New energy automobile thermal pool calibration method based on thermal pool physical neural network model

Abstract

The invention relates to a new energy automobile heat pool calibration method based on a physical neural network model, and the method comprises the steps: dividing a heat pool into a plurality of heat transfer units based on the heat transfer and heat storage distribution of the cross section of the heat pool, and deploying a temperature sensor in each heat transfer unit; installing the hot pool at an automobile exhaust pipe, starting an engine, heating the hot pool by using exhaust waste heat, and collecting data of each temperature sensor to construct a data set; building a hot pool physical neural network model, and training and optimizing the hot pool physical neural network model by using the data set; and deploying a plurality of temperature sensors in a real vehicle, and inputting data collected by the plurality of temperature sensors into the trained and optimized thermal pool physical neural network model to obtain a calibrated thermal pool temperature model. By introducing the equivalent heat capacity method and physical constraint of the phase change material, the nonlinear thermophysical property in the phase change process can be effectively processed, the precision and efficiency of thermal pool temperature calibration are remarkably improved, and the method is suitable for optimization of a thermal management system of a new energy automobile.

Description

Technical Field

[0001] The present invention relates to the field of vehicle technology, and in particular to a thermal pool calibration method for new energy vehicles based on a thermal pool physical neural network model. Background Art

[0002] Traditional heat pool temperature calibration methods rely primarily on experimental measurements and numerical simulations. While experimental methods offer high precision, they are expensive and difficult to cover all operating conditions. Numerical simulation methods (such as finite element analysis) can simulate complex heat conduction processes, but they are computationally intensive and time-consuming, making them difficult to meet real-time requirements. Physically Based Neural Networks (PINNs), as a method that combines physical constraints with data-driven approaches, have demonstrated significant advantages in solving complex physical field problems. By embedding control equations (such as the heat conduction equation) into the loss function of a neural network, PINNs can achieve high-precision modeling with minimal data support while ensuring the physical consistency of the solution.

[0003] However, existing research has largely focused on traditional heat conduction problems, with limited research on temperature calibration of heat pools containing phase change materials. The nonlinear thermophysical properties of phase change materials (such as the dynamic changes in equivalent heat capacity) and the temperature plateau effect during the phase change process further complicate modeling.

[0004] Therefore, developing a heat pool temperature calibration method based on physical neural network can effectively combine experimental data with physical laws to achieve high-precision and high-efficiency temperature distribution prediction, which has important theoretical significance and engineering application value. Summary of the Invention

[0005] The purpose of the present invention is to provide a new energy vehicle thermal pool calibration method based on a thermal pool physical neural network model. This method can address the shortcomings of the existing technology and use the thermal pool physical neural network model to calibrate the energy storage thermal pool of the new energy vehicle, thereby adjusting the thermal management strategy of the new energy vehicle to more efficiently utilize the vehicle's waste heat.

[0006] In order to achieve the above object, the technical solution of the present invention is as follows:

[0007] A new energy vehicle thermal pool calibration method based on a thermal pool physical neural network model comprises the following steps:

[0008] S1. Based on the heat transfer and heat storage distribution of the heat pool cross section, the heat pool is divided into multiple heat transfer units, and a temperature sensor is deployed in each heat transfer unit.

[0009] S2. Install the heat pool at the exhaust pipe of the car, start the engine, use the waste heat of the exhaust gas to heat the heat pool, and collect data from each temperature sensor to build a data set.

[0010] S3. Build a thermal pool physical neural network model and use the data set to train and optimize the thermal pool physical neural network model.

[0011] S4. Deploy several temperature sensors in the actual vehicle, input the data collected by the temperature sensors into the trained and optimized heat pool physical neural network model, and obtain a calibrated heat pool temperature model.

[0012] Furthermore, the loss function of the heat pool physical neural network model includes data error loss, physical equation residual loss, boundary condition loss and phase change constraint loss.

[0013] The loss function of the thermal pool physical neural network model is:

[0014] ;

[0015] in, is the data error loss, is the residual loss of the physical equation, is the boundary condition loss, is the phase change constraint loss, are all weight parameters.

[0016] Furthermore, the data error loss is:

[0017] ;

[0018] in, is the data error loss, is the amount of data for training, is the predicted temperature, is the measured temperature, is the space-time coordinate parameter, Represent the position coordinates in three-dimensional space, Represents the time parameter.

[0019] Furthermore, the residual loss of the physical equation is:

[0020] ;

[0021] in, is the residual loss of the physical equation, is the number of samples lost by the physical equation residual, is the residual, express The first of the samples samples.

[0022] Furthermore, the boundary condition loss is:

[0023] ;

[0024] in, is the boundary condition loss, is the number of boundary points, is the thermal conductivity, is the heat flux density, is the convective heat transfer coefficient, is the phase change material temperature, is the ambient temperature, is the convective heat loss.

[0025] Furthermore, the phase change constraint loss is:

[0026] ;

[0027] in, is the phase transition constraint loss, M is the number of data points in the phase transition interval, is the phase change material temperature, is the indicator function, is the phase transition temperature, is the phase transition temperature range.

