A Physical Information Augmentation Mamba-Driven Reinforcement Learning Approach for Future Thermal Situation Awareness in Vehicle Thermal Management

Abstract

This invention belongs to the field of thermal management control technology for new energy vehicles, and proposes a physical information-enhanced Mamba-driven reinforcement learning method for vehicle thermal management based on future thermal situation awareness. The method first obtains a future state vector within a preset time range based on a thermal situation prediction model trained with physical mechanism constraints. Then, it fuses the current measured battery operating state vector with the future thermal situation prediction vector to construct a reinforcement learning-enhanced state space containing spatiotemporal evolution information. Finally, it inputs the enhanced state space into a reinforcement learning policy network, directly outputting joint control commands for the cooling fan, coolant pump, and cooling circuit valves in the battery cooling system. This invention, by introducing a physical information-enhanced future thermal situation awareness mechanism, expands the observation dimension of the reinforcement learning policy network, realizing a shift in power battery thermal management from passive feedback control to proactive proactive control.

Description

Technical Field

[0001] This invention relates to the field of thermal management control technology for new energy vehicles, and in particular to a physical information-enhanced Mamba-driven reinforcement learning method for future thermal situation awareness in vehicle thermal management, applicable to the intelligent control of power batteries, electric drive systems and vehicle thermal management systems in new energy vehicles or hybrid vehicles. Background Technology

[0002] With the continuous improvement of vehicle electrification and intelligence, vehicle thermal management systems face challenges such as complex operating conditions, frequent fluctuations in heat load, and increasingly higher requirements for control precision. Existing vehicle thermal management control methods mainly have the following shortcomings:

[0003] Feedback lag in PID control: Traditional PID control typically adjusts based on the current temperature deviation, which is a reactive control method. In situations where the vehicle's thermal system has high thermal inertia and significant heat transfer lag, this type of control struggles to respond promptly to rapidly changing heat loads, easily leading to temperature overshoot or system oscillation, thus affecting thermal management effectiveness and energy consumption levels.

[0004] Temporal limitations of traditional reinforcement learning (RL): Most existing reinforcement learning-based vehicle thermal management methods only utilize the current system state for decision-making, lacking explicit modeling of the future evolution trend of the vehicle thermal system. Due to the lack of forward-looking information in the state space, reinforcement learning policy networks struggle to anticipate potential thermal risks, easily leading to control response lag or insufficient stability under complex dynamic conditions.

[0005] To address the aforementioned issues, recent research has attempted to combine data-driven predictive models with reinforcement learning to achieve intelligent control of vehicle thermal management systems. Among these, time-series models such as recurrent neural networks (RNNs) and long short-term memory networks (LSTMs) have been used to predict future temperature trends. However, these models are prone to gradient vanishing or memory decay issues during long-term series modeling, limiting their ability to characterize long-term thermal evolution trends under complex operating conditions.

[0006] Furthermore, while the Transformer architecture has been applied in time series forecasting due to its self-attention mechanism, enabling it to capture long-term dependencies, its computational complexity increases quadratically with the length of the time series. The large model parameter size and high demands on computing and storage resources make it difficult to meet the engineering application requirements of vehicle thermal management systems for real-time performance, low latency, and embedded deployment, thus limiting its widespread application in practical vehicle control systems. On the other hand, some existing methods still use traditional regression models or shallow neural networks for thermal state prediction. These models lack the ability to characterize the dynamic characteristics and temporal correlations of the system, making it difficult to accurately describe the nonlinear evolution process of the vehicle thermal system under complex operating conditions, and resulting in limited stability and generalization ability of the prediction results. Summary of the Invention

[0007] To overcome the problems of large thermal inertia, slow response of traditional feedback control, and lack of foresight in existing vehicle power battery thermal management systems, the present invention aims to provide a physical information augmented Mamba-driven future thermal situation awareness reinforcement learning vehicle thermal management method, which realizes forward-looking collaborative control of thermal management actuators such as electric water pumps, cooling fans, and cooling circuit flow distribution valves.

[0008] To achieve the above objectives, the present invention provides the following solution: a physical information-enhanced Mamba-driven reinforcement learning method for future thermal situation awareness in vehicle thermal management, comprising:

[0009] Step S1: Collect real-time operating data of the vehicle's power battery and thermal management system, construct a current state vector representing the current thermal condition, and construct a historical state set using the current state vectors of the past T moments.

