Automatic control method and device for electric vehicle thermal management system

Abstract

The application discloses an automatic control method and device of an electric vehicle thermal management system, and relates to the field of intelligent control. The method realizes multi-step prediction of future battery and motor heat production and heat load changes by fusing external information such as navigation paths and weather forecasts, plans a thermal management strategy in advance, and improves the foresight of system response. A multi-source disturbance dynamic structured correlation balancing technology is adopted to model the coupling relationship between each heat source, and cross disturbance is compensated through a joint state function to realize multi-heat loop collaborative control. Based on a model predictive control (MPC) framework, temperature tracking error and energy consumption penalty are optimized, the energy consumption of actuators such as compressors and water pumps is reduced while meeting the safe temperature interval, and the endurance is prolonged. The control accuracy and robustness under complex working conditions are enhanced by combining a state estimator with a fusion model and observation data, and finally, the global optimal thermal management considering component life, passenger comfort and energy efficiency is realized.

Description

Technical Field

[0001] This application relates to the field of intelligent control, and more specifically, to an automatic control method and apparatus for an electric vehicle thermal management system. Background Technology

[0002] As a crucial direction for future transportation development, the performance, driving range, and lifespan of electric vehicles largely depend on an efficient and reliable thermal management system. This system is responsible for precisely controlling the temperature of the battery pack, motor, power electronics, and passenger compartment. The battery pack needs to operate within a specific temperature range to ensure optimal charging and discharging performance and cycle life; the motor and power electronics require heat dissipation to prevent overheating damage and efficiency degradation; and the passenger compartment temperature directly affects passenger comfort. These heat loads and demands dynamically change with different vehicle operating conditions, environmental conditions, and user settings. Therefore, building a thermal management system that can intelligently respond and work efficiently is crucial.

[0003] Most existing thermal management control schemes for electric vehicles employ rule-based control, PID control, or simple state feedback control. These methods are typically reactive, adjusting based on current sensor signals such as temperature and pressure. For example, the cooling system is activated when the battery temperature rises to a certain threshold, or the air conditioning is turned on based on a set passenger compartment temperature. While these methods achieve basic thermal management functions, they lack the ability to predict future operating conditions and cannot anticipate upcoming changes in thermal load (such as motor heating due to hill climbing, increased battery heating due to prolonged high-speed driving, or a sudden increase in ambient temperature). This reactive control often lags behind actual needs, resulting in untimely system response when significant heat dissipation is required, or overworking after the thermal load decreases, leading to unnecessary energy consumption, shortened driving range, and difficulty in achieving globally optimal energy management while ensuring comfort and component health. Furthermore, existing methods struggle to effectively coordinate the control of multiple interconnected thermal loops and actuators, handle the coupling relationships between different heat sources (battery, motor, cabin), and achieve a dynamic balance between strict temperature constraints and energy consumption targets.

[0004] In view of the above problems, there is an urgent need for a more advanced and intelligent automatic control method for thermal management of electric vehicles. Summary of the Invention

[0005] This application is made in order to solve the above-mentioned technical problems.

[0006] According to one aspect of this application, an automatic control method for an electric vehicle thermal management system is provided, comprising:

[0007] Acquire real-time vehicle status data and external prediction information, wherein the external prediction information includes navigation path information and weather forecast information for the next N time steps;

[0008] Based on the real-time vehicle status data, construct the current system state vector;

[0009] The external prediction information is transformed into a predictable perturbation sequence;

[0010] Based on the current system state vector and the predictable perturbation sequence, construct an instance of the MPC optimization problem;

[0011] The MPC optimization problem instance is input into the online optimization solution module to obtain the current optimal control command;

[0012] The current optimal control command is sent to the drive actuator of the electric vehicle thermal management system to obtain physical drive signals and updated real-time vehicle status data.

[0013] According to another aspect of this application, an automatic control device for an electric vehicle thermal management system is provided, comprising:

[0014] The data acquisition module is used to acquire real-time vehicle status data and external prediction information, wherein the external prediction information includes navigation path information and weather forecast information for the next N time steps.

[0015] The current system state vector construction module is used to construct the current system state vector based on the real-time vehicle state data;

[0016] An external prediction information conversion module is used to convert the external prediction information into a predictable perturbation sequence;

[0017] The MPC optimization problem instance construction module is used to construct an MPC optimization problem instance based on the current system state vector and the predictable perturbation sequence.

[0018] The optimal control command generation module is used to input the MPC optimization problem instance into the online optimization solution module to obtain the current optimal control command;

[0019] The drive execution module is used to send the current optimal control command to the drive actuator of the electric vehicle thermal management system to obtain physical drive signals and updated real-time vehicle status data.

[0020] Compared with existing technologies, this application provides an automatic control method and device for an electric vehicle thermal management system. By integrating external information such as navigation paths and weather forecasts, it achieves multi-step prediction of future battery and motor heat generation and thermal load changes, enabling advance planning of thermal management strategies and improving system response foresight. It employs multi-source disturbance dynamic structured correlation equilibrium technology to model the coupling relationships between various heat sources and compensates for cross disturbances through joint state functions, achieving coordinated control of multiple thermal loops. Based on a model predictive control (MPC) framework, it optimizes temperature tracking error and energy consumption penalties, reducing the energy consumption of actuators such as compressors and water pumps while meeting safe temperature ranges, thus extending driving range. By combining a state estimator with fusion models and observational data, it enhances control accuracy and robustness under complex operating conditions, ultimately achieving globally optimal thermal management that balances component lifespan, passenger comfort, and energy efficiency. Attached Figure Description

[0021] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0022] Figure 1 This is a flowchart of an automatic control method for an electric vehicle thermal management system according to an embodiment of this application.

[0023] Figure 2 This is a flowchart of step S3 in the automatic control method of the electric vehicle thermal management system according to an embodiment of this application.

[0024] Figure 3 This is a flowchart of step S5 in the automatic control method of the electric vehicle thermal management system according to an embodiment of this application.

[0025] Figure 4 This is a block diagram of an automatic control device for an electric vehicle thermal management system according to an embodiment of this application. Detailed Implementation

[0026] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0027] It should be understood that the steps described in the method embodiments of this disclosure may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of this disclosure is not limited in this respect.

[0028] In view of the problems in the background art mentioned above, the present application proposes an automatic control method for an electric vehicle thermal management system. Figure 1 This is a flowchart of an automatic control method for an electric vehicle thermal management system according to an embodiment of this application. Figure 1 As shown, the automatic control method of the electric vehicle thermal management system according to an embodiment of this application includes: S1, acquiring real-time vehicle state data and external prediction information, wherein the external prediction information includes navigation path information and weather forecast information for the next N time steps; S2, constructing a current system state vector based on the real-time vehicle state data; S3, converting the external prediction information into a predictable disturbance sequence; S4, constructing an MPC optimization problem instance based on the current system state vector and the predictable disturbance sequence; S5, inputting the MPC optimization problem instance into an online optimization solution module to obtain the current optimal control command; S6, sending the current optimal control command to the drive actuator of the electric vehicle thermal management system to obtain a physical drive signal and updated real-time vehicle state data.

[0029] In step S1, real-time vehicle status data and external prediction information are acquired. The external prediction information includes navigation path information and weather forecast information for the next N time steps. Specifically, the navigation path information includes altitude, speed limits, and congestion predictions; the weather forecast information includes ambient temperature, ambient humidity, and solar radiation intensity. It should be understood that to achieve forward-looking and optimized control of the electric vehicle thermal management system, it is necessary not only to accurately grasp the current operating status but also to predict factors that will affect future heat load and heat dissipation conditions. Traditional reactive control strategies adjust only based on the current state, often exhibiting lag and failing to cope with upcoming drastic changes in operating conditions or the environment, resulting in low energy efficiency or poor temperature control. By introducing external prediction information, especially navigation paths and weather forecasts for the next period, the control system can anticipate future driving resistance, heat generation from the motor and battery, and the environmental impact on the cabin's heat load and the battery compartment's heat exchange. Therefore, real-time vehicle status data provides the starting point and feedback for control, while external predictive information provides feedforward disturbances for advanced control algorithms such as MPC, enabling the system to make smarter, more energy-efficient, and more predictive control decisions.

[0030] Specifically, in a specific example of this application, step S1 is implemented as follows: Acquiring real-time vehicle status data refers to collecting the operating parameters of each subsystem of the vehicle at the current moment through various sensors, actuator feedback, and network communication of the electronic control unit (ECU) integrated within the vehicle at the beginning of each control cycle. These parameters constitute the core elements of the current system state vector. This data is rich and diverse, including but not limited to: data provided by the battery management system (BMS) such as the total voltage, total current, individual cell temperature, overall temperature, state of charge (SOC), and state of health (SOH) of the battery pack; data provided by the motor controller (MCU) such as the temperature, speed, torque of the drive motor, and temperature information of power electronic devices (such as inverters); and feedback signals such as the operating status, flow rate, pressure, and temperature from the passenger compartment temperature and humidity sensor, the external ambient temperature sensor, and key components in the thermal management system, such as compressors, water pumps, fans, and valves. This real-time data is transmitted to the computing platform executing the thermal management control algorithm via a high-speed communication network within the vehicle, such as CAN FD or in-vehicle Ethernet. The control platform performs necessary filtering, calibration, and integration on these massive amounts of real-time data to form a multi-dimensional vector that can comprehensively reflect the current operating state of the system, namely the current system state vector.