[0028] Furthermore, the thermal pool physical neural network model includes an input layer, a hidden layer and an output layer; the hidden layer contains four layers, which are respectively set with 64 neurons, 128 neurons, 128 neurons and 64 neurons; the activation function of the hidden layer is the Tanh function.

[0029] Furthermore, in step S3, the data set includes a training set, an optimization set, and a validation set; and using the data set to train and optimize the thermal pool physical neural network model specifically includes:

[0030] S31. Initialize the weight of each connection in the hot pool physical neural network model.

[0031] S32. Input the training set data of the data set into the heat pool physical neural network model, perform propagation calculation through layer-by-layer linear transformation and nonlinear activation, and determine the output of the heat pool physical neural network model.

[0032] S33. Obtain the data error loss between the predicted temperature distribution data and the actual heat pool temperature data, and calculate the physical equation residual loss, phase change constraint loss, boundary condition loss, and determine the total loss function of the heat pool physical neural network model. .

[0033] S34. Starting from the output layer, calculate the gradient of the total loss function with respect to the weights and biases of each layer layer by layer, and adjust the network parameters according to the gradient to achieve reverse propagation of errors and parameter optimization.

[0034] S35. Use the optimization set data to evaluate the current model performance and automatically adjust the weight parameters of each sub-item loss in the total loss function based on the evaluation results. , complete single batch training.

[0035] S36. Repeat the above training and use the validation set to verify the trained hot pool physical neural network model until the result of the predicted total loss function is less than the preset convergence threshold or the number of iterations reaches the set value, thereby obtaining a trained hot pool physical neural network model.

[0036] Furthermore, the step S32 specifically includes the following steps:

[0037] S321. Normalize the input data of the thermal pool physical neural network model.

[0038] Step S321 is to eliminate the dimensional differences of different parameters (such as spatial coordinates, time, temperature, etc.), avoid model training failure due to large differences in numerical ranges, and ensure that the influence weight of each input feature on model training is balanced.

[0039] S322. Each layer of the heat pool physical neural network model performs linear transformation and nonlinear activation on the normalized input data in turn, and the last layer generates a predicted value to obtain the predicted heat pool temperature distribution result.

[0040] S323, for the thermal pool physical neural network model Layer, use the following formula to perform linear transformation calculation, and input data And the weight matrix of this layer Multiply and add the bias , get the inactive output :

[0041] ;

[0042] Among them, the first layer input data This is the original input data.

[0043] S324, based on the Tanh function, use the following formula to convert the inactivated output Convert to active output , to achieve nonlinear mapping:

[0044] ;

[0045] in, Represents the Tanh function.

[0046] S325, output the predicted temperature of the heat pool using the following formula :

[0047] ;

[0048] in, is the connection weight between the fourth hidden layer and the output layer, is the activation output of the fourth hidden layer, is the output layer bias.

[0049] Step S325 is implemented through a node of the output layer, which outputs a specific temperature prediction value and completes the mapping from the input parameters to the heat pool temperature.

[0050] Furthermore, the step S34 specifically includes the following steps:

[0051] S341. Calculate the data error loss gradient, and obtain the weight gradient and bias gradient for calculating the data error loss gradient.

[0052] S342. Calculate the physical equation residual loss gradient, and obtain the weight gradient and bias gradient of the physical equation residual loss gradient.

[0053] S343. Calculate the phase change constraint loss gradient, and obtain the weight gradient and bias gradient of the phase change constraint loss gradient.

[0054] S344. Calculate the boundary condition loss gradient and obtain the weight gradient and bias gradient of the boundary condition loss gradient.

[0055] S345. Obtain the total gradient of the output layer of the physical neural network model, and start from the last hidden layer and calculate the error term and gradient of each hidden layer layer by layer.

[0056] S346. Based on the error terms and gradients of each hidden layer, calculate the hidden layer gradients and weight gradients of the physical neural network model.

[0057] S347. Update the weight matrix and weights of the physical neural network model to complete the adjustment of the network parameters.

[0058] Starting from the output layer, the gradient of the loss function to the network parameters is calculated layer by layer, and an optimization algorithm (such as Adam) is used to update the parameters according to the gradient.

[0059] Furthermore, the step S341 specifically includes the following steps:

[0060] S341. Calculate the data error loss gradient.

[0061] S3411. Calculate the weight gradient of the data error loss gradient using the following formula:

[0062] ;

[0063] in, For data loss, is the connection weight between the fourth hidden layer and the output layer, for The corresponding weight parameters, is the amount of data for training, The model predicts The temperature of a point, The actual measurement The temperature of a point, is the space-time coordinate parameter, Represent the position coordinates in three-dimensional space, represents the time parameter, It is The activation output of the fourth hidden layer.

[0064] S3412. Calculate the bias gradient of the data error loss gradient using the following formula:

[0065] ;

[0066] in, is the output layer bias.

[0067] Furthermore, the step S342 specifically includes the following steps:

[0068] S3421. Calculate the weighted gradient of the physical equation residual using the following formula:

[0069] ;

[0070] in, is the physical residual loss, for The corresponding weight parameters, It is The residual of the sample, is the number of samples lost for the physical equation residual.

