicates the power consumption of the vehicle's heat pump air conditioning system. This indicates the comfort level of the driver's cab. Constructing action parameters for TD3 deep reinforcement learning ,in, This indicates the speed of the vehicle's heat pump air conditioning compressor. Indicates the speed of the air conditioning fan in the driver's cab. The speed of the cooling pump in the battery cooling circuit. The speed of the cooling pump in the drive motor cooling circuit; Step 5: Construct the reward function for TD3 deep reinforcement learning : Step 6: Construct the TD3 deep reinforcement learning mesh, including: an Actor policy network and a target Actor policy network with the same structure, two Critic evaluation networks with the same structure and their respective target Critic evaluation networks; Step 6.1: Initialize the parameters of the Actor policy network as follows: The parameters for initializing the two Critic evaluation networks are as follows: and ; Step 6.2: Set the current time State parameters The input is processed in the Actor policy network, and the current time is obtained using Equation (8). Vehicle motion parameters And calculate the current time. Reward value : (8) In equation (9), For the current moment random noise, Represents the Actor policy network; Step 6.3: When the vehicle executes Then, the next moment is obtained. State parameters Thus with , , as well as Termination status symbol A sample is formed and added to the experience buffer; Step 6.4: Obtain several samples following the process in Steps 6.2-6.3 until the empirical buffer is full; Step 6.5: Randomly sample a batch of data of size [size missing] from the experience playback buffer. The samples are used to train the TD3 deep reinforcement learning grid to obtain the optimal Actor policy network; Step 7: Deploy the optimal Actor network model in the electric vehicle thermal management system control to process the input state parameters and output the vehicle heat pump air conditioning compressor speed, cab air conditioning fan speed, battery cooling circuit cooling pump speed, and drive motor cooling circuit cooling pump speed to optimize cab comfort and air conditioning power consumption.

[0008] The electric vehicle thermal management method based on iTransformer temporal prediction and TD3 reinforcement learning described in this invention is characterized in that step 2 includes the following steps: Step 2.1: Embedding layer pair After transposition, the transposed multidimensional input feature matrix is ​​processed using an MLP layer. Perform mapping processing to obtain the token matrix H; Step 2.2: The multi-head self-attention sub-layer uses equation (1) to obtain the query vector of any attention head. Key vector Value vector Thus, the attention score of any attention head can be obtained using the self-attention mechanism. Then, the attention scores of all attention heads are residually connected to obtain a weighted fusion matrix. Then, layer normalization is performed to obtain the normalized matrix. ; (1) In equation (1), , , These are three projection weight matrices. For activation functions; Step 2.3: The feedforward network sublayer utilizes equation (2) to... A nonlinear transformation is performed to obtain the token matrix. : (2) In equation (2), , These are the two weight matrices of the two-layer perceptron; , These are the two bias matrices of a two-layer perceptron; Indicates the activation function; Step 2.3: Using the MLP layer as the projection head, project the token matrix. The high-dimensional features of each token are projected onto the predicted time step to output the predicted vehicle speed sequence. and acceleration sequence .

[0009] Furthermore, step 5 includes the following steps: Step 5.1: Construct the reward function for comfort using equation (3). : (3) In equation (3), The weighting parameter for comfort; Step 5.2: Construct the reward function for the drive motor and battery using equation (4). : (4) In equation (4), This represents the temperature weighting parameter of the battery. This represents the temperature weighting parameter of the drive motor. For ambient temperature, This indicates the battery's set operating temperature. This indicates the set operating temperature of the drive motor. The set temperature deviation value; Step 5.3: Construct the reward function for air conditioner power consumption using equation (5). : (5) In equation (5), , Set the maximum and minimum power for the air conditioner respectively. Weighting parameters for air conditioner power consumption; Step 5.4: Construct a power consumption reward function considering external illumination using equation (6). : (6) Step 5.5: Construct the reward function using equation (7) : (7).

