Model order reduction acceleration method for HIL simulation of automobile thermal system model

Abstract

The invention relates to a model reduced-order acceleration method for HIL simulation of an automobile thermal system model, and belongs to the technical field of automobile thermal system modeling and simulation, and the method comprises the following steps: building a thermal system model: building the thermal system model, generating an input-output data set covering all working conditions, and setting a control equation of the thermal system model; performing principal component analysis (PCA) data preprocessing: compressing high-dimensional space-time data into a low-dimensional feature space, and reserving a dominant dynamic mode; neural state space modeling: designing a nonlinear dynamic system to describe state evolution after dimensionality reduction, establishing a low-dimensional nonlinear dynamic model, and performing network training and regularization; and model integration simulation: integrating the heat exchanger model after order reduction into a complete thermal system model, and performing real-time simulation optimization.

Description

Technical Field

[0001] The present invention belongs to the technical field of automotive thermal system modeling and simulation, and relates to a model reduction and acceleration method for Hardware-in-the-Loop (HIL) simulation of automotive thermal system models. Background Art

[0002] In recent years, the new energy vehicle industry has developed rapidly, and the demand for its thermal management system has also increased accordingly. The existing automotive thermal system modeling methods mainly rely on physical-based models, which usually include complex heat transfer, fluid mechanics, and heat exchange equations. Although these physical models can provide high accuracy, due to their high computational complexity, they often cannot meet the real-time requirements in actual Hardware-in-the-Loop (HIL) simulations. To improve the simulation efficiency, some studies have tried to use simplified models or look-up tables, but traditional methods often rely on manual experience or overly simplified linear assumptions and cannot significantly improve the simulation speed while ensuring accuracy. In addition, although the existing modeling methods have made certain progress in some fields, their applications in complex systems such as automotive thermal management still face problems of accuracy loss and large computational volume. Especially in the automotive industry, how to find a balance between ensuring the accuracy and real-time performance of the thermal system model remains a technical problem to be solved urgently. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to provide a model reduction and acceleration method for Hardware-in-the-Loop (HIL) simulation of automotive thermal system models to solve the problems of large computational volume, slow simulation speed, and poor simulation real-time performance in automotive thermal system models.

[0004] To achieve the above purpose, the present invention provides the following technical solutions:

[0005] A model reduction and acceleration method for Hardware-in-the-Loop (HIL) simulation of automotive thermal system models, comprising the following steps:

[0006] Thermal system model construction: Construct a thermal system model, generate an input-output data set covering all working conditions, and set the control equations of the thermal system model;

[0007] Principal Component Analysis (PCA) data preprocessing: Compress high-dimensional spatio-temporal data into a low-dimensional feature space and retain the dominant dynamic modes;

[0008] Neural state space modeling: Design a nonlinear dynamic system to describe the state evolution after dimension reduction, establish a low-dimensional nonlinear dynamic model, and perform network training and regularization;

[0009] Model integrated simulation: Integrate the reduced-order heat exchanger model into the complete thermal system model for real-time simulation optimization.

[0010] Furthermore, the method further comprises the following steps:

[0011] Adopt Gaussian noise injection and impulse interference;

[0012] Superimpose zero-mean Gaussian noise on the training data, as shown in the following formula:

[0013] u t ' = u t + N(0, σ 2 ), σ = 5%·max(u t )

[0014] Randomly insert a step change in mass flow rate:

[0015] Q'(t) = Q(t) + ΔQ·H(t - t s ), ΔQ ∈ [-50%, +100%]

[0016] where H(·) is the step function and t s is the random mutation moment

[0017] Compensate the model prediction error in real time during the simulation process;

[0018] Solve the model drift problem by incremental learning to adapt to the changes in the working condition environment.