[0071] S3422. Calculate the bias gradient of the physical equation residual using the following formula:

[0072] .

[0073] Furthermore, the step S343 specifically includes the following steps:

[0074] S3431. Calculate the weight gradient of the phase change constraint loss gradient using the following formula:

[0075] ;

[0076] in, is the phase change constraint loss, for The corresponding weight parameters, Indicates the phase change temperature of the phase change material; is the phase transition temperature range, is the number of data points in the phase transition interval.

[0077] S3432. Calculate the bias gradient of the phase change constraint loss gradient using the following formula:

[0078] ;

[0079] in, is the output layer bias.

[0080] Furthermore, the step S344 specifically includes the following steps:

[0081] S3441. Calculate the weight gradient of the boundary condition loss gradient using the following formula;

[0082] ;

[0083] in, is the boundary condition loss, for The corresponding weight parameters, is the number of boundary points, is the thermal conductivity, is the heat flux density, is the convective heat transfer coefficient, is the phase change material temperature, is the ambient temperature, represents convective heat loss.

[0084] S3442. Calculate the bias gradient of the boundary condition loss gradient using the following formula:

[0085] .

[0086] Furthermore, the step S345 specifically includes the following steps:

[0087] S3451. Calculate the total gradient of the output layer of the physical neural network model using the following formula:

[0088] ;

[0089] ;

[0090] in, is the total loss function, is the connection weight between the fourth hidden layer and the output layer, is the output layer bias, is the data error loss, is the physical residual loss, is the phase change constraint loss, is the boundary condition loss.

[0091] S3452. Starting from the last hidden layer, use the following formula to calculate the error term and gradient of each hidden layer layer by layer:

[0092] ;

[0093] in, Represents the loss function for the Layer inactive output The gradient, represents the transpose of the next layer weight matrix, Represents the loss function for the Layer inactive output The gradient, It represents element-by-element multiplication, and tanh is the hyperbolic tangent function.

[0094] Furthermore, the step S346 specifically includes the following steps:

[0095] S3461. Based on the error term, the hidden layer gradient of the physical neural network model is calculated using the following formula:

[0096] ;

[0097] in, is the total loss function, For the The weight matrix of the layer, For the The transposed matrix of the layer activation outputs.

[0098] S3462. Calculate the weight gradient of the physical neural network model using the following formula:

[0099] ;

[0100] in, is the total loss function, is the data error loss, is the residual loss of the physical equation, is the boundary condition loss, is the phase change constraint loss, They are 、 、 、 The weight parameter of .

[0101] The step S347 specifically includes the following steps:

[0102] The weight matrix and weights of the physical neural network model are updated using the following formula:

[0103] ;

[0104] ;

[0105] ;

[0106] ;

[0107] ;

[0108] in, For the The weight matrix of the layer version, For the The weight matrix of the layer version, is the learning rate, is the weight parameter No. version, is the weight parameter No. version, is the weight parameter No. version, is the weight parameter No. version, is the weight parameter No. version, is the weight parameter No. version, is the weight parameter No. version, is the weight parameter No. version.

[0109] Compared with the prior art, the advantages of the present invention are:

[0110] Compared with existing technologies, the present invention, by incorporating the equivalent heat capacity method and physical constraints of phase change materials, effectively addresses the nonlinear thermal properties of the phase change process, significantly improving the accuracy and efficiency of heat pool temperature calibration. This allows for high-precision modeling with minimal data, making it suitable for optimizing thermal management systems for new energy vehicles. By constructing a heat pool physical information neural network model and embedding the governing equations into the neural network loss function, the present invention predicts the temperature distribution of the heat pool installed on the vehicle's exhaust pipe. This allows for an estimate of the actual available heat, enabling the vehicle's thermal management system to more accurately utilize exhaust waste heat and phase change material heat storage. For example, when the battery is cold, heat released by the phase change material can be quickly utilized for preheating to improve battery performance. Furthermore, waste heat can be prioritized for interior heating to reduce energy consumption on the power battery, thereby improving the overall performance of new energy vehicles. After neural network training, the present invention only requires temperature data from a few key locations to calibrate the heat pool temperature field, significantly simplifying the training process. Furthermore, the temperature field can be calibrated and the neural network model can be corrected in real time via a cloud server, ensuring the accuracy of heat pool temperature estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0111] Figure 1 A schematic diagram of the process of the thermal pool calibration method for new energy vehicles based on the thermal pool physical neural network provided by the present invention;

[0112] Figure 2 is a schematic diagram of the structure of the heat pool;

[0113] Figure 3 This is a structural breakdown diagram of the heat pool;

[0114] Figure 4 This is a schematic diagram of the installation cross-section position of the temperature sensor;

[0115] Figure 5 This is the exploded view of the temperature sensor installation section;

[0116] Figure 6 Schematic diagram of the heat transfer unit division and sensor installation position in section AA;

[0117] Figure 7 Schematic diagram of the heat transfer unit division and sensor installation position in section BB;

[0118] Figure 8 Schematic diagram of the heat transfer unit division and sensor installation position in section CC;

[0119] Figure 9 Schematic diagram of the neural network model constructed in the present invention.