[0031] Simultaneously, the external forecast information is acquired in parallel. This information provides the control system with environmental and operating condition predictions for a future period, enabling the controller to anticipate and proactively address potential changes in heat load. The external forecast information includes navigation path information and weather forecast information for the next N time steps. The navigation path information is provided by the vehicle's navigation system or a connected cloud service, and its acquisition is based on the user-defined destination or currently planned route. The navigation system not only provides real-time location and direction but also predicts road condition characteristics along the route. Specifically, the navigation path information includes altitude, speed limits, and congestion predictions for the next N time steps. Altitude data can be obtained through digital elevation maps to predict future gradient changes; speed limit information reflects the maximum permissible speed for different road sections; congestion predictions are based on historical traffic data and real-time road conditions to predict future traffic flow and speed reductions. This information is provided sequentially, i.e., predicting altitude, speed limits, and congestion status for each of the next N consecutive time steps. The weather forecast information is obtained from an external meteorological service platform through the vehicle's communication module, such as a cellular network or Wi-Fi, and is associated with the vehicle's current geographical location and the planned driving route. The weather forecast information specifically includes ambient temperature, ambient humidity, and solar radiation intensity for the next N time steps. This data service provides weather predictions for the vehicle's expected future location. Ambient temperature and humidity directly affect system heat dissipation and cabin heat load; solar radiation intensity primarily affects the cabin's solar heat gain. Similarly, this information is provided sequentially, predicting the ambient temperature, ambient humidity, and solar radiation intensity at each time point within the next N time steps.

[0032] Specifically, N in the aforementioned future N time steps is a key parameter pre-set as part of the overall MPC-based control system design. N represents the prediction time domain length of the MPC controller, i.e., how far ahead the controller looks when performing optimization calculations. The selection of N is a trade-off: a larger N provides longer-term prediction capabilities, enabling the controller to better plan long-term strategies, but significantly increases the computational load and complexity of model predictions; a smaller N has a lighter computational burden, but may lead to short-sighted control decisions, failing to effectively cope with long-term thermal load changes. Therefore, during the system development phase, the value of N is determined through offline simulation and performance evaluation, combined with the processing power of the onboard computing platform and the requirements for control performance, to find an optimal or acceptable fixed value. The time span of the external prediction information obtained in this step must cover the time range defined by the pre-set N time steps to meet the input requirements of the subsequent MPC optimization problem. For example, N can be set to 90, indicating that the system predicts information within the next 90 seconds. If each time step is set to 1 second, it contains 90 predicted data points for future times. These preset N values ​​and time steps are determined during system design and calibration, and are either used in a fixed manner during vehicle operation or dynamically adjusted according to specific conditions.

[0033] In step S2, the current system state vector is constructed based on the real-time vehicle state data. Correspondingly, since MPC is an advanced control strategy, its core is to predict the future behavior of the system in the prediction time domain based on the system's dynamic model, and obtain the optimal control input at the current moment through optimization calculation. To accurately predict future behavior, the control algorithm must know the precise state of the system at the current moment. Although real-time vehicle state data provides rich sensor information and other operating parameters, this raw data is scattered, diverse, and contains redundancy or noise. It cannot be directly used as variables to describe the internal physical state of the system, such as energy, mass, and temperature distribution, which are the key factors determining the future dynamic response of the system. Therefore, it is necessary to filter, process, and integrate this real-time acquired, multi-source sensor data and information, transforming it into a refined and standardized mathematical representation, namely the current system state vector.

[0034] Specifically, in a specific example of this application, step S2 is implemented as follows:

[0035] The real-time vehicle status data consists of raw operating parameters collected in real time from various subsystems, sensors, and actuators within the vehicle at each control cycle or sampling moment. This real-time data is transmitted to the computing platform responsible for executing the thermal management control algorithm via the vehicle's internal communication network, such as CAN bus or in-vehicle Ethernet.

[0036] After receiving this real-time data from different sampling frequencies and buses, a series of processes are required to construct the current system state vector. First, the received data is time-synchronized to ensure that all acquired data points reflect the system's state at the same moment as closely as possible. Then, the raw sensor data undergoes necessary preprocessing, including but not limited to: filtering to remove measurement noise, calibration to correct sensor bias, and validity checks to eliminate abnormal or erroneous data. For certain state variables, simple calculations or table lookups are needed, such as calculating the average or maximum temperature of the battery pack based on data from multiple temperature sensors.

[0037] Next, key state variables that represent the dynamic characteristics of the thermal management system are selected and extracted from the processed real-time vehicle status data. These state variables are selected based on the physical model and control requirements of the thermal management system; they represent the minimum set necessary to describe the system's current thermodynamic and electrochemical state. State variables include: key battery pack temperatures (such as average and maximum temperatures), drive motor temperatures, power electronics temperatures, passenger compartment interior air temperature, and even the battery's state of charge (SOC). Importantly, these selected variables should represent the system's dynamic state, meaning their future values ​​depend not only on current control inputs and external disturbances but also on their current intrinsic values.

[0038] Finally, these selected, processed, and synchronized key state variables are arranged in a predefined order and assembled into a fixed-dimensional vector, namely the current system state vector. For example, this vector is defined as [average battery temperature, motor temperature, inverter temperature, passenger compartment temperature, battery SOC]. Each component of the vector corresponds to the current value of a specific state variable. This constructed current system state vector concisely and comprehensively quantifies the operating state of the thermal management system at the current moment and serves as the starting point and key input for state estimation and model predictive control algorithms in subsequent steps, used to predict the future behavior of the system and solve for the optimal control command.

[0039] In step S3, the external prediction information is transformed into a predictable disturbance sequence. It should be understood that the core mechanism of MPC is to predict system behavior over a future period through a system dynamic model and make optimization decisions based on this prediction. The future behavior of the system depends not only on the current initial state and applied control inputs, but also on various uncontrollable factors from the external environment and driving conditions. For the thermal management system, these external factors manifest as known or predictable "disturbances" that change dynamically over time, such as changes in the heat generation power of different components and changes in the thermal load of the external environment on the cabin and battery. The external prediction information itself is not a direct value of heat or energy, but rather the fundamental cause or influencing factor leading to heat generation, transfer, or exchange. For example, climbing (altitude change) or acceleration / deceleration (speed limit, congestion impact) will cause changes in the load on the motor and battery, thus affecting their heat generation power; while the external ambient temperature, humidity, and solar radiation directly affect the heat exchange between the cabin and battery compartment and the outside world, as well as the thermal load of solar radiation on the cabin. In order for MPC to use this information to predict future thermodynamic state evolution in its prediction model, these external predictions reflecting future operating conditions and environment need to be transformed into predictable perturbation sequences that can be directly input into the system dynamic model and affect the system state.

[0040] Figure 2 This is a flowchart of step S3 in the automatic control method of the electric vehicle thermal management system according to an embodiment of this application. Specifically, in the embodiments of this application, as... Figure 2 As shown, step S3, which transforms the external prediction information into a predictable disturbance sequence, includes: S31, predicting the battery heat generation power and motor / electronic control heat generation power for the next P steps based on the navigation path information for the next N time steps to obtain the battery heat generation power sequence and the motor / electronic control heat generation power sequence; S32, predicting the environmental heat load on the cabin and the heat exchange volume on the battery compartment for the next P steps based on the weather forecast information for the next N time steps to obtain the heat load sequence and the heat exchange volume sequence; S33, performing multi-source disturbance dynamic structured correlation equilibrium on the battery heat generation power sequence, the motor / electronic control heat generation power sequence, the heat load sequence, and the heat exchange volume sequence respectively to obtain the battery heat generation power equilibrium sequence, the motor / electronic control heat generation power equilibrium sequence, the heat load equilibrium sequence, and the heat exchange volume equilibrium sequence; S34, integrating the battery heat generation power equilibrium sequence, the motor / electronic control heat generation power equilibrium sequence, the heat load equilibrium sequence, and the heat exchange volume equilibrium sequence to obtain the predictable disturbance sequence.

[0041] Specifically, in step S31, based on the navigation path information for the next N time steps, the battery heat generation power and motor / electronic control heat generation power for the next P steps are predicted to obtain battery heat generation power sequences and motor / electronic control heat generation power sequences. Correspondingly, the main heat sources that the thermal management system of an electric vehicle needs to manage are the battery and drive motor and their power electronic controller (ECU). The heat generation power of these components is not constant but highly dependent on the vehicle's current operating conditions, especially the demand for traction and braking power. For example, during high-speed driving, rapid acceleration, hill climbing, or energy recovery, the battery and motor / ECU generate a large amount of heat; while during low-speed constant speed, coasting, or parking, heat generation is relatively low. Traditional reactive thermal management systems determine the cooling intensity based solely on the current component temperature, making it difficult to cope with upcoming drastic changes in operating conditions. Therefore, by acquiring future navigation path information, including altitude, speed limits, and congestion predictions, the system can predict the driving cycles and power demands experienced by the vehicle over a future period and convert them into specific, quantified predicted sequences of battery and motor / ECU heat generation power, providing the MPC controller with early warnings about future heat load changes.

