An Intelligent Control Method for Thermal Management Systems of New Energy Vehicles Based on Artificial Intelligence

Abstract

This invention belongs to the field of new energy vehicle thermal management application technology, and particularly relates to an intelligent control method for a new energy vehicle thermal management system based on artificial intelligence. The method includes constructing a dynamic spatiotemporal graph and multi-physics field perception. Multi-source sensor data, including battery cells, motor windings, and air conditioning vent thermal components, are defined as nodes in the spatiotemporal graph. A dynamic adjacency matrix characterizes the thermal coupling strength, solving the problem of fragmented multi-physics field modeling in traditional methods. A graph attention network and gated recurrent units are used to construct the spatiotemporal graph convolution output spatiotemporal graph features. A differentiable computational fluid dynamics (CFD-PINN) thermal load prediction model is used, embedding the Navier-Stokes equations as hard constraints into the neural network to ensure that the prediction results strictly conform to thermodynamic laws. A meta-policy network generates target weights based on the state, achieving multi-objective adaptive optimization. Lower-level control commands change the actual thermal state of the new energy vehicle and are fed back to update the spatiotemporal graph, forming a closed-loop control.

Description

Technical Field

[0001] This invention belongs to the field of new energy vehicle thermal management application technology, and particularly relates to an intelligent control method for a new energy vehicle thermal management system based on artificial intelligence. Background Technology

[0002] With the development of new energy vehicles, the importance of thermal management systems has become increasingly prominent. Vehicles generate a significant amount of heat during operation, and an efficient thermal management system can effectively regulate this heat, thereby improving overall vehicle performance and extending its driving range. The thermal management systems for new energy vehicles (such as pure electric, hybrid, and fuel cell vehicles) need to simultaneously optimize battery life (temperature uniformity), motor efficiency (cooling requirements), cabin comfort (PMV index), and overall vehicle energy consumption. Traditional methods face challenges such as: multi-physics coupling, strong coupling between battery heat generation, coolant flow, and cabin heat transfer, difficulty in global modeling, dynamic environmental adaptation, and complex and variable operating conditions due to extreme temperatures and driving mode switching. Furthermore, traditional solutions often separate data-driven models from physical equations, leading to unreliable long-term predictions, decision-making dimension explosion, and difficulty in converging multi-objective optimization in continuous state-action spaces. Summary of the Invention

[0003] This invention addresses the technical problems existing in the thermal management system of new energy vehicles by proposing an intelligent control method for the thermal management system of new energy vehicles based on artificial intelligence. This method is reasonable in design, simple in approach, theoretically sound, and capable of realizing multi-physics field coupling and multi-objective coordination in the thermal management system.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is: an intelligent control method for a new energy vehicle thermal management system based on artificial intelligence, comprising the following steps:

[0005] S1. Construct a dynamic spatiotemporal graph and multi-physics sensing. Define the multi-source sensor data of battery cells, motor windings, and air conditioning vent thermal components as nodes in the spatiotemporal graph. The feature vectors of the N nodes at time t contain real-time data of temperature, flow rate, and current, respectively, to obtain the node matrix H at time t. (t) The adjacency matrix is ​​calculated based on thermal resistance, temperature difference, and mass flow rate of the connecting edges to characterize the thermal coupling strength. The formula for calculating the dynamic adjacency matrix is ​​as follows:

[0006]

[0007] in, Let ΔT be the adjacency matrix connecting edges i and j at time t. ij R represents the real-time temperature difference between nodes i and j. ij Let m be the thermal resistance between nodes i and j. ijLet σ() be the mass flow rate connecting edges i and j, σ be the sigmoid function, and w1 and w2 be the weight hyperparameters. A spatiotemporal graph convolution is constructed using a graph attention network and gated recurrent units, and the node matrix H at time t is... (t) The convolution of the adjacency matrix into the spatiotemporal graph outputs the spatiotemporal graph features h. (t) ;