[0120] in:

[0121] 1. First air chamber, 2. Air distribution pipe, 3. Heat conducting plate, 4. Second air chamber, 5. Bellows, 6. Bellows inlet, 7. Round tube, 8. Internal cavity of heat pool shell, 9. Round tube inlet, 10. Round tube outlet, 11. Bellows outlet, 12. Exhaust gas outlet pipe, 13. Exhaust gas inlet pipe. DETAILED DESCRIPTION

[0122] In order to provide a further understanding and appreciation of the structural features and effects achieved by the present invention, a detailed description is provided with reference to preferred embodiments and accompanying drawings as follows:

[0123] like Figure 1 The thermal pool calibration method for new energy vehicles based on the thermal pool physical neural network model shown includes the following steps:

[0124] S1. Based on the heat transfer and storage distribution across the heat pool cross section, the heat pool is divided into multiple heat transfer units, with a temperature sensor deployed in each unit. Based on the heat transfer and storage functions, the heat transfer portion consists of two parts: the bronchial-hot oil pipe nested within the heat transfer pipe and the phase change material within the hot oil pipe. Heat storage is the phase change material portion. To account for the uneven temperature distribution of the phase change material, the phase change material is divided into multiple heat transfer units across the cross section, with the divisions closer to the heat transfer section being more dense. Three cross sections are defined along the longitudinal direction of the heat pool.

[0125] S2. Install the heat pool at the exhaust pipe of the car, start the engine, use the waste heat of the exhaust gas to heat the heat pool, and collect data from each temperature sensor to build a data set.

[0126] S3. Build a thermal pool physical neural network model and use the data set to train and optimize the thermal pool physical neural network model.

[0127] S4. Deploy several temperature sensors in the actual vehicle, input the data collected by the temperature sensors into the trained and optimized heat pool physical neural network model, and obtain a calibrated heat pool temperature model.

[0128] The heat pool temperature model calibrated in step S4 has the following functions:

[0129] (1) Real-time prediction of temperature distribution inside the thermal pool

[0130] Only a small number of key temperature sensors need to be deployed in the actual vehicle. By inputting the collected data into the model, the overall temperature distribution of the heat pool can be quickly obtained. This eliminates the need for a full set of sensors, reducing hardware costs and on-board computing resource consumption.

[0131] (2) Optimizing the thermal management system of new energy vehicles

[0132] By precisely understanding the heat pool's temperature distribution and heat storage status, efficient utilization of vehicle waste heat is achieved. Combined with the heat pool's structural design (a bellows outputs high-temperature thermal oil, and a circular tube outputs phase-change material heat storage), the heat pool temperature model assists in determining the heat distribution of different heat sources, guiding the rational allocation of waste heat for scenarios such as in-vehicle heating and battery preheating. When the battery is cold, the heat released by the phase-change material is quickly utilized to preheat the battery, improving battery performance. When heating the vehicle, exhaust waste heat is prioritized, reducing energy consumption on the power battery and thereby improving the overall performance of new energy vehicles.

[0133] (3) Support dynamic model update and optimization

[0134] When the model prediction error in actual vehicle applications is large (such as a temperature difference of more than 5°C), the neural network model can be retrained and updated through the cloud based on real-time data to ensure the accuracy of the heat pool temperature estimation and continuously adapt to changes in actual operating conditions.

[0135] like Figure 2 and Figure 3 As shown, the heat pool includes a heat pool shell, a first air chamber 1 and a second air chamber 4. The first air chamber 1 and the second air chamber 4 are located at both ends of the heat pool shell and are connected by a number of parallel air distribution pipes 2 to form an exhaust gas circulation channel. The heat pool shell is covered with a skin to form a closed cavity. The internal cavity 8 of the heat pool shell is filled with a phase change material. The heat pool structural design described in the present invention can enable the heat pool to absorb waste heat in the exhaust gas, and release the heat stored in the phase change material and the waste heat in the exhaust gas as needed, thereby realizing integrated heat storage and heat exchange.

[0136] Furthermore, the first air chamber 1 and the second air chamber 4 are connected by a plurality of air distribution pipes 2. The air distribution pipe 2 runs through the internal cavity 8 of the heat pool, and the two ends of the air distribution pipe 2 are respectively connected to the second air chamber 4 and the first air chamber 1, and the bellows 5 are nested inside the air distribution pipe 2. The air distribution pipe 2 is located in the internal cavity 8 of the heat pool shell, and a plurality of air distribution pipes 2 are arranged in parallel. In this embodiment, the number of the air distribution pipes 2 is 7. The air distribution pipe 2 is used to guide the exhaust gas to the inside of the heat pool, and at the same time, heat exchange is performed with the bellows 5 through the pipe wall of the air distribution pipe.