[0042] Specifically, in a concrete example of this application, the prediction process in step S31 is implemented using a machine learning model, and the specific implementation process is as follows: First, a prediction model needs to be established, which can learn the mapping relationship between the vehicle's future driving conditions (reflected by navigation path information) and the heat generation power of the battery and motor / electronic control system. To train this model, a large amount of historical data needs to be collected. This data includes the vehicle's driving records under various driving conditions, containing data corresponding to the future navigation path information (such as changes in altitude, speed, and whether it is in congestion), as well as the simultaneously recorded battery heat generation power and motor / electronic control system heat generation power. The collected data needs to cover different environmental conditions, driving styles, and operating conditions to ensure the model's generalization ability.

[0043] The collected historical data is preprocessed to form a dataset for training the machine learning model. For each segment of historical driving data, it is divided into an input sequence and an output sequence. The input sequence is navigation path information corresponding to the next N time steps (e.g., altitude sequence, speed sequence, congestion status sequence), and the output sequence is the battery heat generation power sequence and the motor / electronic control heat generation power sequence corresponding to the next P time steps. P is the time step at which the model's predicted heat generation power matches the prediction of subsequent MPC control. In practical applications, P is less than or equal to N, indicating that navigation information from the next N steps is used to predict the heat generation power of the most recent P steps. For example, if N is set to 90 and P is set to 60, the model will use navigation data from the next 90 seconds to predict the battery and motor / electronic control heat power for the next 60 seconds.

[0044] Choose a suitable machine learning model architecture for training. Specifically, the prediction model employs an encoder-decoder architecture. The encoder (which can use LSTM, GRU, or Transformer) processes the input sequence of navigation path information for the next N time steps, compressing it into a fixed-length context vector or a higher-level sequence representation. The decoder (also an LSTM, GRU, or Transformer) receives the encoder's output and progressively generates sequences of battery heat generation power and motor / electronic control heat generation power for the next P time steps. During the generation of each output step, the decoder can utilize an attention mechanism to focus on the most relevant parts of the encoder's output, thereby improving prediction accuracy. In other words, considering that both input and output are sequential data and the prediction task has temporal dependencies, sequence models such as RNN, LSTM, or GRU are suitable choices. More advanced model architectures such as Transformer can also be used. The model's input layer receives navigation path information for the next N time steps (e.g., altitude, speed limits, congestion predictions, etc., can be used as feature vectors input at each time step), and the output layer outputs battery heat generation power and motor / electronic control heat generation power for the next P time steps. The model's weights and biases are trained by optimizing algorithms (such as Adam) to minimize the error (e.g., mean squared error) between the predicted output and the actual measured data.

[0045] After model training, it is deployed to the vehicle's thermal management control unit for online prediction. During vehicle operation, the system acquires navigation path information sequences for the next N time steps. These sequences serve as input to the trained machine learning model. Upon receiving these inputs, the model performs internal calculations to predict the battery heat generation power sequences and motor / electronic control heat generation power sequences for the next P time steps. For example, if the input is a prediction sequence of altitude, speed limit, and congestion for the next 90 seconds, the model will output a battery heat generation power prediction sequence (one containing 60 power values) and a motor / electronic control heat generation power prediction sequence (another containing 60 power values) for the next 60 seconds. These prediction sequences are the outputs of this step, providing crucial information for subsequent heat load prediction and balancing. The preset values ​​N and P are determined during the system design phase based on factors such as the availability of prediction information, prediction accuracy requirements, and system computing resources, and are used consistently during system operation.

[0046] Specifically, P represents the prediction time domain length of the Model Predictive Controller (MPC). Unlike N (navigation / weather prediction time domain), P is a core parameter determined during the design of the MPC controller, determining the length of time the controller can predict system dynamics in advance. The choice of P depends on the thermodynamic characteristics of the controlled object (electric vehicle thermal management system), the system's response time, and the computing power of the onboard computing platform. A larger P can better capture and utilize the system's thermal inertia for optimization, such as activating the cooling system in advance to cope with upcoming high heat loads, thereby reducing peak temperature and energy consumption, but also increasing the computational burden. P is less than or equal to N because the navigation path information used to predict heat generation only has N time steps. The preset values ​​N and P are determined during the system design phase based on factors such as the availability of prediction information, prediction accuracy requirements, and system computing resources, and are fixed during system operation.

[0047] Specifically, in step S32, based on the weather forecast information for the next N time steps, the thermal load on the cabin and the heat exchange with the battery compartment in the next P steps are predicted to obtain the thermal load sequence and the heat exchange sequence. It should be understood that the thermal management system of an electric vehicle not only needs to cope with the thermal load of its own heat-generating components (battery, motor, electronic control), but also needs to manage the heat input or output from the external environment, especially the environmental impacts on the passenger compartment and battery compartment. These environmental factors, such as ambient temperature, humidity, and solar radiation intensity, will significantly change the cooling / heating requirements of the cabin and the heat exchange rate between the battery compartment and the external environment. Traditional control methods often only passively adjust the air conditioning and cooling system based on current environmental conditions or cabin interior temperature. This response mode cannot anticipate the thermal load fluctuations caused by future environmental changes, resulting in insufficient system response when environmental conditions are about to deteriorate, affecting passenger comfort or component temperature; conversely, when environmental conditions are about to improve, the system overworks, causing energy waste. Therefore, by utilizing the weather forecast information acquired over N future time steps, the system can anticipate the changing trends of the external environment in the future and transform these future environmental impacts based on weather forecasts into specific and quantified cabin heat load sequences and battery compartment heat exchange sequences, providing the MPC controller with key feedforward information about future environmental disturbances.

[0048] Specifically, in a specific example of this application, the specific implementation of step S32, this prediction process, can also be implemented using a machine learning model, particularly an encoder-decoder architecture suitable for processing sequence inputs and outputs:

[0049] First, a predictive model needs to be established to learn the complex mapping relationship between future weather information and environmentally induced cabin heat load and battery compartment heat exchange. Cabin heat load is mainly affected by factors such as ambient temperature, humidity, solar radiation, window heat transfer, and occupant heat generation, with ambient weather being the primary external disturbance. The heat exchange between the battery compartment and the environment is influenced by ambient temperature, battery compartment structure, cooling duct design, and any existing external cooling or heating devices. To train this model, a historical dataset containing weather forecast information for N future time steps (e.g., ambient temperature, ambient humidity, and solar radiation intensity sequences) is needed as input, along with calculated or measured cabin heat load and battery compartment heat exchange as output.

[0050] After data collection, preprocessing is performed to form the input-output sequence pairs needed to train the machine learning model. The input sequence is weather forecast information for the next N time steps, and the output sequence is the cabin heat load sequence and battery compartment heat exchange rate sequence for the next P time steps. P is the time step predicted by the model, consistent with the P of the thermal power prediction model and subsequent MPC predictions, and less than or equal to N. For example, if N is set to 90 and P is set to 60, the model uses weather information for the next 90 seconds to predict the cabin heat load and battery compartment heat exchange rate for the next 60 seconds.

[0051] This prediction model employs an encoder-decoder architecture. The encoder processes the input weather forecast information sequence for the next N time steps, encoding it into a fixed-length vector or a higher-level sequence representation. The encoder can use a multi-layer LSTM or GRU network, or a Transformer encoder. Taking a GRU encoder as an example, it receives feature vectors such as ambient temperature, ambient humidity, and solar radiation intensity at each time step, and updates the hidden state internally. Ultimately, the hidden state of the encoder at the last time step, or the context vector formed by combining the hidden states of multiple time steps, captures the information of the entire input weather sequence.

[0052] The decoder receives the encoder's output and is responsible for progressively generating the cabin heat load sequence and battery compartment heat exchange volume sequence for the next P time steps. The decoder is also an LSTM or GRU network, or a Transformer decoder. Taking an LSTM decoder as an example, it initially receives the encoder's output context vector as its initial hidden state. Then, at each time step, the decoder uses the output from the previous time step (or forces in the previous time step's true value during training) and the current hidden state as input, predicts the cabin heat load and battery compartment heat exchange volume for the current time step through a linear layer, and updates its own hidden state for prediction in the next time step. To improve prediction accuracy, an attention mechanism can be introduced into the decoder, allowing it to focus on the most relevant parts of the encoder's input sequence when generating each output.

[0053] The entire encoder-decoder model is trained using an optimization algorithm (such as Adam) by minimizing the error (e.g., mean squared error) between the predicted output and the actual heat load and heat exchange. After training, the model is deployed to the vehicle's thermal management control unit. During vehicle operation, a weather forecast sequence for the next N time steps is acquired and input into the trained encoder-decoder model. The encoder processes the input sequence, and the decoder generates a cabin heat load prediction sequence and a battery compartment heat exchange prediction sequence for the next P time steps based on the encoder's output. For example, given a 90-second sequence of ambient temperature, humidity, and solar radiation, the model outputs a 60-second cabin heat load prediction sequence (one containing 60 heat load values) and a battery compartment heat exchange prediction sequence (another containing 60 heat exchange values). Specifically, the principles for determining the preset values ​​N and P are the same as for the thermal power prediction model, requiring a trade-off between prediction range and accuracy, as well as computational resources.