[0010] Furthermore, step 6.5 includes the following steps: Step 6.5.1: Transfer the state parameters from the i-th sample The input to the target Actor policy network is processed, and the parameters of the i-th action are output. Therefore, equation (9) is used to... Add clipping noise to obtain the motion parameters of the i-th motion with added clipping noise. : (9) In equation (10), N is a normal distribution. Indicates variance; Step 6.5.2: [The sentence is incomplete and requires more context to translate accurately.] and The data are input into two target Critic evaluation networks for processing, and the i-th first target is obtained accordingly. value and the i-th second objective value Thus, the i-th target can be obtained using equation (11). value : (10) In equation (11), As a discount factor, For the set goal value; Step 6.5.3: Set the i-th state parameter and the i-th action parameter The data are processed by two separate Critic evaluation networks to obtain the i-th first-ranked data. value and the i-th second value ; Step 6.5.4: Construct the loss function of the first Critic evaluation network using equations (11) and (12) respectively. And the loss function of the second Critic evaluation network. ; (11) (12) Step 6.5.5: Obtain the updated parameters of the two Critic evaluation networks using equations (13) and (14) respectively. and : (13) (14) In equations (13)-(14), The loss gradient of the first Critic evaluation network; The loss gradient of the second Critic evaluation network is used. The learning rate; Step 6.5.6: When the number of updates for the two Critic evaluation networks reaches the set number d, the current state parameters are... The input to the Actor policy network is processed, and the current action parameters are output. and will The input is processed in the first Critic evaluation network after d updates to obtain the first target value. The loss function is calculated and used to backpropagate the Actor policy network, thereby obtaining the updated parameters of the Actor policy network using equation (15). : (15) In equation (15), Let be the loss gradient of the Actor policy network; Step 6.5.7: After updating the Actor policy network, use equations (16), (17), and (18) to adjust the parameters of the target Actor policy network. The primary objective is to evaluate network parameters using the Critic system. The second objective is to evaluate the network parameters using the Critic evaluation algorithm. Perform a soft update to obtain the updated target Actor policy network parameters. Updated first-objective Critic evaluation network parameters Updated second-objective Critic evaluation network parameters : (16) (17) (18) In equations (16), (17), and (18), This is a soft update coefficient; Step 6.5.8: Iterate through the TD3 deep reinforcement learning mesh according to the process of steps 6.5.1-6.5.7 until the set maximum number of iterations is reached. This process continues until the optimal Actor policy network is obtained.

[0011] The present invention provides an electronic device, including a memory and a processor, characterized in that the memory is used to store a program supporting the processor in performing the method described therein, and the processor is configured to execute the program stored in the memory.

[0012] The present invention discloses a computer-readable storage medium storing a computer program, characterized in that the computer program is executed by a processor to perform the steps of the method described thereon.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention integrates cloud-based road condition data and vehicle driving data to construct a feature matrix, and uses iTransformer to complete operating condition prediction and load level classification. It solves the problems of existing solutions that rely solely on local data, lack operating condition classification, and have lagging control. It can pre-adjust the thermal management system in advance, reduce peak power consumption for heat dissipation, improve vehicle range, and make the system adaptable to various dynamic driving conditions.

[0014] 2. This invention constructs a multi-dimensional state and action parameter system and designs a segmented multi-objective composite reward function. Combined with a dual-Critic network and a delayed update strategy, the TD3 algorithm architecture solves the shortcomings of traditional reinforcement learning, such as a single reward dimension, Q-value overestimation in DDPG, and unstable training. It can simultaneously balance component temperature control, cabin comfort, and air conditioning energy consumption, reduce component temperature fluctuations, and improve control stability under complex road conditions.