[0019] Furthermore, the construction of the thermal system model specifically includes the following steps:

[0020] S11: Construct a high-fidelity Simscape model of the heat exchanger, covering the coupling effects of thermodynamics, fluid dynamics and structural mechanics. The goal is to construct a one-dimensional high-precision heat exchanger model to simulate the mass flow rate, temperature, pressure and flow velocity distributions along the flow direction, providing a data basis for model reduction; for one-dimensional simulation requirements, use the Simscape library to establish a one-dimensional thermal fluid model of the heat exchanger, discretize the heat exchanger into multiple control units along the flow direction, and each unit contains mass, momentum and energy conservation equations; the key equations include:

[0021] Three-dimensional unsteady heat conduction equation:

[0022]

[0023] Its physical meaning is to describe the energy conservation relationship of the temperature field T(x, y, z, t) evolving with time, is the heat source term;

[0024] Navier-Stokes equation for incompressible fluids:

[0025]

[0026] It describes the dynamic equilibrium of the fluid velocity field u and the pressure p, fbuoyancy Denoted as buoyancy force;

[0027] Linear elastic structural deformation equation:

[0028]

[0029] Its physical meaning is the relationship between material stress σ and strain ∈, and C is the elastic tensor;

[0030] S12: Parametric data acquisition:

[0031] Generate an input-output data set D covering the full operating conditions through Simscape simulation:

[0032] Input variables: coolant flow rate Q(t), inlet temperature T in (t), ambient temperature T amb , denoted as u = [Q, T in , T amb T ;

[0033] Output variables: heat exchanger temperature field T(x, t) and pressure field p(x, t), discretized into a high-dimensional vector y ∈ R d (d ≥ 10 3 );

[0034] S13: Control equation setting:

[0035] The setting of the model control module includes the various conservations of the refrigerant during the process of transporting energy with air and materials;

[0036] Discrete form mass conservation equation:

[0037]

[0038] where U1 = [(ρA)1, (ρA)2, ···] T , U2 = [(ρuA)1, (ρuA)2, ···] T ; D x is the convection matrix;

[0039] Discrete form momentum conservation equation:

[0040]

[0041] where is the velocity vector, is element-wise division, F is the diagonal matrix of the friction coefficient; P = [p1A, p2A, ···] T is the pressure term, which needs to be associated with the state equation and U1 and the energy variables;

[0042] Energy conservation equation:​

[0043]

[0044] where H is the diagonal matrix of heat transfer coefficients, and U3 = [(ρc p TA)1, (ρc p TA)2, ···] T ;

[0045] Wall heat transfer equation:

[0046]

[0047] where D xx is the second-order diffusion matrix,

[0048] Furthermore, in the PCA data preprocessing step of principal component analysis, the following steps are specifically included:

[0049] S21: Data standardization:

[0050] Normalize the input and output data to eliminate the influence of dimensions:

[0051]

[0052] where are the mean and standard deviation of all unit temperature data respectively, and the same applies to other variables; The time advancement adopts an implicit format, and the Jacobian matrix is constructed to handle the non-linear terms;

[0053] S22: Principal component analysis PCA:

[0054] Compress the high-dimensional spatio-temporal data into a low-dimensional feature space and retain the dominant dynamic modes;

[0055] Data matrix construction: Expand the M groups of time series data into a three-dimensional matrix Y ∈ R M×3N×T , where T is the number of time steps; Convert it into a two-dimensional matrix, and each row corresponds to a standardized state vector at a time step;

[0056] Covariance matrix calculation: To avoid the problem of large matrix storage, the incremental PCA algorithm is used to calculate in batches:

[0057]

[0058] Eigenvalue decomposition and mode selection:

[0059] CW = WA

[0060] where A = diag(λ1, λ2, λ3, ····λ d)The eigenvalues are arranged in descending order, and W is the eigenvector matrix;

[0061] Low-dimensional projection: Retain the first r principal components, where r is much smaller than d:

[0062]

[0063] The energy conservation principle is

[0064] Furthermore, the steps of neural state space modeling specifically include the following steps;

[0065] S31: Construction of the state space equation:

[0066] Design a non-linear dynamic system to describe the state evolution after dimensionality reduction. The goal is to establish a low-dimensional non-linear dynamic model to describe the state z k evolving with the input u k The evolution law is as follows:

[0067] z k+1 = f θ (z k , u k ) + ∈ k

[0068]

[0069] where f θ is the Neural space state, the input is the current state z k and the control quantity u k , and the output is the next state z k+1 ; Add a mass conservation penalty term to the loss function to force the model to satisfy global mass conservation. The physical constraint is embedded as shown below:

[0070]

[0071] where and are the reconstructed density and flow velocity, and λ is the weight coefficient;

[0072] S32: Network training and regularization:

[0073] Improve the long-term stability of the model through loss function design, model learning, optimization algorithm steps prediction, and curriculum learning; Network training and regularization, and the objective function includes the prediction error and the regularization term:

[0074]

[0075] During this process, the Adam optimizer is used, and the learning rate is adaptively adjusted;

[0076] Finally, the Koopman operator theory is used to enhance the non - linear modeling ability:

[0077] K * ψ(z k ) = ψ(z k+1 )

[0078] where K is the Koopman operator, and ψ(·) is a non - linear dimensionality - raising function approximated by a neural network.

[0079] Furthermore, the model integration and simulation steps include the following steps:

[0080] S41: Integrate the reduced - order heat exchanger model into the complete thermal system model and replace the original high - fidelity Simscape heat exchanger sub - module; during the integration process, ensure the continuity of the mass flow rate at the interface, that is, the mass flow rate at the pump outlet is equal to the inlet flow rate of the reduced - order model, as shown in the following formula:

[0081]

[0082] where is the inlet flow velocity of the reduced - order model reconstruction;

[0083] At the same time, ensure the energy continuity, that is, the enthalpy value at the front - end outlet is equal to the inlet enthalpy value of the reduced - order model, as shown in the following formula:

[0084]

[0085] where, is the inlet temperature of the heat exchanger coupling, is the mass flow rate;

[0086] Use the relaxation factor method to enhance the numerical stability. Introduce a relaxation factor α at the interface and gradually update the coupling variables. Check the mass / energy residuals after each coupling. If it exceeds the threshold, recalculate to achieve iterative correction, as shown in the following formula:

[0087]

[0088] S42: Real - time simulation optimization:

[0089] During the real - time simulation optimization process, the goal is to achieve real - time calculation with a 1 - ms time step on HIL hardware, including the following steps:

[0090] First, discretize the state equation into the explicit Euler form:

[0091] z k+1 = z k + Δt·f θ (z k , u k )

[0092] The time step Δt is strictly synchronized with the HIL hardware clock; when generating the code, the trained and coupled system model is converted into C code by MATLAB Coder, and the dynamic memory allocation is disabled and the array size is fixed; the weight matrix is solidified into a pre-computed look-up table to reduce the online computation amount; the timer interrupt service program of the real-time operating system is used to trigger the model calculation once every 1 ms to achieve hardware synchronization; if the single-step calculation times out, linear interpolation compensation is enabled, as shown in the following formula:

[0093] z k+1 = 2z k - z k-1 .

[0094] The beneficial effects of the present invention are as follows: The present invention provides a data-driven model order reduction modeling method for automotive thermal system HIL simulation. By using the PCA+Neural state space method for model order reduction, the model complexity of the one-dimensional high-fidelity heat exchanger model is reduced, the model memory occupancy is decreased, the CPU utilization rate in HIL testing is increased, and the hardware cost is significantly reduced. The hybrid model order reduction architecture, conservation constraint interface and real-time-robust co-optimization overcome the contradictions among speed, accuracy and stability in one-dimensional automotive thermal system HIL simulation, provide key technical support for the rapid control prototype development and testing of intelligent networked vehicles, and have significant engineering application value and market competitiveness.