[0010] Step S2: Input the historical state set of the vehicle thermal management system into the preset thermal state prediction model, and output the future state vector of the preset time step; wherein, the thermal state prediction model is built based on the Mamba architecture and introduces physical mechanism constraints during the training phase, so that the thermal state prediction model satisfies the laws of heat transfer and energy conservation when predicting the long-term evolution of the thermal state of the power battery.

[0011] Step S3: Perform feature fusion on the current state vector and the future state vector to construct a reinforcement learning augmented input vector that contains information on the current thermal state of the system and information on the future thermal situation evolution;

[0012] Step S4: Input the reinforcement learning enhancement input vector into the reinforcement learning policy network to directly calculate the joint control action for the vehicle thermal management system;

[0013] Step S5: The combined control action is used to coordinate and regulate the cooling fan, electric water pump and cooling circuit flow distribution valve, so that the cooling system can establish heat dissipation capacity before the temperature exceeds the safety threshold, thereby realizing the active control of the vehicle thermal management system.

[0014] Compared with the prior art, the significant advantages of this invention are:

[0015] This invention provides a physical information-enhanced Mamba-driven reinforcement learning method for vehicle thermal management based on future thermal situational awareness. By introducing future thermal state prediction information into the reinforcement learning decision-making process, it solves the problem of response lag in traditional feedback control and reinforcement learning control based solely on the current state in high thermal inertia vehicle thermal systems. It achieves forward-looking collaborative control of thermal management actuators such as electric water pumps, radiator fans, and cooling circuit flow distribution valves. Furthermore, by constructing a reinforcement learning-enhanced state space that simultaneously includes information on the system's current thermal state and future thermal evolution trends, it addresses the lack of forward-looking perception capabilities and the difficulty in timely identifying potential overheating risks in existing reinforcement learning thermal management methods. The problem is that the control system can establish adjustment capability in advance before the temperature exceeds the safety threshold; by adopting a thermal state prediction model with physical information constraints, the problem of insufficient physical consistency of pure data-driven prediction models under complex operating conditions and rapid changes in thermal load is solved, which improves the reliability and engineering interpretability of future thermal state prediction results, thereby providing stable and reliable prior information for reinforcement learning decision-making; through a continuous control strategy with joint output of multiple actuators, the problem of independent adjustment and insufficient coordination of actuators in traditional vehicle thermal management methods is solved, so that the thermal management system can still maintain good temperature control stability and energy efficiency optimization effect during complex operating condition changes. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the vehicle thermal management method provided in an embodiment of the present invention.

[0018] Figure 2 A flowchart of a vehicle thermal management method provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The purpose of this invention is to provide a physical information-enhanced Mamba-driven reinforcement learning method for vehicle thermal management, enabling forward-looking collaborative control of thermal management actuators such as electric water pumps, radiator fans, and cooling circuit flow distribution valves.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] Figure 1 This is a schematic diagram of the vehicle thermal management method provided in an embodiment of the present invention. Figure 2 A flowchart of a vehicle thermal management method provided in an embodiment of the present invention is shown below. Figure 1 and Figure 2 As shown, this invention provides a physical information-enhanced Mamba-driven reinforcement learning method for future thermal situation awareness in vehicle thermal management, comprising:

[0023] Step S1: Collect real-time operational feedback data of the vehicle thermal management system and its controlled objects. The feedback data includes information such as power battery temperature, coolant temperature, vehicle operating conditions, and thermal management actuator status. Based on the real-time operational data, construct environmental state information reflecting the current thermal condition of the vehicle and form a current state vector. ;

[0024] During vehicle operation, multi-source sensors deployed in the vehicle's thermal management system collect operational data related to thermal management in real time. The collected data includes, but is not limited to: power battery temperature, coolant inlet and outlet temperatures, vehicle operating parameters, battery power or current, coolant flow rate, system pressure, and the current control status of the electric water pump, radiator fan, and flow distribution valves.

[0025] At time t, the current state vector of the system is represented as follows:

[0026]

[0027] For the power battery temperature; This refers to the coolant temperature. , This refers to the power or current of the battery. This refers to the coolant flow rate; , , These are the control quantities for the cooling fan, the electric water pump, and the flow distribution valve, respectively.

[0028] To capture the dynamic evolution characteristics of the vehicle's thermal system, and simultaneously cache the historical state sequence of the past T time points, forming a historical state set:

[0029]

[0030] This historical state sequence serves as the input to the thermal state prediction model.

