Temperature prediction method and device and electronic equipment

Abstract

The invention discloses a temperature prediction method and apparatus, and an electronic device. The method comprises the steps of obtaining a first feature vector set of a target passenger compartment; performing time sequence feature extraction on temperature influence parameters included in the first feature vector set by using a long short-term memory encoder, and outputting an initial hidden state sequence; the initial hidden state sequence comprises a time dependency relationship and a dynamic change mode of the temperature influence parameters; performing global relation modeling on the initial hidden state sequence by using a self-attention encoder, and outputting a target hidden state sequence; the target hidden state sequence comprises the long-acting influence of the temperature influence parameter on the temperature; and performing temperature prediction based on the target hidden state sequence by using a long short-term memory decoder, and outputting a first predicted temperature curve corresponding to the target passenger compartment. Therefore, the temperature prediction can be realized by combining the advantages of the LSTM in the aspect of capturing the long-term dependency relationship of the time sequence data and the powerful capability of the Transform in the aspects of parallel processing and global information capturing, and the effectiveness of the temperature prediction can be improved.

Description

Technical Field

[0001] This application belongs to the field of temperature prediction technology, and in particular relates to temperature prediction methods, devices and electronic equipment. Background Technology

[0002] Currently, the primary traditional method used in crew cabin temperature prediction is simulation based on physical models. However, while traditional physical models provide theoretical guidance based on thermodynamic principles, their assumptions are often overly simplified, making it difficult to accurately capture dynamic changes under complex operating conditions. In particular, the prediction accuracy of traditional physical models drops significantly under extreme conditions. Therefore, the reliability of existing crew cabin temperature prediction methods is low. Summary of the Invention

[0003] This application provides a temperature prediction method, apparatus, and electronic device that can improve the reliability of model prediction of the temperature of the crew cabin.

[0004] In a first aspect, embodiments of this application provide a temperature prediction method applied to an electronic device, wherein the electronic device is equipped with a target deep learning model; the target deep learning model includes a cascaded long short-term memory encoder, a self-attention encoder, and a long short-term memory decoder; the method includes: Obtain a first feature vector set of the target passenger cabin; wherein, the first feature vector set includes feature vectors corresponding to multiple time steps, and each feature vector includes the temperature influence parameter of the target passenger cabin at the corresponding time step; The first feature vector set is input into the long short-term memory encoder, and the long short-term memory encoder is used to extract time series features from the temperature influence parameters included in the first feature vector set, and outputs an initial hidden state sequence; the initial hidden state sequence contains the time dependence and dynamic change pattern of the temperature influence parameters. The initial hidden state sequence is input into the self-attention encoder, and the self-attention encoder is used to perform global relation modeling on the initial hidden state sequence to output the target hidden state sequence; the target hidden state sequence contains the long-term effect of the temperature influence parameter on temperature; The target hidden state sequence is input into the long short-term memory decoder, and the long short-term memory decoder is used to predict the temperature based on the target hidden state sequence, and outputs the first predicted temperature curve corresponding to the target crew cabin.

[0005] In some embodiments, training the target deep learning model includes: acquiring a first training set; wherein the first training set includes experimental temperature training samples obtained based on experiments and simulated temperature training samples obtained based on simulations; each temperature training sample includes a second feature vector set of a crew cabin and a temperature curve label corresponding to the second feature vector set; and training the target deep learning model using the first training set.

[0006] In some embodiments, the acquisition of the simulated temperature training samples includes: using a crew cabin thermodynamic model to simulate and generate crew cabin temperature curves for each operating condition; wherein the crew cabin thermodynamic model is constructed using a one-dimensional system simulation model and a three-dimensional computational fluid dynamics model; generating simulated temperature training samples corresponding to each operating condition based on each operating condition and the crew cabin temperature curves for each operating condition; wherein the feature vector set of each simulated temperature training sample includes temperature influence parameters for multiple time steps of the corresponding operating condition, and the temperature curve label of each simulated temperature training sample is the crew cabin temperature curve of the corresponding operating condition.