[0015] 3. This invention deploys the trained optimal model on-board to form a closed-loop control, overcoming the shortcomings of traditional PID and threshold control's passive adjustment, realizing adaptive intelligent temperature control, slowing down battery and motor aging, and improving the intelligent level of vehicle thermal management. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a diagram of the iTransformer temporal prediction network structure of the present invention; Figure 3 This is a flowchart of the TD3 deep reinforcement learning method of the present invention. Detailed Implementation

[0017] In this embodiment, as Figure 1As shown, a method for thermal management of electric vehicles based on iTransformer temporal prediction and TD3 reinforcement learning includes the following steps: Step 1: Obtain historical traffic flow data of the road segment where the target vehicle is located from the vehicle-to-everything (V2X) cloud platform, including traffic flow information and road grade. Simultaneously, obtain the target vehicle's own historical driving data, including speed, acceleration, deceleration, and average speed. Perform time alignment and data preprocessing on the historical traffic flow data and historical driving data, dividing the previous 90 seconds into n time segments. Combine the data from each segment with the time segment to define a vector, with a vector dimension of [missing value]. Construct a multidimensional input feature matrix using n sets of vectors. .

[0018] Step 2: As Figure 2 As shown, the iTransformer temporal prediction network is constructed, including: an embedding layer, a multi-head self-attention sub-layer, and a projection layer, to predict the velocity and acceleration sequences of vehicles within the next 60 seconds.

[0019] Step 2.1: Embedding layer pair After transposition, the transposed multidimensional input feature matrix is ​​processed using an MLP layer. The mapping process is performed to obtain the token matrix H.

[0020] Step 2.2: The multi-head self-attention sub-layer uses equation (1) to obtain the query vector of any attention head. Key vector Value vector Thus, the attention score of any attention head can be obtained using the self-attention mechanism. Then, the attention scores of all attention heads are residually connected to obtain a weighted fusion matrix. Then, layer normalization is performed to obtain the normalized matrix. ; (1) In equation (1), , , These are three projection weight matrices. This is the soft function.

[0021] Step 2.3: The feedforward network sublayer utilizes equation (2) to... A nonlinear transformation is performed to obtain the token matrix. : (2) In equation (2), , These are the two weight matrices of the two-layer perceptron; , These are the two bias matrices of a two-layer perceptron; This represents the activation function.

[0022] Step 2.3: Using the MLP layer as the projection head, project the token matrix. The high-dimensional features of each token are projected onto the predicted time step to output the predicted vehicle speed sequence. and acceleration sequence .

[0023] Step 3: Based on and The predicted power consumption level of the drive motor is obtained. The higher the predicted speed and acceleration, the higher the predicted power consumption level of the drive motor, which is used for subsequent control of the vehicle's thermal management system. When the predicted power consumption level of the drive motor is not high, in order to ensure that the overall power consumption of the heat pump air conditioning is within the normal range, the cooling efficiency of the drive motor cooling circuit can be reduced, thereby reducing the overall power consumption. When the predicted power consumption level of the drive motor is high, the cooling efficiency of the drive motor cooling circuit can be increased in advance to prevent a sudden rise in the drive motor temperature later.

[0024] like ,and Then let =1; like ,and Then let =2; like ， and Then let =2; like ,and Then let =3; like Then let =3; in, Indicates the first-level speed threshold. Indicates the second-level speed threshold. Indicates the first-order acceleration threshold. This represents the second-order acceleration threshold, and < .

[0025] Step 4: Constructing the state parameters for TD3 deep reinforcement learning ,in, Indicates light intensity. Indicates the temperature of the drive motor. Indicates battery temperature. This indicates the power consumption of the vehicle's heat pump air conditioning system. This indicates the comfort value of the driver's cab. The comfort value is obtained using a PMV calculator. The PMV value is calculated using real-time air data obtained from sensors such as temperature and humidity sensors in the passenger cabin. The PMV value represents the average index of the group's vote on thermal comfort. There are 7 levels, with values ​​ranging from -3 to 3. Negative values ​​are too cold, and positive values ​​are too hot. The further away from 0, the more uncomfortable it is.