[0031] Step S2: Based on the historical time-series operation data of the vehicle thermal management system, the historical data is input into a preset thermal state prediction model to predict the thermal state evolution process of the vehicle within a preset time range in the future, and to obtain the future state vector. That is, to obtain a predicted sequence of battery temperature and its changing trend within a future preset time range;

[0032] Furthermore, the thermal state prediction model is built using the Mamba architecture to model the long-term evolution of the thermal state of vehicle thermal management objects. The Mamba architecture, based on a state-space model, can model long-term thermal data with linear time complexity, significantly reducing computational complexity while maintaining modeling accuracy. This makes it suitable for applications in vehicle thermal management systems that require real-time prediction of future thermal states, effectively capturing the long-term temporal correlation characteristics and thermal inertia properties of the power battery's thermal state under complex operating conditions.

[0033] Furthermore, the historical state sequence Inputting data into a pre-defined thermal state prediction model, the model outputs a sequence of thermal states within a pre-defined future prediction time domain.

[0034]

[0035] In the formula, They are respectively , Predicted thermal state at all times The preset prediction time length;

[0036] The aforementioned thermal state prediction model is built upon the Mamba architecture. Mamba employs a state-space model (SSM) to model long-term series, and its hidden state update form can be expressed as:

[0037]

[0038]

[0039] in, These are the hidden states of the model; The input thermal state vector; To predict the output thermal state, The hidden state update value at time t+1 is determined by the hidden state at the previous time and the current input. This is the state transition matrix, used to capture the thermal inertia and historical thermal state evolution characteristics of the power battery; The input matrix is ​​responsible for mapping the currently collected environmental state to the hidden state space; The output matrix maps the hidden states back to the predicted thermal states at the physical level. The direct pass matrix represents the direct influence of the input vector on the output prediction.

[0040] Specifically, to ensure the physical realism and engineering interpretability of the future state vector, and to avoid physical distortion in the pure data-driven prediction model under rapid changes in operating conditions and complex operating conditions, a physical information enhancement strategy is introduced during the training process of the thermal situation prediction model. Its joint loss function is expressed as:

[0041]

[0042] in, This represents the error term between the model's predicted output and the actual measured data. Represents the physical residual term; These are the physical constraint weighting coefficients, used to balance data fitting accuracy with physical consistency.

[0043] Furthermore, the physical residual constraint term is constructed based on the heat transfer mechanism of the vehicle's power battery. The temperature evolution process of a single power battery cell or battery module satisfies the following energy conservation differential equation:

[0044]

[0045] in, The equivalent heat capacity of the power battery; This refers to the core temperature of the power battery. This refers to the heat generated by the battery during charging and discharging. The convective heat transfer coefficient between the power battery and the coolant; For effective heat exchange area; This refers to the coolant temperature.

[0046] The physical residual term can be expressed as:

[0047]

[0048] In the formula, The equivalent heat capacity of the power battery; This refers to the core temperature of the power battery. This refers to the heat generated by the battery during charging and discharging. is the convective heat transfer coefficient between the power battery and the coolant; S is the effective heat transfer area. This refers to the coolant temperature.

[0049] By introducing the aforementioned physical consistency constraints during the Mamba model training process, the thermal state prediction model not only relies on historical data to fit the thermal state change trend but is also explicitly constrained by the power battery heat transfer mechanism. This avoids physical distortions in purely data-driven models under sudden changes in operating conditions or out-of-sample scenarios. Therefore, the physically enhanced Mamba thermal state prediction model can still generate future thermal state sequences that conform to the power battery heat transfer mechanism under complex operating conditions such as rapid changes in operating conditions, sudden power changes, and high-rate charging and discharging. This provides stable, reliable, and physically meaningful forward-looking state information for the reinforcement learning policy network, thereby supporting the vehicle thermal management system to achieve forward-looking control decisions based on future heat load changes.

[0050] Step S3: Perform feature fusion on the current state vector and the future state vector to construct a reinforcement learning augmented input vector that simultaneously contains the vehicle's current thermal state information and future thermal evolution trend information;

[0051] Furthermore, to expand the observation horizon of the reinforcement learning policy network, the current state and the predicted future state are fused, and the constructed augmented state input vector is as follows:

[0052]

[0053] in, To enhance the input vector for reinforcement learning; This is the current state vector, used to represent the thermal condition information of the vehicle thermal management system at the current moment; The future state vector describes the thermal evolution trend of the vehicle's thermal management object within a preset time range; α is the confidence weight coefficient, used to adjust the influence of future state prediction information on the reinforcement learning decision-making process. By introducing the future state vector, the reinforcement learning policy network can identify potential risk states where "the current temperature is not exceeded but there is a tendency for overheating in the future" during decision-making.