[0095] Other advantages, objectives and features of the present invention will be described to some extent in the subsequent description, and to some extent, will be obvious to those skilled in the art based on the study of the following text, or can be learned from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be described in detail preferably with reference to the accompanying drawings, where:

[0097] Figure 1 It is a flow chart of the model order reduction acceleration method for automotive thermal system model HIL simulation;

[0098] Figure 2 It is a data generation flow chart;

[0099] Figure 3 It is a heat exchanger model;

[0100] Figure 4 It is a real-time simulation flow chart. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0101] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0102] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and thus the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0103] In the following description, numerous details are discussed to provide a more thorough explanation of the embodiments of the present invention. However, it is obvious to those skilled in the art that the embodiments of the present invention can be implemented without these specific details. In other embodiments, well-known structures and devices are shown in the form of block diagrams rather than in detail to avoid making the embodiments of the present invention difficult to understand.

[0104] The present invention provides that in the simulation of automotive thermal systems, the one-dimensional model is widely used due to its computational efficiency and engineering practicality, but there are still the following bottlenecks:

[0105] Difficulty in nonlinear dynamic modeling: Traditional order reduction methods (such as proper orthogonal decomposition (POD)) rely on linear assumptions and are difficult to capture nonlinear phenomena such as turbulent transitions and phase change heat transfer.

[0106] High computational complexity: The one-dimensional model requires solving thousands of differential equations (such as the mass-momentum-energy conservation equation) simultaneously. A single simulation takes tens of minutes, which cannot meet the millisecond-level real-time requirements of HIL.

[0107] Poor stability of coupling interface: When the reduced-order model is coupled with the original system, it is easy to cause numerical divergence due to mass / energy non-conservation, affecting the simulation reliability.

[0108] How to significantly improve the calculation speed of one-dimensional high-fidelity models while ensuring reliable accuracy and ensure the robustness of reduced-order models under variable working conditions is a core issue that needs to be urgently addressed in the current field.

[0109] The present invention belongs to the technical field of automotive thermal management system simulation, and particularly relates to a hardware-in-the-loop (HIL) real-time simulation method based on one-dimensional high-fidelity models and data-driven reduced order. This method aims at the heat exchanger, a core component in the automotive thermal system. By integrating principal component analysis (PCA) and neural state space modeling techniques, it significantly reduces the model complexity while retaining key dynamic characteristics, ultimately achieving real-time simulation and meeting the strict real-time requirements of HIL tests. To solve the above problems, the present invention provides the following technical solutions, and the overall technical process is as Figure 1 shown.

[0110] Example 1:

[0111] The present invention includes the following steps:

[0112] 1. Provide a data-driven model reduction and modeling method for automotive thermal system HIL simulation, including:

[0113] 1.1 Construct a thermal system model

[0114] Construct a high-fidelity Simscape model of the heat exchanger, covering the coupling effects of thermodynamics, fluid dynamics, and structural mechanics. The goal is to construct a one-dimensional high-precision heat exchanger model to accurately simulate the mass flow rate, temperature, pressure, and flow velocity distributions along the flow direction, providing a data basis for model reduction. For one-dimensional simulation requirements, use the Simscape library to establish a one-dimensional thermal-fluid model of the heat exchanger, discretize the heat exchanger into multiple control units along the flow direction, and each unit contains mass, momentum, and energy conservation equations. The key equations include:

[0115] The three-dimensional unsteady heat conduction equation as shown in Equation 1:

[0116]

[0117] Its physical meaning is to describe the energy conservation relationship of the temperature field T(x, y, z, t) evolving with time, being the heat source term.

[0118] The Navier-Stokes equation for incompressible fluids as shown in Equation 2:

[0119]

[0120] It describes the dynamic balance of the fluid velocity field u and pressure p, f buoyancy representing buoyancy.

[0121] The structural deformation equation for linear elasticity as shown in Equation 3:

[0122]

[0123] Its physical meaning is the relationship between material stress σ and strain ∈, and C is the elastic tensor.

[0124] 1.2 Parametric data acquisition

[0125] Generate an input-output data set D covering the full operating conditions through Simscape simulation:

[0126] Input variables: coolant flow rate Q(t), inlet temperature T in (t), ambient temperature T amb , denoted as u = [Q, T in , T amb T .