[0054] Specifically, in step S33, the battery heat generation power sequence, the motor / electronic control heat generation power sequence, the heat load sequence, and the heat exchange quantity sequence are subjected to multi-source disturbance dynamic structured correlation equilibrium to obtain the battery heat generation power equilibrium sequence, the motor / electronic control heat generation power equilibrium sequence, the heat load equilibrium sequence, and the heat exchange quantity equilibrium sequence, respectively. Correspondingly, when integrating the battery heat generation power sequence, the motor / electronic control heat generation power sequence, the heat load sequence, and the heat exchange quantity sequence, since the obtained predictable disturbance is essentially a multi-source disturbance, representing a complex nonlinear correlation of the thermal management state based on each disturbance source, the integration of these sources can easily lead to an imbalance. Therefore, if the battery heat generation power sequence, the motor / electronic control heat generation power sequence, the heat load sequence, and the heat exchange quantity sequence can be mapped to an correlation space with a stable state structure first, then a balanced expression of the predictable disturbance sequence relative to the aforementioned variables can be achieved.

[0055] Specifically, in the embodiments of this application, the battery heat generation power sequence, the motor / electronic control heat generation power sequence, the heat load sequence, and the heat exchange quantity sequence are respectively subjected to multi-source perturbation dynamic structured correlation equilibrium to obtain the battery heat generation power equilibrium sequence, the motor / electronic control heat generation power equilibrium sequence, the heat load equilibrium sequence, and the heat exchange quantity equilibrium sequence, including:

[0056] The battery heat generation power sequence, the motor / electronic control heat generation power sequence, the heat load sequence, and the heat exchange quantity sequence are each represented as row vectors to obtain a set of state vectors, i.e. , , and The process of calculating the mean vector among the set of state vectors to obtain the state mean vector is expressed as follows: ;in, Add vectors by their positions. It is the state mean vector;

[0057] Based on the state mean vector, the covariance matrix of each state vector in the set of state vectors is calculated to obtain the set of state feature covariance matrices. This process is expressed as follows: ;in, It is the th in the set of state vectors A state vector, It is a vector subtraction by position. It is matrix multiplication. It is a transpose operation. It is the first in the set of state feature covariance matrices The covariance matrix of each state feature;

[0058] Sensitive correlation is performed on each set of state vectors and state feature covariance matrices in the set of state vectors and the set of state feature covariance matrices to obtain a set of joint state function values. This process is expressed as follows: ;in, It is the value of an exponential function with the natural constant e as its base. It is the first in the set of joint state function values A joint state function value to describe each state vector The associated coupling relationship between the joint state and the state;

[0059] Based on each joint state function value in the set of joint state function values, a stable state prediction value is constructed for each corresponding state vector to obtain a set of stable state prediction values. This process is represented as follows: ;in, For L2 norm calculation, It is the th in the set of steady-state predictions A stable state prediction value, that is, to obtain the stable state based on the non-degenerate norm representation by constructing an embedding mapping from the state vector to the joint state function;

[0060] Using each stable state prediction value in the set of stable state prediction values ​​as a compensation coefficient, a state equilibrium correlation is performed on each state vector in the set of state vectors to obtain a set of state equilibrium vectors; wherein, each state equilibrium vector in the set of state equilibrium vectors is the battery heat generation power equilibrium sequence, the motor / electronic control heat generation power equilibrium sequence, the heat load equilibrium sequence, and the heat exchange quantity equilibrium sequence, and this process is expressed as: ;in, For observation coefficients, It is the th in the set of state equilibrium vectors Each state equilibrium vector is a dynamic structured constraint compensation superimposed on the predicted stable state value as a state equilibrium vector.

[0061] In particular, the observed coefficient Its value is dynamically adjusted through experimental calibration or online optimization algorithms, depending on the system's deviation from the state vector (e.g. The sensitivity requirements and noise suppression needs can be determined by fitting historical data or calculating the Kalman filter gain.

[0062] In this way, state associations can be quantified by calculating joint state function values, and non-equilibrium association states caused by joint state degradation can be eliminated based on the calculation of stable state prediction values. Ultimately, each state is dynamically constrained to a structured association space, realizing the balanced expression of the predictable perturbation sequence relative to the above variables.

[0063] Specifically, in step S34, the battery heat generation power balancing sequence, the motor / electronic control heat generation power balancing sequence, the heat load balancing sequence, and the heat exchange balancing sequence are integrated to obtain the predictable disturbance sequence. It should be understood that although the aforementioned steps have separately predicted and processed and balanced predictable thermodynamic disturbances (battery heat generation, motor / electronic control heat generation, cabin heat load, and environmental heat exchange) from different sources, these disturbances ultimately affect the system, collectively influencing the temperature state of key components such as the battery, motor, and cabin. The core of the MPC algorithm is to predict future state evolution based on a unified system dynamic model, which treats all external predictable influences as one or a set of structured disturbance inputs. Therefore, these disturbance sequences from different sources, but after balancing and alignment to the same prediction time domain, with P time steps, need to be logically integrated according to their function in the system model to form a single predictable disturbance sequence containing all predictable thermodynamic input information.

[0064] Specifically, in a specific example of this application, step S34 is implemented as follows:

[0065] For each time step k (from 1 to P) in the future P time steps, extract the k-th value of the battery heat generation power balance sequence, the k-th value of the motor / electronic control heat generation power balance sequence, the k-th value of the heat load balance sequence, and the k-th value of the heat exchange quantity balance sequence. Then, arrange these four values ​​in a predefined order to form a perturbation vector at time step k. For example, this perturbation vector can be defined as [battery heat generation power prediction value_k, motor / electronic control heat generation power prediction value_k, cabin heat load prediction value_k, battery compartment heat exchange quantity prediction value_k]. Repeat the above combination process to construct the corresponding perturbation vectors for the future P time steps (from time step 1 to time step P). In this way, a sequence containing P perturbation vectors is obtained. This sequence, which is formed by concatenating the P perturbation vectors in chronological order, is the predictable perturbation sequence.

[0066] In step S4, an MPC optimization problem instance is constructed based on the current system state vector and the predictable disturbance sequence. Accordingly, the working mechanism of MPC requires solving an optimization problem online in each control cycle, considering the current system state and foreseeable external influences, to determine the optimal control input at the current moment. The current system state vector accurately describes the physical state of the system at the current moment and is the starting point for predicting future states. The predictable disturbance sequence provides information on external disturbances the system will face in the future prediction time domain, including known or predicted influences from driving conditions (battery, motor heat generation) and the environment (cabin heat load, battery compartment heat exchange). To construct a mathematical optimization problem that can guide optimal control decisions, these two need to be integrated.

[0067] Specifically, in this embodiment of the application, step S4, constructing an MPC optimization problem instance based on the current system state vector and the predictable perturbation sequence, includes: S41, performing state estimation optimization on the current system state vector based on the state estimator to obtain an optimized current system state vector; S42, constructing the MPC optimization problem instance based on the optimized current system state vector and the predictable perturbation sequence.

[0068] More specifically, in a specific embodiment of this application, step S41, optimizing the current system state vector based on the state estimator to obtain an optimized current system state vector, includes: S411, inputting the state estimate from the previous moment and the control input from the previous moment into the system dynamic model to predict the current state to obtain a predicted current system state vector and the uncertainty of the predicted current system state vector; S412, inputting the predicted current system state vector, the uncertainty of the predicted current system state vector, and the current system state vector into the observation model to obtain the optimized current system state vector. The state estimator described here is an intelligent algorithm module whose core function is to integrate information from different sources, including the system's mathematical model prediction, the system state estimate from the previous moment, and the actual sensor measurement data at the current moment, to provide the best estimate of the actual internal state of the system. Specifically, the state estimator includes input system dynamic model and observation model.

[0069] Accordingly, the system dynamic model embodies the physical laws governing the evolution of the system state over time. This prediction process, based on the system's intrinsic mechanisms, provides a smooth, physically consistent estimate of the current state, avoiding the noise effects of directly using current raw measurements. However, since the model is not perfect and contains process noise and modeling errors, this model-based prediction carries uncertainty. Quantifying this uncertainty (e.g., through the covariance matrix) is crucial because it reflects the reliability of the model's prediction and provides weighted information for subsequent corrections based on actual measurements. Therefore, the state estimate and control input from the previous time step are input into the system dynamic model to predict the current state, resulting in the predicted current system state vector and the uncertainty of this predicted vector.

[0070] Specifically, in a specific example of this application, the specific implementation process of step S411 is as follows: Here, the state estimate of the previous moment refers to the optimal estimate of the internal state variables of the system obtained by the state estimator after comprehensive data fusion (e.g., combining the predicted value of the previous moment with the measured value of the current moment for filtering and optimization) at the end of the immediately preceding control cycle. It is a vector containing the most reliable numerical representation of the key physical quantities in the system (such as the temperature of each area of ​​the battery pack, the temperature of the motor and electronic control, the temperature and flow rate of the coolant, the cabin temperature, etc.) at the previous moment. This vector is the output of the state estimation process in the previous step (or the previous control cycle). The control input of the previous moment refers to the control quantity or the actual response of these actuators that the controller of the thermal management system actually applied to each actuator in the immediately preceding control cycle. For example, the actual speed or power of the coolant pump, the actual speed of the fan, the actual torque or power of the electric compressor, the heating power of the heater, etc. These values ​​are the actions actually performed by the system in the previous moment to achieve the control objective, and they directly affect the change of the system state in the current control cycle. This control input is also a vector, consisting of the input values ​​of all controlled actuators in the system.