[0007] In some embodiments, the above-described method of using the first training set to train the target deep learning model includes: inputting a temperature training sample from the first training set into the target deep learning model; using the target deep learning model to predict the temperature of a second feature vector set included in the temperature training sample, and outputting a second predicted temperature curve; calculating the mean square error between the second predicted temperature curve and the temperature curve labels of the temperature training sample using a first loss function; calculating the product of a target squared value and a preset hyperparameter using a second loss function; the target squared value being the square of the difference between the temperature at the final moment on the second predicted temperature curve and the ideal temperature at the final moment; determining the target loss function value of the target deep learning model based on the mean square error and the product; and adjusting the model parameters of the target deep learning model if the target loss function value does not meet the training stopping condition, and returning to the step of inputting a temperature training sample from the first training set into the target deep learning model until the target loss function value meets the training stopping condition.

[0008] In some embodiments, determining the target loss function value of the target deep learning model based on the mean squared error and the product includes: determining a first loss function value of the target deep learning model based on the mean squared error and the product; and performing a weighted summation of the first loss function value and the second loss function value to obtain the target loss function value; wherein the second loss function value is determined by the auxiliary quantity curve labels included in the auxiliary quantity training samples and the predicted auxiliary quantity curve obtained by the target deep learning model based on the auxiliary quantity influence parameters included in the auxiliary quantity training samples, and the auxiliary quantity is a physical quantity that has an impact on the temperature of the crew cabin.

[0009] In some embodiments, the above-mentioned passenger compartment is the passenger compartment of a vehicle, and the temperature training sample further includes vehicle model data; after inputting the temperature training sample into the target deep learning model, the method further includes: embedding the vehicle model embedding vector corresponding to the vehicle model data into the target deep learning model; the vehicle model embedding vector is used to represent the vehicle model.

[0010] In some embodiments, after outputting the second predicted temperature curve corresponding to the temperature training sample, and before calculating the mean square error between the second predicted temperature curve and the temperature curve labels included in the temperature training sample using the first loss function, the method further includes: smoothing the second predicted temperature curve if the second predicted temperature curve does not satisfy a preset monotonicity.

[0011] In some embodiments, the target passenger compartment is the passenger compartment of a target vehicle; before inputting the first feature vector set into the long short-term memory encoder, the method further includes: obtaining a second training set, the second training set including multiple vehicle model data associated with the model of the target vehicle; and adjusting the target deep learning model using the second training set.

[0012] Secondly, embodiments of this application also provide a temperature prediction device applied to an electronic device, wherein the electronic device is equipped with a target deep learning model; the target deep learning model includes a long short-term memory encoder, a self-attention encoder, and a long short-term memory decoder connected in series; the device includes: The first acquisition module is used to acquire a first feature vector set of the target crew cabin; wherein, the first feature vector set includes feature vectors corresponding to multiple time steps, and each feature vector includes the temperature influence parameter of the target crew cabin at the corresponding time step; The second acquisition module is used to input the first feature vector set into the long short-term memory encoder, and use the long short-term memory encoder to extract time series features from the temperature influence parameters included in the first feature vector set, and output an initial hidden state sequence; the initial hidden state sequence contains the time dependence and dynamic change pattern of the temperature influence parameters. The third acquisition module is used to input the initial hidden state sequence into the self-attention encoder, use the self-attention encoder to perform global relation modeling on the initial hidden state sequence, and output the target hidden state sequence; the target hidden state sequence includes the long-term effect of the temperature influence parameter on temperature; The fourth acquisition module is used to input the target hidden state sequence into the long short-term memory decoder, use the long short-term memory decoder to perform temperature prediction based on the target hidden state sequence, and output the first predicted temperature curve corresponding to the target crew cabin.

[0013] Thirdly, embodiments of this application also provide an electronic device, which includes: a processor and a memory storing computer program instructions; the processor executes the computer program instructions to implement the temperature prediction method as described in the first aspect.

[0014] Fourthly, embodiments of this application also provide a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the temperature prediction method as described in the first aspect.

[0015] Fifthly, embodiments of this application also provide a computer program product, wherein instructions in the computer program product, when executed by the processor of a model training device, cause the model training device to perform the temperature prediction method as described in the first aspect.