[0026] Constructing action parameters for TD3 deep reinforcement learning ,in, This indicates the speed of the vehicle's heat pump air conditioning compressor. Indicates the speed of the air conditioning fan in the driver's cab. Cooling pump speed for battery cooling circuit To drive the cooling pump speed of the motor cooling circuit, the overall cooling effect of the vehicle thermal management system is adjusted by regulating the speed of the vehicle heat pump air conditioning compressor. The cab temperature can be adjusted to a certain extent by adjusting the speed of the cab air conditioning fan. The battery cooling pump speed and the drive motor cooling pump speed can respectively regulate the battery temperature and the motor temperature.

[0027] Step 5: Construct the reward function for TD3 deep reinforcement learning : Step 5.1: Construct the reward function for comfort using equation (3). : (3) In equation (3), The weighting parameter for comfort.

[0028] Step 5.2: Construct the reward function for the drive motor and battery using equation (4). : (4) In equation (4), This represents the temperature weighting parameter of the battery. This represents the temperature weighting parameter of the drive motor. For ambient temperature, This indicates the battery's set operating temperature. This indicates the set operating temperature of the drive motor. This is the set temperature deviation value.

[0029] When the predicted power level is Level 1 and Level 2, the temperature of the drive motor is relatively low after prediction, so only the deviation between the battery temperature and the set value is considered as the bonus value. At Level 2, the bonus value is increased when the same battery temperature deviation value is increased. When the predicted power level is Level 3, the temperature of the drive motor is considered.

[0030] The optimal operating temperature range for power batteries is 15℃ to 35℃; the safe operating range for normal vehicle operation is 0℃ to 40℃. The maximum operating temperature for ternary lithium batteries should not exceed 60℃, and the maximum operating temperature for lithium iron phosphate batteries should not exceed 65℃; there is a risk of thermal runaway when the cell temperature remains above 60℃. In low-temperature environments, high-current charging and discharging are not permitted below -20℃; the tolerable storage temperature range for batteries is -40℃ to 85℃.

[0031] The optimal steady-state operating temperature range for water-cooled drive motors is 60℃~90℃; the maximum continuous operating temperature of F-class insulated motor windings commonly used in passenger vehicles is 155℃, and the maximum continuous operating temperature of H-class insulated motor windings is 180℃. During vehicle road testing, the motor temperature must not exceed 175℃, and the maximum operating temperature of the matching motor controller is 95℃; the motor can only withstand a peak temperature of up to 200℃ for a short period of 30 seconds.

[0032] Comprehensive , , .

[0033] Step 5.3: Construct the reward function for air conditioner power consumption using equation (5). : (5) In equation (5), , Set the maximum and minimum power for the air conditioner respectively. This is a weighted parameter for air conditioner power consumption.

[0034] Step 5.4: Construct a power consumption reward function considering external illumination using equation (6). : (6) Step 5.5: Construct the reward function using equation (7) : (7) Step 6: Construct the TD3 deep reinforcement learning mesh, including: an Actor policy network and a target Actor policy network with the same structure, two Critic evaluation networks with the same structure and their respective target Critic evaluation networks.

[0035] Step 6.1: As Figure 3 As shown, the parameters for initializing the Actor policy network are: The parameters for initializing the two Critic evaluation networks are as follows: and .

[0036] Step 6.2: Set the current time State parameters The input is processed in the Actor policy network, and the current time is obtained using equation (8). Vehicle motion parameters And use equation (7) to calculate the current time. Reward value : (8) In equation (9), For the current moment random noise, This represents the Actor policy network.

[0037] Step 6.3: When the vehicle executes Then, the next moment is obtained. State parameters Thus with , , as well as Termination status symbol A sample is formed and added to the experience buffer.

[0038] Step 6.4: Obtain several samples following the process in Steps 6.2-6.3 until the experience buffer is full.