[0127] Output variables: heat exchanger temperature field T(x, t) and pressure field p(x, t), discretized into a high-dimensional vector y ∈ R d (d ≥ 10 3 ).

[0128] The data generation process is as Figure 2 shown.

[0129] 1.3 Control equation setting

[0130] The setting of the model control module mainly includes various conservations of the refrigerant with air and materials during the energy transfer process.

[0131] Discrete form of the mass conservation equation:

[0132]

[0133] where U1 = [(ρA)1, (ρA)2, ···] T , U2 = [(ρuA)1, (ρuA)2, ···] T

[0134] D x is the convection matrix.

[0135] Discrete form of the momentum conservation equation:

[0136]

[0137] where is the velocity vector, is element-wise division, and F is the diagonal matrix of the friction coefficient. P = [p1A, p2A, ···] T is the pressure term, which needs to be related to U1 and the energy variables by the state equation.

[0138] Energy conservation equation:

[0139] ​

[0140] where H is the diagonal matrix of heat transfer coefficients, and U3 = [(ρc p TA)1, (ρc p TA)2, ···] T .

[0141] Wall heat transfer equation:

[0142]

[0143] where D xx is the second-order diffusion matrix,

[0144] 2. PCA data preprocessing

[0145] 2.1 Data standardization

[0146] Normalize the input and output data to eliminate the influence of dimensions:

[0147]

[0148] where are the mean and standard deviation of all unit temperature data respectively, and the same applies to other variables. The time advancement often uses an implicit format, and the Jacobian matrix needs to be constructed to handle the non-linear terms.

[0149] 2.2 Principal component analysis (PCA)

[0150] Compress the high-dimensional spatio-temporal data into a low-dimensional feature space and retain the dominant dynamic modes.

[0151] Data matrix construction: Expand the M groups of time series data into a three-dimensional matrix Y ∈ R M×3N×T , where T is the number of time steps. Convert it into a two-dimensional matrix, and each row corresponds to the standardized state vector at a time step.

[0152] Covariance matrix calculation: To avoid the problem of large matrix storage, the incremental PCA algorithm is used to calculate in batches.

[0153]

[0154] Eigenvalue decomposition and mode selection:

[0155] CW = WA (10)

[0156] where A = diag(λ1, λ2, λ3, ····λ d ) is the eigenvalue arranged in descending order, and W is the eigenvector matrix.

[0157] Low-dimensional projection: Retain the first r principal components (r is much smaller than d):

[0158]

[0159] The principle of energy conservation is

[0160] 3. Neural state space reduced-order modeling

[0161] 3.1 Construction of state space equation

[0162] Design a nonlinear dynamic system to describe the state evolution after dimension reduction. The goal is to establish a low-dimensional nonlinear dynamic model to describe the state z k evolving with the input u k . The state equation is as follows:

[0163] z k+1 = f θ (z k , u k ) + ∈ k

[0164]

[0165] where f θ is the Neural space state, the input is the current state z k and the control quantity u k , and the output is the next state z k+1 . Add a mass conservation penalty term to the loss function to force the model to satisfy global mass conservation. The physical constraint is embedded as shown in the following formula:

[0166]

[0167] where and are the reconstructed density and flow velocity, and λ is the weight coefficient

[0168] 3.2 Network training and regularization

[0169] This step improves the long-term stability of the model through multi-step prediction and curriculum learning, including loss function design, model learning, and optimization algorithms. For network training and regularization, the objective function includes the prediction error and the regularization term:

[0170]

[0171] During this process, the Adam optimizer is used, and the learning rate is adaptively adjusted.

[0172] Finally, use the Koopman operator theory to enhance the nonlinear modeling ability:

[0173] K * ψ(zk ) = ψ(z k+1 ) (15)

[0174] where K is the Koopman operator, ψ(·) is a non-linear dimensionality-raising function approximated by a neural network.