[0016] The temperature prediction method, apparatus, device, storage medium, and product of this application embodiment obtain a first feature vector set of the target passenger cabin; input the first feature vector set into a long short-term memory encoder, and use the long short-term memory encoder to extract time-series features of the temperature influence parameters included in the first feature vector set, outputting an initial hidden state sequence; the initial hidden state sequence contains the time dependency and dynamic change pattern of the temperature influence parameters; input the initial hidden state sequence into a self-attention encoder, and use the self-attention encoder to perform global relationship modeling on the initial hidden state sequence, outputting a target hidden state sequence; the target hidden state sequence contains the long-term effect of the temperature influence parameters on the temperature; input the target hidden state sequence into a long short-term memory decoder, and use the long short-term memory decoder to perform temperature prediction based on the target hidden state sequence, outputting a first predicted temperature curve corresponding to the target passenger cabin. Thus, by adopting a hybrid deep learning architecture that integrates LSTM and Transformer, it can combine the advantages of LSTM in capturing long-term dependencies in time-series data with the powerful capabilities of Transformer in parallel processing and capturing global information. This fusion architecture not only improves the model's understanding and prediction capabilities for complex sequence data, but also significantly enhances training efficiency, while strengthening the model's generalization ability and robustness. Therefore, using a target deep learning model that includes a cascaded long short-term memory encoder, a self-attention encoder, and a long short-term memory decoder for temperature prediction can improve the reliability of temperature prediction. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic flowchart of the temperature prediction method provided in the embodiments of this application; Figure 2 This is an overall architecture diagram of model training provided in the embodiments of this application; Figure 3 This is a flowchart of the preprocessing procedure before model training provided in the embodiments of this application; Figure 4 This is a schematic diagram of the vehicle model embedding and adaptive optimization strategy provided in the embodiments of this application; Figure 5 This is a schematic diagram of the temperature prediction device provided in the embodiments of this application; Figure 6 This is a schematic diagram of the structure of an embodiment of the electronic device provided in this application. Detailed Implementation

[0019] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0020] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0021] This application provides a temperature prediction method applied to an electronic device. The electronic device is equipped with a target deep learning model, which can be a trained model. The target deep learning model includes a long short-term memory encoder, a self-attention encoder, and a long short-term memory decoder connected in series. The method obtains a first feature vector set of the target passenger compartment. This first feature vector set includes feature vectors corresponding to multiple time steps. Each feature vector includes temperature influence parameters of the target passenger compartment at the corresponding time step. For example, if the multiple time steps include 5400 seconds, the first feature vector set includes temperature influence parameters corresponding to each moment within those 5400 seconds. The first feature vector set is then input into the long short-term memory decoder... The long short-term memory encoder extracts time-series features from the temperature-affecting parameters in the first feature vector set, outputting an initial hidden state sequence. This initial hidden state sequence contains the time dependence and dynamic change patterns of the temperature-affecting parameters. The initial hidden state sequence is then input into a self-attention encoder, which performs global relationship modeling on the initial hidden state sequence, outputting a target hidden state sequence. This target hidden state sequence contains the long-term effects of the temperature-affecting parameters on the temperature. The target hidden state sequence is then input into a long short-term memory decoder, which performs temperature prediction based on the target hidden state sequence, outputting the first predicted temperature curve corresponding to the target crew cabin.

[0022] The electronic device can be a body controller, a vehicle terminal, a cloud server, or other servers or vehicles connected to the vehicle, or it can be an electronic product that includes a body controller, a vehicle terminal, a cloud server, or other servers or vehicles connected to the vehicle. The temperature prediction method can be executed by the aforementioned electronic device.

[0023] For ease of understanding, a detailed explanation of the temperature prediction method is provided here. (Refer to...) Figure 1 , Figure 1 This is a schematic flowchart of the temperature prediction method provided in the embodiments of this application.

[0024] The target deep learning model comprises a cascaded Long Short-Term Memory (LSTM) encoder, a self-attention encoder, and an LSM decoder. The LSM encoder excels at capturing long-term and short-term dependencies between time steps, handling sequential dependency modeling and compressing long-term historical information into one or more context vectors. The self-attention encoder is responsible for capturing key information and modeling global dependencies; it receives the hidden states of all time steps from the LSM encoder and analyzes the mutual importance between these states. The LSM decoder is responsible for sequence generation; it utilizes the "information-enhanced" context provided by the self-attention encoder and combines it with previously predicted temperature values ​​to generate future temperature sequences step by step. Therefore, the target deep learning model obtained by cascading the LSM encoder, self-attention encoder, and LSM decoder represents a powerful synergy strategy, combining the robustness of LSM in sequence modeling and point-by-point information extraction with the sensitivity of the self-attention encoder in identifying global dependencies and key information.

[0025] like Figure 1 As shown, the temperature prediction method specifically includes the following steps S101-S104.

[0026] Step S101: Obtain the first feature vector set of the target crew cabin.