[0008] S2. The differentiable computational fluid dynamics (CFD)-PINN heat load prediction model discretizes the Navier-Stokes equations into differentiable operators, which serve as hard constraints for the neural network. It is then jointly trained with the physical neural network PINN, and the loss function is defined as follows:

[0009]

[0010] in, For each data item, N is the number of nodes, and T is the number of nodes. pred To predict temperature, T real For actual measured temperature, Let represent the residuals of the Navier-Stokes equations, u be the fluid velocity field, p be the pressure field, ρ be the fluid density, and v be the dynamic viscosity. The equation is a continuity equation, where λ and μ are physical constraint weighting coefficients, and the spatiotemporal graph features h are used. (t) Input a differentiable computational fluid dynamics (CFD) model for predicting heat load; output the heat load distribution Q for the next 30 seconds. pred and predicting the fluid velocity field u pred ;

[0011] S3, Meta-reinforcement learning multi-objective dynamic decision-making, extracting heat load distribution Q pred Maximum temperature max(T) pred ), mean flow velocity (u) pred and battery temperature difference ΔT bat The remaining battery charge (SOC) and driving mode are used as inputs to construct meta-reinforcement learning state variables. The meta-policy network outputs dynamic target weights, which are then used to generate lower-level control commands to set the coolant flow rate.

[0012] The lower-level control commands S4 and S3 change the actual thermal state of the new energy vehicle, and the new multi-source sensor data is fed back to S1 to update the spatiotemporal diagram, forming a closed-loop control.

[0013] Preferably, the mass flow rate of the connecting edge in step S1 is the mass flow rate of the coolant, the spatiotemporal graph convolution uses a spatial aggregation graph attention network to fuse neighbor node information, solves the local coupling in the heterogeneous graph structure, and the gated loop unit captures the node state evolution.

[0014] Preferably, the dynamic target weights output by the meta-policy network include vehicle energy consumption optimization weights w.能耗 Battery temperature uniformity weight w 温差 and passenger cabin thermal comfort weight w 舒适 Satisfying w 能耗 +w 温差 +w 舒适 =1.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0016] This invention proposes an intelligent control method for the thermal management system of new energy vehicles based on artificial intelligence. This method includes constructing a dynamic spatiotemporal graph and multi-physics field perception. Multi-source sensor data, including battery cells, motor windings, and air conditioning vent thermal components, are defined as nodes in the spatiotemporal graph. A dynamic adjacency matrix characterizes the thermal coupling strength, addressing the problem of fragmented multi-physics field modeling in traditional methods. Spatial aggregation GAT and GRU units are used to construct spatiotemporal graph convolution and output spatiotemporal graph features. A differentiable computational fluid dynamics (CFD)-PINN thermal load prediction model is used, embedding the Navier-Stokes equations as hard constraints into the neural network to ensure that the prediction results strictly conform to thermodynamic laws. A meta-policy network generates target weights based on the state, achieving multi-objective adaptive optimization. Lower-level control commands change the actual thermal state of the new energy vehicle and are fed back to update the spatiotemporal graph, forming a closed-loop control. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the structure of an intelligent control method for a new energy vehicle thermal management system based on artificial intelligence, provided in an embodiment of the present invention. Detailed Implementation

[0019] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0020] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0021] Examples, such as Figure 1Considering the significant heat generated during vehicle operation, an efficient thermal management system can effectively regulate this heat, thereby improving overall vehicle performance and extending its driving range. The thermal management system for new energy vehicles (such as pure electric, hybrid, and fuel cell vehicles) needs to simultaneously optimize battery life (temperature uniformity), motor efficiency (cooling requirements), cabin comfort (PMV index), and overall vehicle energy consumption. Traditional methods face challenges such as: multi-physics coupling, strong coupling between battery heat generation, coolant flow, and cabin heat transfer, difficulty in global modeling, dynamic environmental adaptation, and complex and variable operating conditions due to extreme temperatures and driving mode switching. Traditional solutions also suffer from the separation of data-driven models from physical equations, unreliable long-term predictions, decision dimension explosion, and difficulty in converging multi-objective optimization in continuous state-action space. This invention proposes an intelligent control method for the thermal management system of new energy vehicles based on artificial intelligence. Traditional modeling methods for new energy vehicle thermal management systems mainly rely on physical equations or purely data-driven models, which have the following limitations:

[0022] (1) Multiphysics field separation modeling: Processes such as battery heat generation, coolant flow, and cabin heat transfer are often modeled independently, ignoring their coupling effects. For example, battery temperature rise will change coolant temperature, which in turn affects motor heat dissipation efficiency. However, traditional finite element simulation requires solving different fields separately, which is computationally expensive and difficult to update in real time.

[0023] (2) Static topology assumption: Traditional methods assume that the system topology is fixed (e.g., the cooling pipe layout remains unchanged), which cannot adapt to dynamic scenarios (e.g., passengers opening and closing car doors change the airflow path in the cabin).

[0024] Considering that graph convolutional neural networks (GCNNs) extract features from graph-structured data using graph convolution methods, they can automatically learn node features and the relationships between nodes, using information from other nodes in the graph to deduce the information of that node. Spatiotemporal graph neural networks define components such as batteries, motors, and air conditioners as graph nodes, with edge weights reflecting heat conduction / convection intensity, and dynamic adjacency matrices quantifying real-time coupling relationships. Rapid acceleration and fast charging cause sudden changes in heat load, requiring millisecond-level updates to the system state. Traditional CFD simulations take several minutes, failing to meet real-time requirements. Classical graph neural networks mainly consist of two phases: aggregation and propagation. In the aggregation phase, the graph neural network aggregates the information of each node's neighboring nodes to form a new node representation. Aggregation operations typically employ weighted aggregation or concatenation of the representations of each neighboring node. In the propagation phase, each node broadcasts its calculated and updated representation to its neighboring nodes to update their representations. The propagation phase can be iterated multiple times, allowing information to be propagated and aggregated throughout the graph. Graph Attention Network (GAT) is used to capture the structural feature space, and Gate Recurrent Unit (GRU) is used to capture temporal features. Aggregating GAT and GRU units can reduce inference latency.

[0025] Therefore, by constructing a dynamic spatiotemporal graph and multiphysics sensing, the data from multiple sensor cells, including battery cells, motor windings, and air conditioning vent thermal components, are defined as nodes in the spatiotemporal graph. The feature vectors of the N nodes at time t contain real-time data of temperature, flow rate, and current, respectively, resulting in the node matrix H at time t. (t) The adjacency matrix is ​​calculated based on thermal resistance, temperature difference, and mass flow rate of the connecting edges to characterize the thermal coupling strength. The formula for calculating the dynamic adjacency matrix is ​​as follows:

[0026]

[0027] in, Let ΔT be the adjacency matrix connecting edges i and j at time t. ij R represents the real-time temperature difference between nodes i and j. ij Let m be the thermal resistance between nodes i and j. ij Let σ() be the mass flow rate connecting edges i and j, σ be the sigmoid function, and w1 and w2 be the weight hyperparameters. A spatiotemporal graph convolution is constructed using a graph attention network and gated recurrent units, and the node matrix H at time t is... (t) The convolution of the adjacency matrix into the spatiotemporal graph outputs the spatiotemporal graph features h. (t) The mass flow rate of the connecting edge is the mass flow rate of the coolant. The spatiotemporal graph convolution uses spatial aggregation GAT to fuse neighbor node information, solving the local coupling in the heterogeneous graph structure. The GRU unit captures the node state evolution.