[0175] 4. Model Coupling and Real-Time Simulation

[0176] 4.1 Coupling Interface Design

[0177] The complete thermal system includes one-dimensional components such as pumps, valves, and pipelines. In the present invention, the reduced-order heat exchanger model is integrated into the complete thermal system model to replace the original high-fidelity Simscape heat exchanger sub-module. The schematic diagram of the discrete heat exchanger model is as shown in Figure 3 shown.

[0178] During the integration process, the mass flow rate at the interface is ensured to be continuous, that is, the mass flow rate at the pump outlet is equal to the inlet flow rate of the reduced-order model, as shown in the following equation:

[0179]

[0180] where is the inlet flow velocity reconstructed by the reduced-order model.

[0181] At the same time, the energy is ensured to be continuous, that is, the enthalpy value at the front-end outlet is equal to the inlet enthalpy value of the reduced-order model, as shown in the following equation:

[0182]

[0183] where,[[]] is the inlet temperature of the heat exchanger coupling,[[]] is the mass flow rate.

[0184] Meanwhile, the numerical stability is enhanced by the relaxation factor method. A relaxation factor α is introduced at the interface, and the coupling variables are gradually updated. After each coupling, the mass / energy residuals are checked. If they exceed the threshold, recalculation is performed to achieve iterative correction. As shown in the following equation:

[0185]

[0186] 4.2 Real-Time Simulation Optimization

[0187] During the real-time simulation optimization process, the goal is to achieve real-time calculation with a 1 ms time step on the HIL hardware. The real-time simulation flow chart of the coupled system is as shown in Figure 4 shown.

[0188] First, the state equation is discretized into the explicit Euler form:

[0189] z k+1 = z k+Δt·f θ (z k ,u k ) (19)

[0190] The time step Δt is strictly synchronized with the HIL hardware clock. When generating the code, MATLAB Coder is used to convert the trained and coupled system model into C code. By disabling dynamic memory allocation, the array sizes are fixed. The weight matrices are solidified into pre-computed look-up tables to reduce the online computation. Using the timer interrupt service program of the real-time operating system, the model calculation is triggered every 1 ms to achieve hardware synchronization. If the single-step calculation times out, linear interpolation compensation is enabled, as shown in the following formula:

[0191] z k+1 =2z k -z k-1 (20)

[0192] 5. Robustness Enhancement Techniques

[0193] 5.1 Noise Injection

[0194] To improve the adaptability of the model to sensor noise and sudden changes in operating conditions and enable the model to realistically simulate real-world operating conditions, Gaussian noise injection and pulse interference are adopted.

[0195] Zero-mean Gaussian noise is superimposed on the training data, as shown in the following formula:

[0196] u t '=u t +N(0,σ 2 ),σ=5%·max(u t ) (21)

[0197] Randomly insert a step change in the mass flow rate:

[0198] Q'(t)=Q(t)+ΔQ·H(t-t s ),ΔQ∈[-50%,+100%] (22)

[0199] where H(·) is the step function and t s is the random mutation time.

[0200] At the same time, during the simulation process, the model prediction error is compensated in real time to prevent error accumulation. The incremental learning is used to adapt to the changes in the operating conditions to solve the problem of model drift caused by the changes in the long-term operating environment.

[0201] Example 2:

[0202] For an electric vehicle motor cooling circuit, the length of the heat exchanger L = 2m, which is discretized into N = 100 units, and the simulation duration is 1000s.

[0203] The computing speed is increased by 200 times, and the CPU occupancy rate is reduced to less than 25%.

[0204] Under sudden traffic changes and noise interference, the model still maintains high precision (RMSE ≤ 2.5%).

[0205] The results are shown in Table 1.