[0071] The state estimate and control input from the previous time step are fed into the system dynamic model. This system dynamic model is the foundation for state estimation and MPC prediction, capturing the physical laws or empirical relationships of the evolution of the internal state of the thermal management system over time. For physics-based modeling methods, the system dynamic model consists of a set of mathematical equations, such as differential or difference equations describing energy conservation, heat transfer, fluid dynamics, and the conversion of electrical energy into heat. The parameters in these equations (such as the heat capacity, thermal conductivity, convective heat transfer coefficient of materials, the flow characteristic curve of pumps, the efficiency curve of fans, and the relationship between resistive heating power and current) constitute the core content of the model, characterizing the physical properties of each component in the system and their interactions. For data-driven or hybrid modeling methods, the model consists of state transition functions or parameterized relationships learned from a large amount of historical data. These mathematically describe, in form (not necessarily explicit physical equations), how inputs (previous state, control input, disturbances) lead to state changes.

[0072] Specifically, the state estimate from the previous moment is taken as the initial state of the system, and the control input from the previous moment is taken as the external excitation applied within the time interval. Then, the system dynamic model is used to calculate how the system state will change after one control cycle (i.e., the time interval from the previous moment to the current moment). If the system dynamic model is in the form of differential equations, this is solved using numerical integration methods (such as the Euler method, Runge-Kutta method, etc.); if the model is in the form of difference equations, a one-step recursive calculation is performed directly. This calculation process simulates physical processes such as how heat flows inside the system and how temperature responds to control actions.

[0073] Besides predicting the state vector itself, this step also requires obtaining the uncertainty in predicting the current system state vector. This is because the system dynamic model may not be perfectly accurate (model error), and the system itself may contain unmodeled random disturbances (process noise). In probabilistic state estimation algorithms (such as Kalman filters and their variants), the system dynamic model includes not only the state transition equations but also the statistical properties describing the process noise (covariance matrix). During prediction, not only is the mean of the state predicted (i.e., the current system state vector), but also, based on the linearized approximation of the state transition equations (or methods such as unscented transformations to handle nonlinearity) and the known process noise covariance, how the current state estimation covariance matrix propagates over time is predicted, thus obtaining the uncertainty in predicting the current system state vector, represented as a covariance matrix. The diagonal elements of this covariance matrix represent the variance of the predicted values ​​of each state variable, and the off-diagonal elements represent the covariance between them. This part of the system dynamic model (the process noise covariance matrix and its propagation method) is also an important component of the model. The model parameters (such as heat transfer coefficient, heat capacity, etc.) and process noise covariance matrix are set through system identification experiments or reference component specifications and physical principles.

[0074] In a specific example: the state estimation vector at the previous moment contains an estimated battery temperature of 42°C and an estimated coolant temperature of 38°C. The control input vector at the previous moment contains a cooling pump power command of 50%. The system dynamic model consists of a set of difference equations describing battery heat conduction and convective heat transfer. Its parameters (such as the battery equivalent heat capacity, thermal resistance from the battery to the cooling plate, and the relationship between the convective heat transfer coefficient from the cooling plate to the coolant and the coolant flow rate (determined by the pump power)) are determined based on battery pack design and experimental data. The model also includes a process noise covariance describing the uncertainties in internal battery heat generation and other unmodeled disturbances.

[0075] The model receives the previous time step's estimated battery temperature (42°C) and coolant temperature (38°C), along with the previous coolant pump power command of 50%. Internally, the model's difference equations, based on these inputs and utilizing their included thermophysical parameters, calculate how the battery and coolant temperatures will change within the current control cycle (e.g., 100 milliseconds), considering processes such as battery self-heating (potentially related to battery current, internal resistance, etc., where current may also be an input or predicted disturbance), heat transfer from the battery to the cooling plate, and heat transfer from the cooling plate to the coolant. This calculation predicts the current system state vector; for example, a predicted battery temperature of 42.1°C and a coolant temperature of 38.05°C. Simultaneously, the model, based on its description of process noise and considering the uncertainty of the previous state estimate (if available), calculates and outputs the uncertainty of the predicted current system state vector; for example, a predicted variance of 0.02 for the battery temperature and 0.01 for the coolant temperature. These values ​​represent the possible fluctuation range of the predicted values. These predicted values ​​and their uncertainties will be used in subsequent state update steps.

[0076] It is understandable that predicting the current system state vector is based on the forward extrapolation of the system's dynamic model, reflecting the physical laws governing the evolution of the system state over time. However, it does not utilize real-time sensor data at the current moment, and therefore cannot correct model errors or address disturbances not modeled. The current system state vector, on the other hand, originates directly from sensor measurements and contains the system's true information at the current moment, although it may contain noise or be incomplete. Therefore, the predicted current system state vector, its uncertainty, and the current system state vector are input into the observation model.

[0077] Specifically, in a specific example of this application, the specific implementation process of step S412 is as follows:

[0078] The observation model described can be understood as a machine learning model specifically trained for data fusion. It learns how to calculate the best estimate of the system state at the current moment based on the model's predicted state, the uncertainty of the prediction, and the actual sensor measurements. This model can be a feedforward neural network, also known as a multilayer perceptron. Its architecture includes an input layer, several hidden layers, and an output layer.

[0079] The model's input layer receives a combination of three pieces of information: first, the predicted current system state vector, which is the dynamic model's inference of the current state based on information from the previous time step; second, the uncertainty of the predicted current system state vector, which quantifies the reliability of the dynamic model's prediction; and third, the current system state vector itself, which is the raw sensor measurement value acquired by the system at the current time step, containing true information but also accompanied by noise. These three vectors are concatenated or otherwise combined into a single input vector and fed into the input layer of the feedforward neural network. The hidden layer processes the input signal through a nonlinear activation function (such as ReLU), learning the complex relationship between the predicted value, uncertainty, and measurement value, and how they collectively influence the final state estimate. The output layer then produces a vector, namely the optimized current system state vector. This output vector is the optimal estimate of all key state variables of the system at the current time step; it integrates predicted and measured information, filters out noise, and may include inferences about states that cannot be directly measured.

[0080] This machine learning observation model, based on a feedforward neural network architecture, requires offline training to determine its internal parameters (weights and biases). The training data includes a large amount of historical data: the predicted current system state vector at the same time point, the uncertainty of the predicted current system state vector (derived from the output of a trained dynamic model), the current system state vector (raw sensor data), and the real current system state vector used for supervised learning (obtained through high-precision simulation or experimental calibration). The model is trained by minimizing the error (e.g., mean squared error) between its optimized current system state vector and the corresponding real current system state vector. During training, the model learns how to intelligently weight and combine predicted and measured values ​​based on the confidence level of the prediction (the magnitude of uncertainty) and the characteristics of the measurements (the noise distribution in the training data) to obtain the optimal estimate.

[0081] More specifically, in one embodiment of this application, step S42, constructing the MPC optimization problem instance based on the optimized current system state vector and the predictable disturbance sequence, includes: S421, using the optimized current system state vector and the predictable disturbance sequence to predict the system state evolution of the electric vehicle thermal management system if a series of control inputs are applied in the next P steps to obtain a predicted state trajectory and a predicted control trajectory; S422, defining a cost function for the predicted state trajectory and the predicted control trajectory, the cost function including temperature tracking error, energy consumption and control quantity change penalties and weighting coefficients; S423, applying constraints to the cost function, the constraints including state constraints and input constraints to obtain the MPC optimization problem instance. Accordingly, MPC is a control strategy based on online optimization. Its working mechanism requires constructing and solving a mathematical optimization problem for the current system state and expected future external influences in each control cycle to obtain the optimal control command at the current moment. Therefore, in this application, based on the optimized current system state vector and the predictable disturbance sequence, an instance of the MPC optimization problem is constructed. This step is the core link in integrating thermal management requirements, system capabilities, and future prediction information into a solvable mathematical problem.

[0082] It's understandable that MPC's predictive ability stems from its ability to anticipate system state changes over a future period. To achieve this, MPC requires a mathematical model describing the system's dynamic behavior, upon which simulations are performed. The optimized current system state vector serves as the starting point for the simulation, as it represents the most reliable estimate of the system's current state. Simultaneously, known future external influences (represented by predictable perturbation sequences) are incorporated into the model input to ensure the predictions closely reflect reality.

[0083] Specifically, in a specific example of this application, the specific implementation process of step S421 is as follows:

[0084] First, the optimized current system state vector is the most accurate estimate of the system's current internal state output by the state estimator at the current time step (called time t). It is a vector containing key state information that has been fused and filtered, such as the currently optimal estimated temperatures of various battery components, motor temperature, cabin temperature, and coolant status. This vector will serve as the starting point for future predictions.