[0028] Traditional CFD has limitations, including high computational costs (high-precision 3D fluid simulations can take hours and cannot be deployed on vehicles), strong parameter dependence (mesh generation and turbulence model selection significantly impact results), and poor generalization. Purely data-driven approaches also have drawbacks, leading to unreliable long-term predictions, lack of physical consistency, and potential violations of mass conservation or energy equations. Therefore, we consider discretizing the Navier-Stokes equations into differentiable operators and embedding a neural network loss function to achieve "hard constraint" physical consistency.

[0029] Traditional control strategies have shortcomings. Rule-based control (such as PID control) relies on manual parameter tuning and cannot adapt to multi-objective conflicts (such as the competition between battery heating and cabin heating energy consumption at low temperatures). Single-layer reinforcement learning (such as DQN) struggles to converge in high-dimensional continuous action spaces. Passengers have significant differences in comfort preferences (e.g., the elderly prefer high temperatures, while children are more susceptible to cold), making it difficult for fixed strategies to meet their needs. Therefore, this design adopts a hierarchical decision-making architecture. In terms of technology selection, the upper-layer meta-policy network generates target weights, while the lower layer executes specific control, solving the dimensionality explosion problem. Innovative value: Supports dynamic weight adjustment (e.g., reducing weights when a passenger is detected wearing a down jacket). 舒适 (Reduce air conditioning energy consumption).

[0030] Physics-Informed Neural Network (PINN) is a method for solving partial differential equations. This algorithm combines deep learning and traditional numerical methods. By embedding physical constraints into the neural network, it makes full use of physical information as prior knowledge and trains the model with little or no labeled data to achieve the goal of accurately solving differential equations.

[0031] Therefore, a differentiable computational fluid dynamics (CFD)-PINN heat load prediction model is constructed. The Navier-Stokes equations are discretized into differentiable operators, which serve as hard constraints for the neural network. This model is then jointly trained with the physical neural network PINN, and the loss function is defined as follows:

[0032]

[0033] in, For each data item, N is the number of graph nodes, and T is the number of nodes in the graph. pred To predict temperature, T real For actual measured temperature, Let represent the residuals of the Navier-Stokes equations, u be the fluid velocity field, p be the pressure field, ρ be the fluid density, and v be the dynamic viscosity. The equation is a continuity equation, where λ and μ are physical constraint weighting coefficients, and the spatiotemporal graph features h are used. (t) Input a differentiable computational fluid dynamics (CFD) model for predicting heat load; output the heat load distribution Q for the next 30 seconds. predand predicting the fluid velocity field u pred .

[0034] The goal of meta-reinforcement learning training is to train a meta-learner from a set of learning tasks to quickly optimize a policy to solve a novel but relevant task. In essence, meta-reinforcement learning explores how to solve a series of tasks, rather than a single task as in traditional reinforcement learning.

[0035] Meta-reinforcement learning for multi-objective dynamic decision-making, extracting heat load distribution Q pred Maximum temperature max(T) pred ), mean flow velocity (u) pred and battery temperature difference ΔT bat The remaining battery charge (SOC) and driving mode are used as inputs to construct a meta-reinforcement learning state variable. A meta-policy network outputs dynamic target weights, which generate lower-level control commands to set the coolant flow rate. These lower-level control commands alter the actual thermal state of the new energy vehicle. New multi-source sensor data is fed back to S1 to update the spatiotemporal graph, forming a closed loop. The dynamic target weights output by the meta-policy network include vehicle energy consumption optimization weights w. 能耗 Battery temperature uniformity weight w 温差 and passenger cabin thermal comfort weight w 舒适 Satisfying w 能耗 +w 温差 +w 舒适 =1.