[0206] Table 1

[0207] Index High-fidelity Simscape model Reduced-order coupling model Single-step running time 10 ms 0.05 ms Maximum temperature error - 1.2％ HIL CPU occupancy rate 90％ 22％ Overshoot of step response - ≤3％ RMSE in noise environment - ≤2.5％

[0208] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the present technical solution, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A model reduction acceleration method for HIL simulation of an automotive thermal system model, characterized by: The following steps are involved: Thermal system model construction: Build a thermal system model, generate input-output data sets covering all operating conditions, and set the control equations of the thermal system model; Principal component analysis (PCA) data preprocessing: compress high-dimensional spatiotemporal data into low-dimensional feature space and retain the dominant dynamic mode; Neural state space modeling: Design nonlinear dynamic systems to describe the state evolution after dimensionality reduction, establish low-dimensional nonlinear dynamic models, and perform network training and regularization; Model integration simulation: Integrate the reduced-order heat exchanger model into the complete thermal system model for real-time simulation optimization.

2. The model reduction acceleration method for HIL simulation of an automotive thermal system model according to claim 1, characterized in that: The method further comprises the following steps: Using Gaussian noise injection and pulse interference; Zero-mean Gaussian noise is superimposed on the training data, as shown in the following formula: you t '=u t +N(0,σ 2 ),σ=5%·max(u t ) Randomly insert mass flow rate step changes: Q'(t)=Q(t)+ΔQ·H(t-t s ),ΔQ∈[-50%,+100%] Where H(·) is a step function, t s Random mutation moment Compensate for model prediction errors in real time during the simulation process; The model drift problem can be solved by incremental learning to adapt to changes in working conditions and environment.

3. The model order reduction acceleration method for HIL simulation of an automobile thermal system model according to claim 1, characterized in that: The thermal system model construction specifically includes the following steps: S11: Build a high-fidelity Simscape model of the heat exchanger, covering the coupling effects of thermodynamics, fluid dynamics and structural mechanics. The goal is to build a one-dimensional high-precision heat exchanger model to simulate the mass flow, temperature, pressure and flow velocity distribution along the flow direction, and provide a data basis for order reduction. In response to the one-dimensional simulation requirements, the Simscape library is used to establish a one-dimensional thermal fluid model of the heat exchanger, discretizing the heat exchanger into multiple control units along the flow direction, each of which contains mass, momentum and energy conservation equations; key equations include: Three-dimensional unsteady heat conduction equation: Its physical meaning is to describe the energy conservation relationship of the temperature field T(x,y,z,t) evolving over time. is the heat source term; Navier-Stokes equations for incompressible fluids: It describes the dynamic balance between the fluid velocity field u and the pressure p, f buoyancy Expressed as buoyancy; The structural deformation equation for linear elasticity is: Its physical meaning is the relationship between material stress σ and strain ∈, and C is the elastic tensor; S12: Parametric data acquisition: Generate input-output data set D covering all working conditions through Simscape simulation: Input variables: coolant flow rate Q(t), inlet temperature T in (t), ambient temperature T amb , denoted as u=[Q,T in ,T amb ] T ; Output variables: heat exchanger temperature field T(x, t) and pressure field p(x, t), discretized into high-dimensional vector y∈R d (d≥10 3 ); S13: Control equation setting: The setting of the model control module includes the major conservation of refrigerant with air and materials in the process of transporting energy; Discrete form of the mass conservation equation: where U1=[(ρA)1,(ρA)2,···] T , U2=[(ρuA)1,(ρuA)2,···] T ;D x is the convection matrix; Discrete form of the momentum conservation equation: in is the velocity vector, is element-by-element division, F is the diagonal matrix of friction coefficients; P = [p1A, p2A, ···] T is the pressure term, which requires the state equation to be associated with U1 and the energy variable; Energy conservation equation: Where H is the diagonal matrix of heat transfer coefficients, And U3=[(ρc p TA)1,(ρc p TA)2,···] T ; Wall heat transfer equation: Where D xx is the second-order diffusion matrix, 4. The model order reduction acceleration method for HIL simulation of an automobile thermal system model according to claim 1, characterized in that: The principal component analysis PCA data preprocessing step specifically includes the following steps: S21: Data Standardization: Normalize the input and output data to eliminate the dimension effect: in are the mean and standard deviation of all unit temperature data respectively, and the same is true for other variables; the time advancement adopts implicit format and constructs Jacobian matrix to deal with nonlinear terms; S22: Principal Component Analysis PCA: Compress high-dimensional spatiotemporal data into low-dimensional feature space and retain the dominant dynamic modes; Data matrix construction: Expand M groups of time series data into a three-dimensional matrix Y∈R by time step M×3N×T , where T is the number of time steps; converted to a two-dimensional matrix, each row corresponds to the standardized state vector of a time step; Covariance matrix calculation: To avoid large matrix storage problems, the incremental PCA algorithm is used to calculate in batches: Eigenvalue decomposition and mode selection: CW=WA where A = diag (λ1, λ2, λ3,····λ d ) is the descending order of eigenvalues, W is the eigenvector matrix; Low-dimensional projection: retain the first r principal components, where r is much smaller than d: The energy conservation principle is 5. The model reduction acceleration method for HIL simulation of an automobile thermal system model according to claim 1, characterized in that: The neural state space modeling step specifically includes the following steps; S31: State space equation construction: Design a nonlinear dynamic system to describe the state evolution after dimensionality reduction. The goal is to establish a low-dimensional nonlinear dynamic model to describe the state z k With input u k The evolution law of , the state equation is as follows: z k+1 =f θ (z k ,u k )+∈ k where f θ is the Neural space state, and the input is the current state z k and the control quantity u k , the output is the next state z k+1 ; Add a mass conservation penalty term to the loss function to force the model to satisfy global mass conservation, and the physical constraints are embedded as shown in the following formula: in and is the reconstructed density and flow velocity, λ is the weight coefficient; S32: Network training and regularization: Improve the long-term stability of the model through loss function design, model learning, optimization algorithm step prediction and course learning; network training and regularization, the objective function includes prediction error and regularization term: In this process, the Adam optimizer is used and the learning rate is adaptively adjusted; Finally, the Koopman operator theory is used to enhance the nonlinear modeling capability: F*ψ(z k )=ψ(z k+1 ) Where K is the Koopman operator and ψ(·) is a nonlinear dimensionality-raising function, which is approximated by a neural network.