[0039] Step 6.5: Randomly sample a batch of data of size [size missing] from the experience playback buffer. The samples are used to train the TD3 deep reinforcement learning grid to obtain the optimal Actor policy network; Step 6.5.1: Transfer the state parameters from the i-th sample The input to the target Actor policy network is processed, and the parameters of the i-th action are output. Therefore, equation (9) is used to... Add clipping noise to obtain the motion parameters of the i-th motion with added clipping noise. : (9) In equation (10), N is a normal distribution. Indicates variance; Step 6.5.2: [The sentence is incomplete and requires more context to translate accurately.] and The data are input into two target Critic evaluation networks for processing, and the i-th first target is obtained accordingly. value and the i-th second objective value Thus, the i-th target can be obtained using equation (11). value : (10) In equation (11), As a discount factor, For the set goal value; Step 6.5.3: Set the i-th state parameter and the i-th action parameter The data are processed by two separate Critic evaluation networks to obtain the i-th first-ranked data. value and the i-th second value ; Step 6.5.4: Construct the loss function of the first Critic evaluation network using equations (11) and (12) respectively. And the loss function of the second Critic evaluation network. ; (11) (12) Step 6.5.5: Obtain the updated parameters of the two Critic evaluation networks using equations (13) and (14) respectively. and : (13) (14) In equations (13)-(14), The loss gradient of the first Critic evaluation network; The loss gradient of the second Critic evaluation network is used. The learning rate; Step 6.5.6: When the number of updates for the two Critic evaluation networks reaches the set number d, the current state parameters are... The input to the Actor policy network is processed, and the current action parameters are output. and will The input is processed in the first Critic evaluation network after d updates to obtain the first target value. The loss function is calculated and used to backpropagate the Actor policy network, thereby obtaining the updated parameters of the Actor policy network using equation (15). : (15) In equation (15), Let be the loss gradient of the Actor policy network; Step 6.5.7: After updating the Actor policy network, soft update the target Actor policy network parameters using equations (16), (17), and (18) respectively. The primary objective is to evaluate network parameters using the Critic system. The second objective is to evaluate the network parameters using the Critic system. ; (16) (17) (18) In equations (16), (17), and (18), For the updated target Actor policy network parameters, To evaluate the network parameters for the updated first objective Critic, To evaluate the network parameters for the updated second objective, Critic. , This refers to the soft update coefficient. Step 6.5.8: Iterate through the TD3 deep reinforcement learning mesh according to the process of steps 6.5.1-6.5.7 until the set maximum number of iterations is reached. This process continues until the optimal Actor policy network is obtained. Step 7: Lighten the trained optimal Actor policy network and deploy it to the electric vehicle's onboard edge computing controller; during actual vehicle operation, collect real-time data on illumination, motor temperature, battery temperature, air conditioning power consumption, and in-vehicle PMV comfort, and combine this data with the predicted power consumption level output in real time by the iTransformer network to form a state vector which is then fed into the optimal Actor policy network model.

[0040] The model outputs four types of continuous control commands in real time: heat pump compressor speed, air conditioning fan speed, battery cooling pump speed, and motor cooling pump speed. These commands are then sent to the corresponding actuators to synchronously adjust the air conditioning cooling power and the heat dissipation capacity of the dual cooling circuits. The vehicle's sensors continuously collect data on the temperature, power consumption, and in-vehicle comfort of each component in real time, forming a closed-loop feedback loop to continuously and dynamically adjust the output of the thermal management system.

[0041] This solution relies on iTransformer to predict the future driving load of the vehicle in advance, and combines TD3 multi-objective reinforcement learning to collaboratively optimize the in-vehicle driving comfort and the power consumption of the vehicle's air conditioning system. Under the premise of meeting the safe operating temperature range of the battery and motor, it effectively reduces the energy consumption of the vehicle's thermal management system.