[0054] Step S4: Input the reinforcement learning enhancement input vector into the preset reinforcement learning policy network to make forward-looking intelligent decisions and output joint control actions for the vehicle thermal management system;

[0055] Furthermore, the reinforcement learning enhancement input vector is input into the reinforcement learning policy network for decision-making. When the battery temperature corresponding to the current state vector has not yet exceeded the set safety threshold, but the future state vector indicates a rapid temperature rise trend within a preset future time range, the reinforcement learning policy network will output a joint control action to enhance cooling capacity in advance. By utilizing the thermal inertia characteristics of the vehicle's thermal management system, cooling capacity is established in advance before the temperature rises significantly, thereby suppressing heat accumulation and reducing the risk of temperature overshoot.

[0056] Furthermore, the state space, action space, and reward function of the reinforcement learning policy network are defined. The state space includes not only the thermal condition state variables of the vehicle thermal management system at the current moment, but also the future thermal state information generated by the thermal situation prediction model. The action space is a continuous control space used to describe the continuous control commands to the vehicle thermal management actuators. The reward function is constructed around the thermal safety of the thermal management object and the optimization goals of system energy efficiency.

[0057] The reinforcement learning policy network adopts a three-stage architecture of "perception-prediction-decision". A pre-trained Mamba prediction model is integrated into the front end of the policy network. The input layer receives the current thermal condition vector in real time, the Mamba prediction layer infers the state trajectory of the next preset number of steps based on the historical sequence, and finally the fusion layer concatenates the current observation and the future predicted state to synthesize a reinforcement learning augmented input vector containing look-ahead information.

[0058] A policy trajectory is constructed using composite data pairs containing "current state, Mamba predicted state, and joint control action." The policy network is trained using the alternative objective function of the PPO algorithm. By comparing the predicted state with the actual evolution results during training, the agent learns how to react in advance when Mamba predicts a warming trend, thereby achieving globally optimal temperature control logic in policy updates.

[0059] Real-time inference is performed using a trained reinforcement learning policy network. If the current battery temperature is within a safe range, but Mamba predictions indicate a rapid temperature rise in the future, the reinforcement learning policy network will utilize the system's thermal inertia to output joint control actions that enhance cooling capabilities in advance; if the temperature is stable, it will output energy-efficient commands adapted to the current operating conditions.

[0060] The final result is a sequence of combined control actions for the actuators, including water pump flow regulation, fan speed control, and valve allocation ratio. Through coordinated and continuous control of multiple thermal management actuators, the system ensures optimal stability, foresight, and energy efficiency in temperature control under complex operating conditions.

[0061] Preferably, the reinforcement learning policy network employs a policy gradient algorithm, such as the Proximal Policy Optimization (PPO) algorithm. The agent adjusts its strategy based on the environmental state. Output control decision The reward function is defined as follows:

[0062]

[0063] in,

[0064] Furthermore, to comprehensively evaluate control performance, overshoot and convergence time are introduced as evaluation indicators. By comparing the deviation between the actual operating value and the expected value in real time, the strategy network is guided to ensure that the temperature overshoot is within acceptable limits. Approaching the expected value At the same time, control the convergence time Shorten to the expected value This mechanism incentivizes the agent to maximize rewards during training. It is applied to policy updates and multi-step decision-making in the PPO algorithm, and, in conjunction with look-ahead information from the Mamba model, ensures that the final generated combined control actions of the fan, pump, and valve achieve both rapid and stable active cooling, improving the system's stability and temperature control accuracy under complex operating conditions.

[0065] set up:

[0066] This is the actual temperature overshoot. The expected overshoot;

[0067] To control convergence time, The expected convergence time;

[0068] when and At that time, the reward value approaches its maximum value.

[0069] Preferably, reinforcement learning strategy updates and multi-step predictive decision-making

[0070] Let the state-action pairs from the past T time steps constitute the policy trajectory:

[0071]

[0072] The parameters of the reinforcement learning policy network are updated using an alternative objective function based on the PPO algorithm, with the specific formula as follows:

[0073]

[0074] Furthermore, predict the sequence of control decision parameters for the next P steps:

[0075]

[0076] And generate joint control actions:

[0077]

[0078] in, ∈[0, 1] represents the electronic water pump control quantity, which is used to adjust the coolant circulation flow rate via PWM; ∈[0, 1] represents the control parameters for the cooling fan, which enhances heat exchange capacity through speed control; ∈[0, 1] represents the flow distribution valve control value, used to control the distribution ratio of coolant between different cooling branches.