[0036] This invention constructs a dynamic spatiotemporal graph and multi-physics sensing. It defines multi-source sensor data, including battery cells, motor windings, and air conditioning vent thermal components, as nodes in the spatiotemporal graph. A dynamic adjacency matrix characterizes the thermal coupling strength, addressing the problem of fragmented multi-physics modeling in traditional methods. Spatial aggregation GAT and GRU units are used to construct spatiotemporal graph convolutional outputs spatiotemporal graph features. A differentiable computational fluid dynamics (CFD)-PINN thermal load prediction model is used, embedding the Navier-Stokes equations as hard constraints into the neural network to ensure that the prediction results strictly conform to thermodynamic laws. A meta-policy network generates target weights based on the state, achieving multi-objective adaptive optimization. Lower-level control commands change the actual thermal state of the new energy vehicle, feeding back to update the spatiotemporal graph and forming a closed-loop control.

[0037] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An intelligent control method for a new energy vehicle thermal management system based on artificial intelligence, characterized in that, Includes the following steps: S1. Construct a dynamic spatiotemporal graph and multi-physics sensing. Define the multi-source sensor data of battery cells, motor windings, and air conditioning vent thermal components as nodes in the spatiotemporal graph. The feature vectors of the N nodes at time t contain real-time data of temperature, flow rate, and current, respectively, to obtain the node matrix H at time t. (t) The adjacency matrix is ​​calculated based on thermal resistance, temperature difference, and mass flow rate of the connecting edges to characterize the thermal coupling strength. The formula for calculating the dynamic adjacency matrix is ​​as follows: in, Let ΔT be the adjacency matrix connecting edges i and j at time t. ij R represents the real-time temperature difference between nodes i and j. ij Let m be the thermal resistance between nodes i and j. ij Let σ() be the mass flow rate connecting edges i and j, σ be the sigmoid function, and w1 and w2 be the weight hyperparameters. A spatiotemporal graph convolution is constructed using a graph attention network and gated recurrent units, and the node matrix H at time t is... (t) The convolution of the adjacency matrix into the spatiotemporal graph outputs the spatiotemporal graph features h. (t) ; S2. The differentiable computational fluid dynamics (CFD)-PINN heat load prediction model discretizes the Navier-Stokes equations into differentiable operators, which serve as hard constraints for the neural network. It is then jointly trained with the physical neural network PINN, and the loss function is defined as follows: in, For each data item, N is the number of nodes, and T is the number of nodes. pred To predict temperature, T real For actual measured temperature, Let represent the residuals of the Navier-Stokes equations, u be the fluid velocity field, p be the pressure field, ρ be the fluid density, and v be the dynamic viscosity. The equation is a continuity equation, where λ and μ are physical constraint weighting coefficients, and the spatiotemporal graph features h are used. (t) Input a differentiable computational fluid dynamics (CFD) model for predicting heat load; output the heat load distribution Q for the next 30 seconds. pred and predicting the fluid velocity field u pred ; S3, Meta-reinforcement learning multi-objective dynamic decision-making, extracting heat load distribution Q pred Maximum temperature max(T) pred ), mean flow velocity (u) pred and battery temperature difference ΔT bat The remaining battery charge (SOC) and driving mode are used as inputs to construct meta-reinforcement learning state variables. The meta-policy network outputs dynamic target weights, which are then used to generate lower-level control commands to set the coolant flow rate. The lower-level control commands S4 and S3 change the actual thermal state of the new energy vehicle, and the new multi-source sensor data is fed back to S1 to update the spatiotemporal diagram, forming a closed-loop control.

2. The intelligent control method for a new energy vehicle thermal management system based on artificial intelligence according to claim 1, characterized in that, The mass flow rate of the connecting edge in step S1 is the mass flow rate of the coolant. The spatiotemporal graph convolution uses a spatial aggregation graph attention network to fuse neighbor node information, solves the local coupling in the heterogeneous graph structure, and the gated loop unit captures the node state evolution.

3. The intelligent control method for a new energy vehicle thermal management system based on artificial intelligence according to claim 1, characterized in that, The dynamic target weights output by the meta-policy network include the vehicle energy consumption optimization weights w. 能耗 Battery temperature uniformity weight w 温差 and passenger cabin thermal comfort weight w 舒适 , satisfy w 能耗 +w 温差 +w 舒适 =1.

Images