Claims

1. A method for thermal management of electric vehicles based on iTransformer temporal prediction and TD3 reinforcement learning, characterized in that, Includes the following steps: Step 1: Obtain historical traffic flow data of the road segment where the target vehicle is located from the vehicle-to-everything (V2X) cloud platform, and simultaneously obtain the target vehicle's own historical driving data. Perform time alignment and data preprocessing on the historical traffic flow data and the historical driving data to construct a multi-dimensional input feature matrix. ;in, For time steps, The number of variables; Step 2: Construct the iTransformer temporal prediction network, including: embedding layer, multi-head self-attention sub-layer, projection layer, and perform... The process is performed to obtain the predicted vehicle speed sequence. and acceleration sequence ; Step 3: Based on and The predicted power consumption level of the drive motor is obtained. ; like ,and Then let =1; like ,and Then let =2; like ， and Then let =2; like ,and Then let =3; like Then let =3; in, Indicates the first-level speed threshold. This represents the second-order velocity threshold, and < , Indicates the first-order acceleration threshold. This represents the second-order acceleration threshold, and < ; Step 4: Constructing the state parameters for TD3 deep reinforcement learning ,in, Indicates light intensity. Indicates the temperature of the drive motor. Indicates battery temperature. This indicates the power consumption of the vehicle's heat pump air conditioning system. This indicates the comfort level of the driver's cab. Constructing action parameters for TD3 deep reinforcement learning ,in, This indicates the speed of the vehicle's heat pump air conditioning compressor. Indicates the speed of the air conditioning fan in the driver's cab. The speed of the cooling pump in the battery cooling circuit. The speed of the cooling pump in the drive motor cooling circuit; Step 5: Construct the reward function for TD3 deep reinforcement learning : Step 6: Construct the TD3 deep reinforcement learning mesh, including: an Actor policy network and a target Actor policy network with the same structure, two Critic evaluation networks with the same structure and their respective target Critic evaluation networks; Step 6.1: Initialize the parameters of the Actor policy network as follows: The parameters for initializing the two Critic evaluation networks are as follows: and ; Step 6.2: Set the current time State parameters The input is processed in the Actor policy network, and the current time is obtained using Equation (8). Vehicle motion parameters And calculate the current time. Reward value : (8) In equation (9), For the current moment random noise, Represents the Actor policy network; Step 6.3: When the vehicle executes Then, the next moment is obtained. State parameters Thus with , , as well as Termination status symbol A sample is formed and added to the experience buffer; Step 6.4: Obtain several samples following the process in Steps 6.2-6.3 until the empirical buffer is full; Step 6.5: Randomly sample a batch of data of size [size missing] from the experience playback buffer. The samples are used to train the TD3 deep reinforcement learning grid to obtain the optimal Actor policy network; Step 7: Deploy the optimal Actor network model in the electric vehicle thermal management system control to process the input state parameters and output the vehicle heat pump air conditioning compressor speed, cab air conditioning fan speed, battery cooling circuit cooling pump speed, and drive motor cooling circuit cooling pump speed to optimize cab comfort and air conditioning power consumption.

2. The electric vehicle thermal management method based on iTransformer temporal prediction and TD3 reinforcement learning according to claim 1, characterized in that, Step 2 includes the following steps: Step 2.1: Embedding layer pair After transposition, the transposed multidimensional input feature matrix is ​​processed using an MLP layer. Perform mapping processing to obtain the token matrix H; Step 2.2: The multi-head self-attention sub-layer uses equation (1) to obtain the query vector of any attention head. Key vector Value vector Thus, the attention score of any attention head can be obtained using the self-attention mechanism. Then, the attention scores of all attention heads are residually connected to obtain a weighted fusion matrix. Then, layer normalization is performed to obtain the normalized matrix. ; (1) In equation (1), , , These are three projection weight matrices. For activation functions; Step 2.3: The feedforward network sublayer utilizes equation (2) to... A nonlinear transformation is performed to obtain the token matrix. : (2) In equation (2), , These are the two weight matrices of the two-layer perceptron; , These are the two bias matrices of a two-layer perceptron; Indicates the activation function; Step 2.3: Using the MLP layer as the projection head, project the token matrix. The high-dimensional features of each token are projected onto the predicted time step to output the predicted vehicle speed sequence. and acceleration sequence .