[0137] Furthermore, the heat pool also includes a bellows 5, and the bellows 5 includes three sections connected in sequence, the first section is located in the second air chamber 4, the second section passes through each air distribution pipe 2 in sequence, and the third section is located in the first air chamber 1. Preferably, the first section of the bellows 5 is spirally coiled in the second air chamber 4, the second section is serpentinely inserted in the air distribution pipe 2, and the third section is spirally coiled in the first air chamber 1. The second air chamber 4 is provided with a bellows inlet 6 connected to the first section of the bellows 5, and the first air chamber 1 is provided with a bellows outlet 11 connected to the second section of the bellows 5. The bellows 5 connected to the bellows inlet 6 starts from the second air chamber 4, passes through each air distribution pipe 2 in the internal cavity 8 of the heat pool in sequence, and finally reaches the first air chamber 1 and is connected to the bellows outlet 11. The two ends of the bellows 5 are coiled in the first air chamber 1 and the second air chamber 4, and the middle part is distributed in a serpentine shape in each air distribution pipe in the heat pool shell. This design can increase the contact area between the bellows and the exhaust gas. The bellows 5 is filled with heat transfer oil. The heat transfer oil enters the bellows 5 from the bellows inlet 6, passes through the middle part of the bellows 5, and is finally discharged from the bellows outlet 11. The heat transfer oil in the bellows 5 is used to absorb the waste heat in the second air chamber 4, the first air chamber 1 and the air distribution pipe 2. The bellows 5 increases the contact area with the exhaust gas through the coiling design. The heat transfer oil inside the bellows 5 absorbs the waste heat in the air distribution pipe 2 and the air chamber and outputs it.

[0138] Furthermore, a circular tube 7 is provided in the heat pool shell, and a heat conducting plate 3 is provided on the surface of the circular tube 7. The heat conducting plate 3 is used to increase the contact area between the phase change material and the circular tube 7, thereby improving the heat exchange efficiency. Preferably, there are multiple circular tubes 7, and the multiple circular tubes 7 are staggered with the multiple gas distribution pipes 2. This design is to achieve uniformity in heat exchange. The multiple circular tubes 7 are distributed in the phase change material, and the circular tubes 7 are filled with heat transfer oil. The heat transfer oil enters the circular tube 7 from the circular tube inlet 9 of the circular tube 7 and flows out from the circular tube outlet 10. It is used to absorb the heat stored in the phase change material in the heat pool shell and release it to other systems.

[0139] In this embodiment, when the heat pool is working, the exhaust gas intake pipe 13 is connected to the exhaust pipe of the automobile, and the exhaust gas enters the second air chamber 4 from the exhaust gas intake pipe 13, and after heat exchange through the air distribution pipe 2, enters the first air chamber 1, and finally is discharged into the atmosphere from the exhaust gas outlet pipe 12. On the one hand, the waste heat of the exhaust gas is exchanged with the heat transfer oil in the bellows 5 through the wall of the air distribution pipe 2, and on the other hand, it is used to heat the phase change material to melt it and store heat; when the system needs heat, the phase change material solidifies and releases heat, which is released through the heat transfer oil in the circular pipe 7, realizing an integrated cycle of heat absorption, heat storage and heat release. The heat transfer oil in the bellows 5 can fully absorb the heat of the exhaust gas through the coiled design, and the heat transfer oil in the circular pipe 7 releases the stored heat when the phase change material solidifies, realizing the storage and on-demand release of waste heat. The present invention combines the equivalent heat capacity design of the phase change material to improve the thermal management efficiency. Preferably, the high-temperature heat-conducting oil output by the bellows 5 can be directly used for heating in the vehicle, and the heat output by the circular tube 7 can be used for battery preheating, thereby realizing graded utilization of waste heat.

[0140] The following describes the division of heat transfer units and the location of temperature sensors in this embodiment in conjunction with the accompanying drawings. Figures 4 to 8 As shown, considering the temperature difference between the front and rear ends of the heat pool entity, a total of Figure 4 、 Figure 5 In the three cross sections. Figure 4 and Figure 5 In the figure, AA, BB, and CC represent the cross-sectional locations of the temperature sensors. Each cross-sectional heat transfer unit is divided as follows: In one cross-sectional area, the gas distribution pipe 2 and the corrugated pipe 5 nested in the gas distribution pipe 2 are divided into the same heat transfer unit, for a total of 7 heat transfer units, such as Figures 6 to 8 The hot oil pipe (hot oil pipe i.e. circular pipe 7) inside the phase change material in the internal cavity 8 of the heat pool shell is divided into 12 heat transfer units, i.e. the cross section of the 12 circular pipes 7 in the cross section of the heat pool is divided into 12 heat transfer units, as shown in the orange part. Figures 6 to 8 The phase change material in the inner cavity 8 of the heat pool shell is divided into 18 heat transfer units according to its relative position with the gas distribution pipe 2 and the circular pipe 7, as shown in the blue part. Figures 6 to 8 The yellow part is shown in the figure. There are slight differences between the three sections, which have no effect on the implementation of the calibration method of the present invention. In this embodiment, when the constructed neural network model is trained, the heat transfer units are divided according to the above method. Each heat transfer unit is equipped with a temperature sensor, each section is equipped with 37 temperature sensors, and the three cross sections have a total of 111 temperature sensors. Figures 6 to 8 In the figure, you can see the temperature sensors and their labels. The number and placement of temperature sensors should be tailored to the specific structure, with a maximum number placed close to the bronchial tubes and thermal oil pipes. In this embodiment, a temperature sensor is placed in each bronchial tube and bellows nested heat transfer unit, each heat pool internal circular tube heat transfer unit, and each phase change material heat transfer unit.