[0085] Secondly, a predictable disturbance sequence refers to the prediction of the system's external environment and operating conditions over P future time steps (from time t to t+P-1). This can be a vector sequence, with each vector representing a predicted disturbance at a future moment.

[0086] The task is to train a machine learning system model. This model learns how the system state evolves over time given the current state, control inputs, and external disturbances. This model can be a sequence-to-sequence neural network or a recurrent neural network (RNN) capable of multi-step prediction, such as a Long Short-Term Memory (LSTM) network or a gated recurrent unit (GRU).

[0087] The model's input structure can be designed such that at each prediction step (from step size k=0 to P-1), it receives the predicted state from the previous step, the control input for the current step, and the predicted perturbation for the current step. Specifically, for the first prediction step (k=0), the model receives the optimized current system state vector, the first element of the future control input sequence, and the first element of the predictable perturbation sequence. The model predicts the state at time t+1. Then, for the second prediction step (k=1), the model receives the predicted state, the second element of the future control input sequence, and the second element of the predicted perturbation sequence. The model predicts the state at time t+2, and so on, performing P-step predictions.

[0088] The internal architecture of this machine learning system model can be a multi-layered LSTM or GRU unit. These recurrent units can capture the dynamic relationships and nonlinear behavior within the system, predicting future state changes by learning the relationships between states, controls, and perturbations in historical data. At each prediction step, the recurrent unit processes the input of the current step (the predicted state of the previous step, the current control, and the current perturbation), updates its internal state, and then outputs a predicted value of the system state at the end of the current step through a fully connected layer.

[0089] Training this machine learning system model requires a large amount of historical data, including records of the system's state evolution under different operating conditions and control inputs, as well as corresponding disturbance data. The model learns to predict the true state at the next moment by inputting the true state at a given moment, the control input at that moment, and the disturbance. This is extended to multi-step prediction, optimizing the model parameters by minimizing the error between the entire predicted trajectory and the true trajectory.

[0090] In the MPC prediction process, this is not a single prediction, but rather a repeated P-step prediction process for a series of alternative future control input sequences. The MPC optimizer generates multiple sets of different sets of control input sequences (i.e., different predicted control trajectories). For each set of control sequences, it is input into the machine learning system model along with the optimized current system state vector and the predictable disturbance sequence. The model then outputs the corresponding predicted state trajectory.

[0091] Here's a concrete example: The optimized current system state vector represents a maximum battery temperature of 45°C, a motor temperature of 70°C, and a cabin temperature of 25°C. Predict the future P = 5 time steps (each step 10 seconds). The predictable disturbance sequence is the ambient temperature prediction for the next 5 10-second intervals, for example, [30°C, 31°C, 32°C, 31°C, 30°C]. The MPC optimizer generates a future control input sequence (i.e., a predicted control trajectory), for example, a coolant pump power command sequence [80%, 80%, 70%, 60%, 50%] and a fan speed command sequence [70%, 70%, 60%, 50%, 40%]. The initial state [45, 70, 25], the disturbance sequence [30, 31, 32, 31, 30], and this set of control sequences [80, 80, 70, 60, 50] and [70, 70, 60, 50, 40] are input into the trained machine learning system model. The model first receives [45,70,25] (current state) and [80,70] (the first step of the control and disturbance combination input) to predict the state at the next time step, for example, [45.2,70.5,25.3]. Then, using this predicted state [45.2,70.5,25.3] and the next set of control [80,70] and disturbance

[31] , it predicts the state at the next time step after that, repeating this process for 5 steps. Finally, a predicted state trajectory is obtained, for example, the battery temperature sequence is [45.2,45.3,45.2,45.1,45.0]. The MPC optimizer evaluates multiple such predicted state trajectories (generated by different control sequences), selects the control sequence that best satisfies the control objective, and uses its first element as the actual control command applied in the current cycle.

[0092] Accordingly, optimization requires a clear cost function to measure the merits of different control strategies. The goal of MPC is to find a control input sequence that optimizes overall performance over the future prediction time domain. To achieve multi-objective optimization of a thermal management system, multiple performance indicators, such as temperature control accuracy (e.g., bringing battery and cabin temperatures as close to target values ​​as possible), system energy consumption, and the smoothness of control actions, need to be quantified and integrated into a single numerical indicator—this is the cost function. By using the predicted state trajectory and the predicted control trajectory as inputs, the cost function can calculate the total cost incurred if the system operates according to these trajectories.

[0093] Specifically, in a specific example of this application, the specific implementation process of step S422 is as follows:

[0094] The cost function is constructed as the sum of instantaneous costs over P future time steps, multiplied by corresponding weighting coefficients. First, there is the temperature tracking error term: for each key temperature variable in the predicted state trajectory, the difference between its value and the set target temperature is calculated at each moment in the P future time steps. These differences are squared to penalize larger deviations, and then summed over the P future time steps. Second, there is the energy consumption term: based on the predicted control trajectory, the total energy consumed by the system actuators (such as compressors, pumps, and fans) over the P future time steps is calculated. This is done by mapping the control input at each time step to the actuator power and then summing them. Finally, there is the control quantity change penalty term: the difference in control input between adjacent time steps in the predicted control trajectory is calculated, squared, and then summed over all control inputs and all future time steps to smooth the control action. The total temperature tracking error, total energy consumption, and total control quantity change penalty obtained from the above summation are multiplied by the preset weighting coefficients. These weighting coefficients are positive values ​​used to balance the importance of various performance indicators. The final cost function value is the sum of these three weighted terms. Specifically, during the system design and calibration phase, engineers set initial weights based on the specific control objectives and performance requirements of the electric vehicle thermal management system. Then, they conduct extensive simulations by running the system model under various typical driving cycles and environmental conditions. By observing the system's temperature control effectiveness, energy consumption, and the smoothness of control actions under different weight combinations, the weight coefficients are iteratively adjusted until the desired performance balance is achieved.

[0095] Specifically, in a specific example of this application, step S423 is implemented as follows: A series of mathematical constraints are defined based on the physical characteristics of the thermal management system, component safety specifications, and actuator specifications. These constraints are applied to each time step of the predicted state trajectory and predicted control trajectory generated in the first two sub-steps. For example, state constraints include: for each time step k in the prediction time domain, the predicted temperature of a critical component must be greater than or equal to its lower safety limit and less than or equal to its upper safety limit. Input constraints include: for each time step k in the prediction time domain, the predicted control input value of an actuator must be greater than or equal to its minimum output capability and less than or equal to its maximum output capability. Using the cost function as the objective function of the optimization problem, and simultaneously using the above state constraints and input constraints as limitations of the optimization problem, the MPC optimization problem instance is formed.

[0096] In step S5, the MPC optimization problem instance is input into the online optimization solution module to obtain the current optimal control command. Figure 3 This is a flowchart of step S5 in the automatic control method of the electric vehicle thermal management system according to an embodiment of this application. Specifically, in the embodiments of this application, as... Figure 3As shown, step S5, inputting the MPC optimization problem instance into the online optimization solution module to obtain the current optimal control command, includes: S51, using a numerical optimization algorithm to solve the MPC optimization problem instance to obtain the optimal control input sequence; S52, extracting the control input at the current moment from the optimal control input sequence as the current optimal control command. The online optimization solution module is a specialized algorithm or software library used to receive the definition of the above MPC optimization problem instance, solve the problem under real-time requirements, and find a series of future control input sequences that minimize (or maximize) the cost function while satisfying all constraints.

[0097] As can be understood, the MPC optimization problem instance constructed in the previous step is a mathematical problem. It describes how to select a sequence of control inputs for the next P steps, while adhering to all physical constraints, such that a cost function measuring the future system performance is minimized. To find this optimal sequence of control inputs, powerful mathematical tools—numerical optimization algorithms—are needed. These algorithms can iteratively search for combinations of control inputs, evaluate the future state trajectories predicted by the system model, calculate the corresponding cost function values, and check whether all state and input constraints are satisfied.

[0098] Specifically, in a specific example of this application, the specific implementation process of step S51 is as follows: The MPC optimization problem instance is a formal mathematical description of the control decision problem to be solved at the current time step. It contains all the necessary information for the numerical optimization algorithm to solve: First, it contains the current system state vector for optimization, which serves as the starting point for prediction and the initial state in the optimization process.

[0099] Secondly, it includes a predictable sequence of disturbances, these external influences being predicted over the next P time steps and considered when predicting the evolution of the system state. Thirdly, it includes a description of the system's dynamic model, which can be a system of physical equations or its simplified form, used to predict the state at the next moment based on the current state, control inputs, and disturbances. The core part is the definition of the cost function, a mathematical expression quantifying control performance and cost; the algorithm aims to minimize this function. For example, the cost function might be defined as a weighted sum (e.g., sum of squares) of the deviations between the predicted system state (e.g., temperature) and the desired setpoint over the next P time steps, plus a weighted sum (e.g., the sum of squares of control inputs or the sum of squares of changes in control inputs) of the "costs" incurred by the magnitude or rate of change of the applied control inputs. The weights reflect the priority of different optimization objectives, such as energy saving or precise temperature control. Finally, it includes constraints, a set of satisfied mathematical inequalities or equations. They describe the physical limitations and operational requirements of the system. For example, state variables (such as battery temperature and motor temperature) need to be maintained within a safe or comfortable range, and control inputs (such as cooling pump power and fan speed) need to be within the upper and lower limits that the actuators can physically allow.