[0079] By jointly and continuously controlling the aforementioned multiple thermal management actuators, the vehicle thermal management system can actively adjust its heat dissipation capacity under different operating conditions, making full use of the system's thermal inertia characteristics to achieve optimal stability, foresight, and energy efficiency in temperature control.

[0080] Step S5: Based on the joint control action, the actuators such as the electronic water pump, cooling fan and cooling circuit flow distribution valve in the vehicle thermal management system are coordinated and regulated to achieve proactive management of the vehicle's thermal state.

[0081] In this invention, the thermal situation prediction model is constructed using the Mamba architecture. The Mamba architecture is a long-sequence modeling structure based on a state-space model, capable of efficiently modeling long-term time-series data with linear time complexity. Compared to prediction models based on recurrent neural networks or Transformer architectures, it significantly reduces computational complexity while maintaining modeling accuracy, making it suitable for thermal state evolution processes in vehicle thermal management systems that exhibit significant thermal inertia and strong temporal correlation. This model allows for continuous prediction of temperature change trends of the vehicle's thermal management object over multiple future time steps, thereby providing forward-looking thermal situation information for the reinforcement learning policy network.

[0082] To improve the physical reliability and engineering interpretability of future thermal situation prediction results, physical mechanism constraints are introduced during the training process of the Mamba thermal situation prediction model. Specifically, during the model training phase, the heat transfer mechanism of the vehicle's thermal management object is embedded as prior knowledge into the model's joint loss function. This ensures that the model's prediction results not only fit historical operating data but also satisfy the basic physical laws of vehicle heat transfer processes. This avoids the problem of physical distortion or prediction drift that can occur in purely data-driven prediction models under complex operating conditions and rapid changes in heat load.

[0083] This invention proposes a physical information-enhanced Mamba-driven reinforcement learning-based vehicle thermal management method for future thermal situation awareness. By constructing a Mamba architecture thermal situation prediction model with physical mechanism constraints and combining it with a reinforcement learning decision-making mechanism, the control of the vehicle thermal management system is transformed from traditional lagging feedback regulation based on the current state to proactive control based on future thermal evolution trends. The Mamba architecture has the ability to efficiently model long-term thermal states, accurately characterizing the dynamic evolution characteristics of the power battery's thermal state while maintaining low computational complexity, thereby significantly improving the adaptability of the control strategy to power battery systems with significant thermal inertia.

[0084] This invention integrates future thermal state information output by the Mamba thermal situation prediction model with the vehicle's current thermal state information to construct an enhanced state space containing spatiotemporal evolution features. This allows the reinforcement learning policy network to identify potential overheating risks before the battery temperature exceeds a safe threshold and to coordinately regulate multiple thermal management actuators, such as the electric water pump, cooling circuit valves, and heat dissipation execution units. This effectively reduces the risk of temperature overshoot and improves the safety and stability of the vehicle's thermal management system. Furthermore, this invention introduces constraints related to the battery heat transfer mechanism during the training process of the Mamba architecture thermal situation prediction model. This ensures that the model can still output a future thermal state sequence that conforms to physical laws under complex and rapidly changing operating conditions, guaranteeing the physical consistency and reliability of the prediction results. This provides engineering-interpretable forward-looking state information support for reinforcement learning decision-making, thereby enhancing the robustness and generalization ability of the vehicle's thermal management system under multiple operating conditions and loads.