[0141] The heat pool inputs heat from the vehicle's exhaust pipe. A phase-change material is embedded within the heat pool for heat storage. Oil pipes are placed within the phase-change material for heat absorption and heat transfer, achieving integrated heat storage and release. The heat pool has an air chamber in front and behind, with seven air pipes located between the two chambers. Exhaust gas flows through the phase-change material within the heat pool, the air pipes, and the hot oil pipes nested within the air chambers, achieving integrated heat storage and heat exchange. The sensor installation layout is divided into three equal sections longitudinally and arranged transversely based on the relative positions of the pipes. Each cross-section of the sensor is divided into three equal sections, with 37 temperature sensors placed within the exhaust pipe, the phase-change material, and the hot oil pipe within the phase-change material, for a total of 111 temperature sensors. The heat transfer units are divided based on the cross-section of the heat pool. The exhaust pipe and the nested hot oil pipe are grouped into the same heat transfer unit, for a total of 7 units. The hot oil pipe within the phase-change material is divided into a separate heat transfer unit, for a total of 12 units. The phase-change material is divided into 18 units based on its relative position to the exhaust pipe and the oil pipe. Factors affecting the heat transfer coefficient in each heat transfer unit include the material, wall thickness, and shape of the exhaust pipe and hot oil pipe, the heat conducting plate arranged in the hot oil pipe within the phase change material, and the gap between the phase change material and the exhaust pipe and hot oil pipe.

[0142] In order to realize the calibration of the heat pool of new energy vehicles based on the physical neural heat pool model, the present invention constructs the following Figure 9 The thermal pool physical neural network model shown in FIG, the neural network model includes an input layer, a hidden layer and an output layer. The input layer contains spatial coordinates , time, exhaust gas temperature , exhaust gas flow rate , mass flow , ambient temperature and phase transition temperature , with a total of 9 input nodes. The hidden layer contains 4 layers, with the number of neurons in the 4 layers being 64, 128, 128, and 64, respectively. The output layer has 1 node, outputs temperature T, and uses a linear activation function; the activation function of this neural network is a nonlinear function, such as the Tanh function.

[0143] The physical constraints of the heat pool physical neural network model include the energy conservation equation, phase change temperature constraints, and boundary conditions. The constraints of the heat pool physical neural network model, which ensures compliance with physical laws and boundary conditions, include the energy conservation law, phase change heat storage and release, exhaust gas heat source input, hot oil pipeline heat output, and heat dissipation at the heat pool boundary. The total loss function of the heat pool physical neural network model includes the physical equation residual loss, phase change constraint loss, boundary condition loss, and data error loss.

[0144] The following is an introduction to the calculation process of each loss in the total loss function of the hot pool physical neural network model:

[0145] (1) Physical equation residual loss

[0146] (11) Incorporating latent heat into heat capacity, the energy conservation equation based on the equivalent heat capacity method is established using the following formula:

[0147] ;

[0148] in, represents the density of the phase change material, is the equivalent heat capacity of the phase change material, is the temperature of the phase change material, For time, is the nabla operator, , is the thermal conductivity of the phase change material, is the heat input to the heat pool, is the heat output of the thermal pool, is the rate of change of thermal energy inside the heat pool, represents the heat conduction term.

[0149] (12) Calculate the equivalent heat capacity of the phase change material using the following formula: :

[0150] ;

[0151] in, represents the specific heat capacity in the solid state, represents the temperature of the phase change material, is the phase transition temperature, is the phase transition temperature range, is latent heat, represents the specific heat capacity of the liquid phase.

[0152] (13) Calculate the heat input of the heat pool using the following formula: :

[0153] ;

[0154] in, is the exhaust gas mass flow rate; is the specific heat capacity of the exhaust gas; is the exhaust gas temperature; For reference temperature, set to ambient temperature.

[0155] (14) Calculate the heat output of the heat pool using the following formula: :

[0156] ;

[0157] in, is the convective heat transfer coefficient; is the surface area of ​​the heat pool; is the surface temperature of the heat pool; is the ambient temperature.

[0158] (15) The residual based on the energy conservation equation is calculated using the following formula:

[0159] ;

[0160] (16) Calculate the residual loss of the physical equation using the following formula: :

[0161] ;

[0162] Where N is the number of samples, For the The residuals of the samples.

[0163] (2) Phase change constraint loss

[0164] When obtaining the phase change temperature constraint, it is important to note that within the phase change range, the temperature change rate of the phase change material approaches zero. The phase change constraint loss is calculated using the following formula: :

[0165] ;

[0166] in, is the phase transition constraint loss, M is the number of data points in the phase transition interval, is the temperature of the phase change material, is the phase transition temperature, is the phase transition temperature range, is the indicator function, which is used to ensure that only when This item is only counted as loss when it is within the phase change range.