[0100] Numerical optimization algorithms are used to find iterative processes that minimize (or maximize) the cost function while satisfying the constraints mentioned above. For the MPC problem, the goal is to find a sequence of control inputs for the next P time steps such that, starting from the current state and following the evolution of the system's dynamic model, the state trajectory for the next P steps minimizes the cost function while satisfying all state and control constraints.

[0101] Specifically, numerical optimization algorithms include: an evaluation mechanism for cost and constraint functions. In each iteration, the algorithm needs to simulate and predict the corresponding future P-step state trajectory using a system dynamic model, based on a future P-step control input sequence to be evaluated. Based on these trajectories and the control sequence, the current value of the cost function is calculated, and all constraints are checked for satisfaction (or violation). A sensitivity information calculation method is also required. Most efficient numerical optimization algorithms need to understand the "sensitivity" of the cost and constraint functions to the control input, i.e., how quickly they change with the control input. This involves calculating the partial derivatives (i.e., gradients or Jacobian matrices) of the cost and constraint functions with respect to the future P-step control input sequence, and sometimes even the second-order partial derivatives (Hessian matrices). These partial derivatives can be calculated analytically, automatically differentiated, or numerically (e.g., finite difference). Finally, a search direction determination strategy is needed. Based on the current cost function value, constraint violations, and calculated sensitivity information, the algorithm needs to decide in which "direction" the control input should be adjusted to more effectively reduce the cost function value or satisfy the constraints, based on the current control input sequence. Different algorithms employ different direction-determining strategies. For example, gradient descent follows the negative gradient direction, Newton's method uses second-order information, and methods like Sequential Quadratic Programming (SQP) determine the direction by solving a quadratic programming subproblem. Step size selection determines how far to proceed along the calculated search direction. A suitable step size must ensure both a decrease in the cost function (for unconstrained optimization) or improvement of constraints, while also guaranteeing the algorithm's stability and convergence. This involves techniques such as line search or trust regions. Iterative update rules update the current control input sequence based on the determined search direction and step size, generating a new and better estimate of the control input sequence. Stopping criteria define when a satisfactory solution is found and iteration terminates. Common stopping criteria include: the change in cost function value is less than a certain threshold, the degree of constraint violation is less than a certain tolerance, the norm of the gradient is sufficiently small, or the preset maximum number of iterations has been reached.

[0102] Specifically, in a concrete example of this application, the MPC optimization problem instance includes: a current cabin temperature estimate of 26°C, a target cabin temperature of 22°C, a predicted ambient temperature sequence for the next three steps [30°C, 31°C, 32°C], and upper and lower limits for PTC heater power [0W, 1000W] and air conditioning compressor power [0W, 2000W]. The system dynamic model describes the thermodynamic relationship of how cabin temperature changes with heater power, compressor power, ambient temperature, and occupant heat generation (potentially treated as a known perturbation). The cost function is set to minimize the sum of squares of the deviations between the predicted cabin temperature for the next three steps and the target value of 22°C, plus the sum of squares of the heater and compressor power for the next three steps (as energy cost). The constraints are that the predicted cabin temperature must be within the range of [21°C, 24°C], and the heater and compressor power must be within their respective upper and lower limits. An SQP solver in the online optimization solution module receives this information. The solver initializes a control sequence for the next three steps, such as heater power [0W, 0W, 0W] and compressor power [500W, 500W, 500W]. It then iterates: simulating the system's dynamic model to predict the cabin temperature trajectory; calculating the cost function value and constraint violations; calculating the gradient with respect to the power sequence; solving the QP subproblem to determine the search direction and step size; updating the power sequence; repeating this process until an optimal sequence is found, such as heater power [0W, 0W, 0W] and compressor power [800W, 750W, 700W], which allows the predicted cabin temperature trajectory to quickly and comfortably reach and remain near the target, while satisfying all constraints and consuming minimal energy. The solver ultimately returns this optimal control input sequence [[0W, 800W], [0W, 750W], [0W, 700W]].

[0103] Accordingly, the MPC algorithm employs a predictive control strategy, calculating the optimal control sequence for a future time period in each control cycle. However, only the first control instruction in this sequence is actually applied to the system actuator. This is because MPC executes cyclically: by the time the next control cycle arrives, the system state has changed, and new predictive disturbance information is available. At this point, a new MPC optimization problem is reconstructed and solved. Therefore, in the future control sequence calculated in the previous cycle, except for the first instruction, the remaining instructions are no longer optimal due to the new information. This "cyclic optimization, rolling execution" characteristic of MPC allows the controller to continuously utilize the latest system state and predictive information to correct its control strategy, thereby improving the adaptability and robustness of the control. Therefore, from the entire calculated optimal control input sequence, only the control instruction corresponding to the current moment (the first time step in the predictive time domain) is extracted and applied, which is the current optimal control instruction.

[0104] Specifically, in a specific example of this application, step S52 is implemented as follows: The optimal control input sequence is obtained, each corresponding to the optimal control command within P future time steps. The control vector located at the first position (corresponding to the first time step starting from the current moment) is extracted from this sequence. This extracted control vector is the current optimal control command. This control command is a multi-dimensional vector containing the setpoint or action command that needs to be sent to each actuator (such as a water pump, compressor, electronic expansion valve, fan, etc.) of the thermal management system within the current control cycle. This command is sent to the drive actuators of the thermal management system, converted into physical drive signals, and applied to the system.

[0105] In step S6, the current optimal control command is sent to the drive actuators of the electric vehicle thermal management system to obtain physical drive signals and updated real-time vehicle status data. It should be understood that the output of the control algorithm must act on the actual physical system to achieve system regulation. The ultimate goal of the aforementioned complex calculation and optimization process is to generate this current optimal control command. This command is a digital signal representing how to operate the various actuators of the thermal management system at the current moment to achieve optimal performance. However, actuators (such as water pumps, fans, compressors, electronic expansion valves, etc.) cannot directly understand digital commands; they need to receive specific physical signals (such as voltage, current, PWM signals, CAN bus commands, etc.) to drive their actions. Therefore, the digital control command needs to be converted into physical drive signals that the actuators can recognize and respond to. When these physical drive signals act on the actuators, the actuators adjust their operating state according to the command (such as changing the water pump speed, compressor power, valve opening, etc.), thereby changing the coolant flow rate, refrigerant flow rate, air volume, etc., and thus affecting the temperature, pressure, and other states of various components in the thermal management system. Changes in the system state are measured by sensors, generating updated real-time vehicle state data that reflects the current state of the system. This updated data serves as the sensing input for the next control cycle, used to construct a new current system state vector, driving the state estimation, disturbance prediction, and MPC optimization processes of the next cycle, thus forming closed-loop control.

[0106] Specifically, in a specific example of this application, step S6 is implemented as follows: The current optimal control command from the previous step is sent to the thermal management system controller or related drive interface module of the electric vehicle. This current optimal control command is a vector containing the target values ​​that each actuator in the thermal management system should set within the current control cycle. For example, the command includes the target water pump speed, target compressor torque or speed, target electronic expansion valve opening, target fan speed, etc. After receiving these digital commands, the thermal management system controller or drive interface module converts these digital values ​​into corresponding physical drive signals according to the preset drive strategy or interface specification. For example, the target water pump speed is converted into a specific PWM signal or voltage signal and sent to the water pump driver; the target compressor torque or speed is converted into a corresponding current or CAN bus command and sent to the compressor controller; the target valve opening is converted into a stepper motor pulse count or voltage signal and sent to the electronic expansion valve actuator. These physical drive signals directly act on each actuator of the thermal management system, causing them to execute the actions specified in the command.

[0107] When the actuators change their operating state according to physical drive signals, the physical state of the thermal management system (such as the temperature, pressure, and flow rate of various components) changes. These changes are measured in real time by sensors (temperature sensors, pressure sensors, flow sensors, etc.) deployed throughout the system. The raw analog or digital signals generated by these sensors are acquired by the vehicle's bus system (such as the CAN bus) and transmitted to the main controller or thermal management domain controller. After necessary signal processing, filtering, and format conversion, the acquired raw sensor data constitutes updated real-time vehicle state data reflecting the current actual operating status of the system. This data is updated periodically and is used at the beginning of the next control cycle to construct a new current system state vector, thereby driving the iterative execution of the entire MPC control cycle.

[0108] In summary, the automatic control method for the electric vehicle thermal management system based on the embodiments of this application is explained. By integrating external information such as navigation paths and weather forecasts, it achieves multi-step prediction of future battery and motor heat generation and thermal load changes, enabling advance planning of thermal management strategies and improving system response foresight. A multi-source disturbance dynamic structured correlation equilibrium technology is employed to model the coupling relationships between various heat sources, and cross-disturbances are compensated through joint state functions to achieve multi-thermal loop collaborative control. Based on the model predictive control (MPC) framework, temperature tracking error and energy consumption penalties are optimized to reduce the energy consumption of actuators such as compressors and water pumps while meeting safe temperature ranges, thus extending driving range. Combining the state estimator with the fusion model and observation data enhances control accuracy and robustness under complex operating conditions, ultimately achieving globally optimal thermal management that balances component lifespan, passenger comfort, and energy efficiency.