[0085] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0086] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A physical information-enhanced Mamba-driven reinforcement learning method for future thermal situation awareness in vehicle thermal management, characterized in that, include: Step S1: Collect real-time operating data of the vehicle's power battery and thermal management system, construct a current state vector representing the current thermal condition, and construct a historical state set using the current state vectors of the past T moments. Step S2: Input the historical state set of the vehicle thermal management system into the preset thermal state prediction model, and output the future state vector at a preset time step; wherein, the thermal state prediction model is built based on the Mamba architecture, and physical mechanism constraints are introduced during the training phase to ensure that the thermal state prediction model satisfies the laws of heat transfer and energy conservation when predicting the long-term evolution of the thermal state of the power battery. The thermal state prediction model uses a state-space model to model the long-term series, and the hidden state update form of the state-space model is as follows: ； ； in, These are the hidden states of the model; The input thermal state vector; To predict the output thermal state; The hidden state update value at time t+1; This is the state transition matrix; The input matrix; This is the output matrix; For direct matrix transmission; The joint loss function used in the thermal situation prediction model is: ； in, This represents the error term between the model's predicted output and the actual measured data. Represents the physical residual term; These are the physical constraint weighting coefficients; specifically, the physical residual term is: ； In the formula, The equivalent heat capacity of the power battery; This refers to the core temperature of the power battery. This refers to the heat generated by the battery during charging and discharging. is the convective heat transfer coefficient between the power battery and the coolant; S is the effective heat transfer area. This refers to the coolant temperature. Step S3: Perform feature fusion on the current state vector and the future state vector to construct a reinforcement learning augmented input vector that contains information on the current thermal state of the system and information on the future thermal situation evolution; Step S4: Input the reinforcement learning enhancement input vector into the reinforcement learning policy network to directly calculate the joint control action for the vehicle thermal management system; Step S5: The combined control action is used to coordinate and regulate the cooling fan, electric water pump and cooling circuit flow distribution valve, so that the cooling system can establish heat dissipation capacity before the temperature exceeds the safety threshold, thereby realizing the active control of the vehicle thermal management system.

2. The physical information augmentation Mamba-driven reinforcement learning vehicle thermal management method for future thermal situation awareness according to claim 1, characterized in that, The historical state set is specifically as follows: ； In the formula, Let be the current state vector at time t, specifically: ； In the formula, For the power battery temperature; This refers to the coolant temperature. , This refers to the power or current of the battery. This refers to the coolant flow rate; , , These are the control quantities for the cooling fan, the electric water pump, and the flow distribution valve, respectively.

3. The physical information augmentation Mamba-driven reinforcement learning method for vehicle thermal management based on claim 1, characterized in that, The reinforcement learning enhancement of input vectors is specifically as follows: ； in, To enhance input vectors for reinforcement learning; This is the current state vector; α represents the future state vector; α is the confidence weight coefficient.

4. The physical information augmentation Mamba-driven reinforcement learning vehicle thermal management method for future thermal situation awareness according to claim 1, characterized in that, The reinforcement learning policy network inputs the reinforcement learning enhancement input vector to the reinforcement learning policy network for decision-making. When the battery temperature corresponding to the current state vector has not exceeded the set safety threshold, and the future state vector indicates that there is a temperature rising trend within a preset time range in the future, the reinforcement learning policy network outputs a joint control action to enhance the cooling capacity.

5. The physical information augmentation Mamba-driven reinforcement learning vehicle thermal management method for future thermal situation awareness according to claim 1, characterized in that, The state space of the reinforcement learning policy network includes the current thermal operating state variables of the vehicle thermal management system and the future thermal state information generated by the thermal situation prediction model; the action space is used to describe the continuous control commands to the vehicle thermal management actuators; the reward function is specifically: ； in, , The reward value representing time t; This is the currently measured temperature of the power battery. The preset target operating temperature; The rate of change of temperature over time; and These represent the real-time power consumption of the cooling fan and the electric water pump, respectively. , As weighting factors, they balance the three core control objectives of temperature control accuracy, operational stability, and overall system energy efficiency.

6. The physical information augmentation Mamba-driven reinforcement learning method for vehicle thermal management based on claim 1, characterized in that, The specific process of reinforcement learning policy update and multi-step prediction decision-making in the reinforcement learning policy network is as follows: Let the state-action pairs from the past T time steps constitute the policy trajectory: ； In the formula, This represents the state over the past T time points. These are actions taken at times T in the past. The parameters of the reinforcement learning policy network are updated using an alternative objective function based on the PPO algorithm, with the specific formula as follows: ； In the formula, This represents the network parameters of the updated reinforcement learning policy; It is the set of policy trajectories consisting of state-action pairs from the past T time points; It is a probability ratio, representing the ratio of the probability of choosing the same action under the same conditions to the probability of choosing the same action under the new strategy and the old strategy; For the dominant function, Indicates state, Indicates an action; It is a pruning term in the PPO algorithm. It is a limiting factor; Based on the updated policy parameters, the sequence of control decision parameters for the next P steps is predicted, and the joint control action is dynamically adjusted according to the dynamic changes of the future state vector.

7. The physical information-enhanced Mamba-driven reinforcement learning method for vehicle thermal management based on claim 1, characterized in that, The joint control action is specifically as follows: ； in, ∈[0, 1] represents the electronic water pump control quantity; ∈[0, 1] represents the control value for the cooling fan; ∈[0, 1] represents the flow distribution valve control quantity.

Images