[0167] (3) Boundary condition loss

[0168] For heat dissipation at the heat pool boundary, this embodiment uses Newton's law of cooling to perform boundary constraints. The boundary condition loss is calculated using the following formula: :

[0169] ;

[0170] in, is the boundary condition loss, is the number of boundary points, is the thermal conductivity, is the heat flux density, is the convective heat transfer coefficient, is the phase change material temperature, is the ambient temperature, is the convective heat loss.

[0171] (4) Data error loss

[0172] ;

[0173] in, is the amount of data for training, is the predicted temperature, is the measured temperature, is the space-time coordinate parameter, Represent the position coordinates in three-dimensional space, Represents the time parameter.

[0174] Based on the above losses, the loss function of the thermal pool physical neural network model is determined as follows:

[0175] ;

[0176] in, Weight parameters are used to balance data loss and physical loss. During training, Adam optimization algorithm or L-BFGS algorithm is used to optimize the parameters. Optimization is performed to make the trained physical neural network more accurate, and by optimizing and adjusting the weight parameters, it is possible to avoid training failure due to a certain loss function dominating and not converging.

[0177] In this embodiment, after the physical neural network model of the heat pool is constructed, data collection and processing are performed first. First, the engine is started, the heat pool is heated by the waste heat of the exhaust gas, and the temperature data of each sensor is recorded. Secondly, the sampling frequency is increased within the phase change interval to ensure the accuracy of the temperature platform effect. Then, the abnormal data in the data set is eliminated to improve the accuracy of the constructed neural network. Finally, the collected data is divided into a training set, a validation set, and a test set, with the proportions of each set being 70%, 15%, and 15%, respectively. When the waste heat of the exhaust gas is used to heat the heat pool, the data of each sensor is recorded, and the data is divided into a training set, a test set, and a validation set for the construction, optimization, and verification of the neural network respectively; the sampling frequency is increased within the phase change interval of the phase change material to more accurately predict the temperature distribution of the heat pool.

[0178] Next, the constructed thermal pool physical neural network model is trained and optimized. Initialize the weights and biases of the thermal pool physical neural network model. The weights are initialized using Xavier and the biases are initialized to zero. Input the training set into the neural network to obtain the predicted temperature distribution. Calculate the data error loss, physical equation residual loss, boundary condition loss, and phase change constraint loss to determine the loss function of the thermal pool physical neural network model. Update the network parameters through backpropagation to optimize the loss function. Use the Adam optimizer to update the parameters. The initial learning rate is set to 0.001. The number of training data input each time is 3300 and the number of iterations is set to 50,000. During the training process, the weight parameters Dynamically adjust based on the relative size of each loss term. Initial values ​​are set to 1, 0.1, 1.0, and 0.5, and updated according to the loss ratio after each round of training to ensure a balance between data error and physical constraints. Calculate the total loss on the validation set every 5 rounds. If the total loss does not decrease for 3 consecutive times, stop training. Stop training when the result of the total loss function is less than the preset convergence threshold or the number of iterations reaches the preset number of steps. Calculate the MAE, RMSE, and physical residual of the temperature prediction to evaluate the performance of the neural network model. If the model performance does not meet expectations, retrain and update the model.

[0179] To verify the reliability of the heat pool physical neural network model under actual operating conditions, this example was applied to a real vehicle. A small number of sensors were installed in the real vehicle, and temperature sensors were used to collect real-time temperature data. This real-time data was fed into a trained neural network model, which then output a predicted heat pool temperature distribution model. This model then used the model to predict the temperature distribution within the heat pool in real time. The predicted temperatures were compared with the temperatures collected by the actual temperature sensors (this data was not used in neural network training and was reserved for determining the magnitude of neural network error). If the temperature difference exceeded 5°C, the error was considered significant. If the model prediction error was significant, the cloud-based neural network model was retrained and updated. Only a few key sensors were installed in the real vehicle (rather than the 111 sensors used in the laboratory), reducing hardware costs (sensor procurement and wiring) and on-board computing resource consumption, making it suitable for the engineering requirements of mass-produced vehicles. Furthermore, leveraging the generalization capabilities of the physical neural network, the overall temperature can be estimated using a small amount of sensor data, avoiding the communication bandwidth pressure associated with full data transmission and ensuring real-time system performance.

[0180] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A new energy vehicle thermal pool calibration method based on a thermal pool physical neural network model, characterized in that: The method comprises the following steps: S1. Based on the heat transfer and heat storage distribution of the heat pool cross section, the heat pool is divided into multiple heat transfer units, and a temperature sensor is deployed in each heat transfer unit; S2. Install the heat pool at the exhaust pipe of the car, start the engine, use the exhaust waste heat to heat the heat pool, and collect data from each temperature sensor to build a data set; S3. Build a thermal pool physical neural network model and use the data set to train and optimize the thermal pool physical neural network model; S4. Deploy several temperature sensors in the actual vehicle, input the data collected by the temperature sensors into the trained and optimized heat pool physical neural network model, and obtain a calibrated heat pool temperature model.