[0109] Figure 4 This is a block diagram of an automatic control device for an electric vehicle thermal management system according to an embodiment of this application. Figure 4 As shown, the automatic control device 100 of the electric vehicle thermal management system according to an embodiment of this application includes: a data acquisition module 110, used to acquire real-time vehicle state data and external prediction information, wherein the external prediction information includes navigation path information and weather forecast information for the next N time steps; a current system state vector construction module 120, used to construct a current system state vector based on the real-time vehicle state data; an external prediction information conversion module 130, used to convert the external prediction information into a predictable disturbance sequence; an MPC optimization problem instance construction module 140, used to construct an MPC optimization problem instance based on the current system state vector and the predictable disturbance sequence; an optimal control command generation module 150, used to input the MPC optimization problem instance into an online optimization solution module to obtain a current optimal control command; and a drive execution module 160, used to send the current optimal control command to the drive actuator of the electric vehicle thermal management system to obtain a physical drive signal and updated real-time vehicle state data.

[0110] Here, those skilled in the art will understand that the specific operation of each step in the automatic control device of the above-mentioned electric vehicle thermal management system has been referenced above. Figures 1 to 3 The automatic control method of the electric vehicle thermal management system is described in detail in the description, and therefore, its repeated description will be omitted.

[0111] In summary, the above detailed description is intended to be illustrative rather than restrictive, and it should be understood that these embodiments are for illustrative purposes only and not for limiting the scope of protection of the invention. After reading the description of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent changes and modifications also fall within the scope defined by the claims of this invention.

Claims

1. An automatic control method of an electric vehicle thermal management system, characterized by, The method comprises the following steps: acquiring real-time vehicle state data and external prediction information, wherein the external prediction information comprises navigation path information and weather forecast information for N future time steps; constructing a current system state vector based on the real-time vehicle state data; transforming the external prediction information into a predictable disturbance sequence, comprising: predicting battery heat generation power and motor / electronic control heat generation power for P future steps based on the navigation path information for N future time steps to obtain a battery heat generation power sequence and a motor / electronic control heat generation power sequence; predicting thermal load on the cabin and heat exchange amount on the battery compartment for P future steps based on the weather forecast information for N future time steps to obtain a thermal load sequence and a heat exchange amount sequence; performing multi-source disturbance dynamic structured correlation balancing on the battery heat generation power sequence, the motor / electronic control heat generation power sequence, the thermal load sequence and the heat exchange amount sequence respectively to obtain a battery heat generation power balanced sequence, a motor / electronic control heat generation power balanced sequence, a thermal load balanced sequence and a heat exchange amount balanced sequence, comprising: representing the battery heat generation power sequence, the motor / electronic control heat generation power sequence, the thermal load sequence and the heat exchange amount sequence as row vectors respectively to obtain a set of state vectors; calculating a mean vector between the set of state vectors to obtain a state mean vector; based on the state mean vector, calculating a covariance matrix of each state vector in the set of state vectors respectively to obtain a set of state characteristic covariance matrices; performing sensitive correlation on each corresponding state vector and state characteristic covariance matrix in the set of state vectors and the set of state characteristic covariance matrices to obtain a set of joint state function values; based on each joint state function value in the set of joint state function values, constructing a stable state prediction value of the state vector corresponding thereto to obtain a set of stable state prediction values; taking each stable state prediction value in the set of stable state prediction values as a compensation coefficient to perform state balancing correlation on each state vector in the set of state vectors to obtain a set of state balancing vectors; wherein each state balancing vector in the set of state balancing vectors is the battery heat generation power balanced sequence, the motor / electronic control heat generation power balanced sequence, the thermal load balanced sequence and the heat exchange amount balanced sequence; integrating the battery heat generation power balanced sequence, the motor / electronic control heat generation power balanced sequence, the thermal load balanced sequence and the heat exchange amount balanced sequence to obtain the predictable disturbance sequence; constructing an MPC optimization problem instance based on the current system state vector and the predictable disturbance sequence; inputting the MPC optimization problem instance into an online optimization solving module to obtain a current optimal control instruction; sending the current optimal control instruction to a driving actuator of the electric vehicle thermal management system to obtain a physical driving signal and updated real-time vehicle state data.

2. The automatic control method of the electric vehicle thermal management system according to claim 1, characterized in that, The navigation path information comprises altitude, speed limit and congestion prediction; the weather forecast information comprises ambient temperature, ambient humidity and solar radiation intensity.

3. The automatic control method of the electric vehicle thermal management system according to claim 1, characterized in that, constructing an MPC optimization problem instance based on the current system state vector and the predictable disturbance sequence, comprising: performing state estimation optimization on the current system state vector based on a state estimator to obtain an optimized current system state vector; constructing the MPC optimization problem instance based on the optimized current system state vector and the predictable disturbance sequence.

4. The automatic control method of the electric vehicle thermal management system according to claim 3, characterized in that, performing state estimation optimization on the current system state vector based on a state estimator to obtain an optimized current system state vector, comprising: inputting the state estimation of the last time and the control input of the last time into a system dynamic model to predict the state of the current time to obtain a predicted current system state vector and an uncertainty of the predicted current system state vector; inputting the predicted current system state vector, the uncertainty of the predicted current system state vector and the current system state vector into an observation model to obtain the optimized current system state vector.

5. The automatic control method of the electric vehicle thermal management system according to claim 4, characterized in that, constructing the MPC optimization problem instance based on the optimized current system state vector and the predictable disturbance sequence, comprising: inputting the optimized current system state vector and the predictable disturbance sequence to predict the system state evolution of the electric vehicle thermal management system within future P steps if a series of control inputs are applied to obtain a predicted state trajectory and a predicted control trajectory; defining a cost function for the predicted state trajectory and the predicted control trajectory, the cost function comprising temperature tracking error, energy consumption and control amount change penalty and weight coefficients; applying constraints to the cost function, the constraints comprising state constraints and input constraints to obtain the MPC optimization problem instance.

6. The automatic control method of the electric vehicle thermal management system according to claim 1, characterized in that, inputting the MPC optimization problem instance into an online optimization solving module to obtain a current optimal control instruction, comprising: solving the MPC optimization problem instance using a numerical optimization algorithm to obtain an optimal control input sequence; extracting the control input of the current time from the optimal control input sequence as the current optimal control instruction.

7. An automatic control device of an electric vehicle thermal management system, characterized by, comprising: a data acquisition module configured to acquire real-time vehicle state data and external prediction information, wherein the external prediction information comprises navigation path information and weather forecast information for future N time steps; a current system state vector construction module configured to construct a current system state vector based on the real-time vehicle state data; an external prediction information conversion module configured to convert the external prediction information into a predictable disturbance sequence, comprising: predicting battery heat generation power and motor / electronic control heat generation power for future P steps based on the navigation path information for future N time steps to obtain a battery heat generation power sequence and a motor / electronic control heat generation power sequence; predicting thermal load of the cabin and heat exchange amount of the battery compartment for future P steps based on the weather forecast information for future N time steps to obtain a thermal load sequence and a heat exchange amount sequence; The battery heat generation power sequence, the motor / electronic control heat generation power sequence, the heat load sequence and the heat exchange amount sequence are respectively subjected to multi-source disturbance dynamic structural correlation balancing to obtain a battery heat generation power balanced sequence, a motor / electronic control heat generation power balanced sequence, a heat load balanced sequence and a heat exchange amount balanced sequence, including: representing the battery heat generation power sequence, the motor / electronic control heat generation power sequence, the heat load sequence and the heat exchange amount sequence as row vectors respectively to obtain a set of state vectors; calculating a mean vector between the set of state vectors to obtain a state mean vector; based on the state mean vector, calculating a covariance matrix of each state vector in the set of state vectors respectively to obtain a set of state characteristic covariance matrices; performing sensitive correlation on each corresponding state vector and state characteristic covariance matrix in the set of state vectors and the set of state characteristic covariance matrices to obtain a set of joint state function values; based on each joint state function value in the set of joint state function values, constructing a stable state prediction value of the state vector corresponding thereto to obtain a set of stable state prediction values; taking each stable state prediction value in the set of stable state prediction values as a compensation coefficient, performing state balancing correlation on each state vector in the set of state vectors to obtain a set of state balancing vectors; wherein each state balancing vector in the set of state balancing vectors is the battery heat generation power balanced sequence, the motor / electronic control heat generation power balanced sequence, the heat load balanced sequence and the heat exchange amount balanced sequence; The battery heat generation power balanced sequence, the motor / electronic control heat generation power balanced sequence, the heat load balanced sequence and the heat exchange amount balanced sequence are integrated to obtain the predictable disturbance sequence; An MPC optimization problem instance construction module is configured to construct an MPC optimization problem instance based on the current system state vector and the predictable disturbance sequence; An optimal control instruction generation module is configured to input the MPC optimization problem instance into an online optimization solving module to obtain a current optimal control instruction; A driving execution module is configured to send the current optimal control instruction to a driving actuator of the electric vehicle thermal management system to obtain a physical driving signal and updated real-time vehicle state data.

Images