3. The electric vehicle thermal management method based on iTransformer temporal prediction and TD3 reinforcement learning according to claim 2, characterized in that, Step 5 includes the following steps: Step 5.1: Construct the reward function for comfort using equation (3). : (3) In equation (3), The weighting parameter for comfort; Step 5.2: Construct the reward function for the drive motor and battery using equation (4). : (4) In equation (4), This represents the temperature weighting parameter of the battery. This represents the temperature weighting parameter of the drive motor. For ambient temperature, This indicates the battery's set operating temperature. This indicates the set operating temperature of the drive motor. The set temperature deviation value; Step 5.3: Construct the reward function for air conditioner power consumption using equation (5). : (5) In equation (5), , Set the maximum and minimum power for the air conditioner respectively. Weighting parameters for air conditioner power consumption; Step 5.4: Construct a power consumption reward function considering external illumination using equation (6). : (6) Step 5.5: Construct the reward function using equation (7) : (7)。 4. The electric vehicle thermal management method based on iTransformer temporal prediction and TD3 reinforcement learning according to claim 3, characterized in that, Step 6.5 includes the following steps: Step 6.5.1: Transfer the state parameters from the i-th sample The input to the target Actor policy network is processed, and the parameters of the i-th action are output. Therefore, equation (9) is used to... Add clipping noise to obtain the motion parameters of the i-th motion with added clipping noise. : (9) In equation (10), N is a normal distribution. Indicates variance; Step 6.5.2: [The sentence is incomplete and requires more context to translate accurately.] and The data are input into two target Critic evaluation networks for processing, and the i-th first target is obtained accordingly. value and the i-th second objective value Thus, the i-th target can be obtained using equation (11). value : (10) In equation (11), As a discount factor, For the set goal value; Step 6.5.3: Set the i-th state parameter and the i-th action parameter The data are processed by two separate Critic evaluation networks to obtain the i-th first-ranked data. value and the i-th second value ; Step 6.5.4: Construct the loss function of the first Critic evaluation network using equations (11) and (12) respectively. And the loss function of the second Critic evaluation network. ; (11) (12) Step 6.5.5: Obtain the updated parameters of the two Critic evaluation networks using equations (13) and (14) respectively. and : (13) (14) In equations (13)-(14), The loss gradient of the first Critic evaluation network; The loss gradient of the second Critic evaluation network is used. The learning rate; Step 6.5.6: When the number of updates for the two Critic evaluation networks reaches the set number d, the current state parameters are... The input to the Actor policy network is processed, and the current action parameters are output. and will The input is processed in the first Critic evaluation network after d updates to obtain the first target value. The loss function is calculated and used to backpropagate the Actor policy network, thereby obtaining the updated parameters of the Actor policy network using equation (15). : (15) In equation (15), Let be the loss gradient of the Actor policy network; Step 6.5.7: After updating the Actor policy network, use equations (16), (17), and (18) to adjust the parameters of the target Actor policy network. The primary objective is to evaluate network parameters using the Critic system. The second objective is to evaluate the network parameters using the Critic system. Perform a soft update to obtain the updated target Actor policy network parameters. Updated first-objective Critic evaluation network parameters Updated second-objective Critic evaluation network parameters : (16) (17) (18) In equations (16), (17), and (18), This is a soft update coefficient; Step 6.5.8: Iterate through the TD3 deep reinforcement learning mesh according to the process of steps 6.5.1-6.5.7 until the set maximum number of iterations is reached. This process continues until the optimal Actor policy network is obtained.

5. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports a processor in executing the method of any one of claims 1-4, the processor being configured to execute the program stored in the memory.

6. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program is executed by the processor to perform the steps of the method according to any one of claims 1-4.

Images