2. The method for calibrating the thermal pool of a new energy vehicle based on the thermal pool physical neural network model according to claim 1 is characterized in that: The loss function of the heat pool physical neural network model includes data error loss, physical equation residual loss, boundary condition loss and phase change constraint loss; The loss function of the thermal pool physical neural network model is: ； in, is the data error loss, is the residual loss of the physical equation, is the boundary condition loss, is the phase change constraint loss, are all weight parameters.

3. The new energy vehicle thermal pool calibration method based on the thermal pool physical neural network model according to claim 2 is characterized in that: The data error loss is: ； in, is the data error loss, is the amount of data for training, is the predicted temperature, is the measured temperature, is the space-time coordinate parameter, Represent the position coordinates in three-dimensional space, Represents the time parameter.

4. The method for calibrating the thermal pool of a new energy vehicle based on the thermal pool physical neural network model according to claim 2 is characterized in that: The physical equation residual loss is: ； in, is the residual loss of the physical equation, is the number of samples lost by the physical equation residual, is the residual, express The first of the samples samples.

5. The method for calibrating the thermal pool of a new energy vehicle based on the thermal pool physical neural network model according to claim 2 is characterized in that: The boundary condition loss is: ； in, is the boundary condition loss, is the number of boundary points, is the thermal conductivity, is the heat flux density, is the convective heat transfer coefficient, is the phase change material temperature, is the ambient temperature, is the convective heat loss.

6. The method for calibrating the thermal pool of a new energy vehicle based on the thermal pool physical neural network model according to claim 2 is characterized in that: The phase change constraint loss is: ； in, is the phase transition constraint loss, M is the number of data points in the phase transition interval, is the phase change material temperature, is the indicator function, is the phase change temperature of the phase change material, is the phase transition temperature range.

7. The method for calibrating the thermal pool of a new energy vehicle based on a thermal pool physical neural network model according to claim 1, characterized in that: The thermal pool physical neural network model includes an input layer, a hidden layer and an output layer; the hidden layer contains four layers, which are respectively set with 64 neurons, 128 neurons, 128 neurons and 64 neurons; the activation function of the hidden layer is a Tanh function.

8. The method for calibrating the thermal pool of a new energy vehicle based on a thermal pool physical neural network model according to claim 1, characterized in that: In step S3, the data set includes a training set, an optimization set, and a validation set; and using the data set to train and optimize the thermal pool physical neural network model specifically includes: S31, initializing the weight of each connection in the hot pool physical neural network model; S32, inputting the training set data of the data set into the heat pool physical neural network model, performing propagation calculation through layer-by-layer linear transformation and nonlinear activation, and determining the output of the heat pool physical neural network model; S33. Obtain the data error loss between the predicted temperature distribution data and the actual heat pool temperature data, and calculate the physical equation residual loss, phase change constraint loss, boundary condition loss, and determine the total loss function of the heat pool physical neural network model. ; S34. Starting from the output layer, calculate the gradient of the total loss function with respect to the weights and biases of each layer layer by layer, and adjust the network parameters according to the gradient; S35. Use the optimization set data to evaluate the current model performance and automatically adjust the weight parameters of each sub-item loss in the total loss function based on the evaluation results. , complete single batch training; S36. Repeat the above training and use the validation set to verify the trained hot pool physical neural network model until the result of the predicted total loss function is less than the preset convergence threshold or the number of iterations reaches the set value, thereby obtaining a trained hot pool physical neural network model.

9. The method for calibrating the thermal pool of a new energy vehicle based on the thermal pool physical neural network model according to claim 8, characterized in that: The step S32 specifically includes the following steps: S321, normalizing the input data of the thermal pool physical neural network model; S322, each layer of the heat pool physical neural network model sequentially performs linear transformation and nonlinear activation on the normalized input data, and the last layer generates a predicted value to obtain a predicted heat pool temperature distribution result; S323, for the thermal pool physical neural network model Layer, use the following formula to perform linear transformation calculation to get the unactivated output : ； Among them, the first layer input data That is the original input data; S324, based on the Tanh function, use the following formula to convert the inactivated output Convert to active output , to achieve nonlinear mapping: ； in, represents the Tanh function; S325, output the predicted temperature of the heat pool using the following formula : ； in, is the connection weight between the fourth hidden layer and the output layer, is the activation output of the fourth hidden layer, is the output layer bias.

10. The method for calibrating the thermal pool of a new energy vehicle based on the thermal pool physical neural network model according to claim 9 is characterized in that: The step S34 specifically includes the following steps: S341. Calculate the data error loss gradient, and obtain the weight gradient and bias gradient for calculating the data error loss gradient; S342. Calculate the physical equation residual loss gradient, and obtain the weight gradient and bias gradient of the physical equation residual loss gradient; S343. Calculate the phase change constraint loss gradient, and obtain the weight gradient and bias gradient of the phase change constraint loss gradient; S344. Calculate the boundary condition loss gradient and obtain the weight gradient and bias gradient of the boundary condition loss gradient; S345, obtaining the total gradient of the output layer of the physical neural network model, and starting from the last hidden layer, calculating the error term and gradient of each hidden layer layer by layer; S346. Calculate the hidden layer gradient and weight gradient of the physical neural network model based on the error terms and gradients of each hidden layer; S347. Update the weight matrix and weights of the physical neural network model.

Images