6. The model order reduction acceleration method for HIL simulation of an automobile thermal system model according to claim 1, characterized in that: The model integration simulation step comprises the following steps: S41: Integrate the reduced-order heat exchanger model into the complete thermal system model, replacing the original high-fidelity Simscape heat exchanger submodule; during the integration process, ensure that the mass flow of the interface is continuous, that is, the mass flow of the pump outlet is equal to the inlet flow of the reduced-order model, as shown in the following formula: in Reconstructed inlet flow rate for the reduced-order model; At the same time, energy continuity is ensured, that is, the front-end outlet enthalpy is equal to the reduced-order model inlet enthalpy, as shown in the following formula: in, is the inlet temperature of the heat exchanger coupling, is the mass flow rate; The relaxation factor method is used to enhance numerical stability. A relaxation factor α is introduced at the interface, and the coupling variables are updated step by step. The mass / energy residual is checked after each coupling step. If it exceeds the threshold, it is recalculated to achieve iterative correction, as shown in the following formula: S42: Real-time simulation optimization: During the real-time simulation optimization process, the goal is to achieve real-time calculation with a time step of 1 ms on the HIL hardware, which includes the following steps: First, the state equation is discretized into an explicit Euler form: z k+1 =z k +Δt·f θ (z k ,u k ) The time step Δt is strictly synchronized with the HIL hardware clock. When generating code, MATLAB Coder is used to convert the trained coupled system model into C code. Dynamic memory allocation is disabled and the array size is fixed. The weight matrix is ​​solidified into a pre-calculated lookup table to reduce the amount of online calculation. The timer interrupt service program of the real-time operating system is used to trigger the model calculation every 1ms to achieve hardware synchronization. If the single-step calculation times out, linear interpolation compensation is enabled, as shown in the following formula: With k+1 =2z k -With k-1 。

Images