[0027] The first feature vector set mentioned above includes feature vectors corresponding to multiple time steps, and each feature vector includes temperature influence parameters of the target crew cabin at the corresponding time step.

[0028] The temperature-affecting parameters mentioned above are those that affect the temperature of the target passenger compartment. These temperature-affecting parameters can be parameters that change over time (such as vehicle speed and illumination, which may change over time) or constant key initial parameters (such as initial temperature and vehicle characteristics).

[0029] The aforementioned temperature-affecting parameters may include at least one of the following: initial heat load inside the target passenger compartment, light intensity (solar radiation intensity), ambient temperature outside the passenger compartment, passenger compartment travel speed, airflow speed and volume at the air conditioning vents, initial temperature inside the passenger compartment, vehicle volume, and insulation parameters.

[0030] Step S102: Input the first feature vector set into the long short-term memory encoder, use the long short-term memory encoder to extract time series features from the temperature influence parameters included in the first feature vector set, and output the initial hidden state sequence.

[0031] The aforementioned initial hidden state sequence contains the time dependence and dynamic change pattern of the temperature-affected parameters. The initial hidden state sequence refers to the sequence of all hidden states output by the Long Short-Term Memory (LSTM) encoder after processing the entire input sequence.

[0032] The aforementioned time dependency can refer to the intrinsic connection between a latent state and its past latent states. It refers to how the "present" is determined and shaped by the "past," reflecting the causal chain of latent states in the time dimension.

[0033] The aforementioned dynamic change pattern describes the overall behavior, trajectory trend, or macroscopic form of a latent state over time. It focuses on the identifiable, regular changes in a sequence of states over long time scales, such as whether it is continuous growth, periodic oscillation, convergence to equilibrium, or divergence into out-of-control behavior.

[0034] The following section explains the LSTM encoder, which is used for time series feature extraction.

[0035] LSTM (Long Short-Term Memory) networks are well-suited for processing time-series data, capable of capturing short-term dependencies and periodic patterns in temperature variations. Multiple stacked LSTM layers can be used as the core of a time-series model.

[0036] The core components of an LSTM encoder are as follows: For processing the input features (temperature-affected parameters), the aforementioned multi-dimensional input feature sequence (the first feature vector set, such as representing the changes in heat load, illumination, vehicle speed, etc. over time) is input into the LSTM together at each time step. For those constant features (such as vehicle type and fixed parameters), they can be input as additional features at the initial time step, and then repeatedly filled into the feature vector corresponding to each time step.

[0037] For hidden layer design, multi-layer LSTMs are used to improve modeling capabilities. Two-layer LSTMs can be used, with each layer having a certain number of hidden units (128 units per layer, which can be adjusted experimentally). Multi-layer structures can extract sequence features layer by layer, and higher-level LSTMs capture longer-term trends.

[0038] For the internal state update mechanism, LSTM can update its internal state using the following equation: ; ; ; ; ; ; in, This represents the input vector at time t (temperature-affected parameters, including dynamic features such as heat load, illumination, and vehicle speed, as well as additional static features). This represents the hidden layer output (hidden state) at time t. This represents the cell state at time t; This represents the activation vector of the input gate, which controls the rate at which the current input updates the cell state. The activation vector represents the forget gate, which controls the proportion of the cell state forgotten from the previous time step. This represents the activation vector of the output gate, used to control the effect of the cell state on the output of the hidden layer; Indicates the candidate cell state. This represents the sigmoid activation function; Represents the hyperbolic tangent activation function; These are the weight matrices for the input gate, forget gate, output gate, and candidate cell states, respectively. These are the bias vectors for the corresponding gate and candidate cell states, respectively; This represents the output of the hidden layer at the previous moment. This indicates the cell state at the previous moment.

[0039] For the output, after processing by two layers of LSTM encoders, the model will generate a hidden state sequence (initial hidden state sequence) {h1, h2, ..., h...} T (If multiple time steps include 5400s, then T=5400), this hidden state sequence contains the temporal dependencies and dynamic change patterns of the input features.

[0040] Step S103: Input the initial hidden state sequence into the self-attention encoder, use the self-attention encoder to perform global relation modeling on the initial hidden state sequence, and output the target hidden state sequence.

[0041] The aforementioned target hidden state sequence contains the long-term effect of temperature-influencing parameters on temperature.

[0042] The following section describes the self-attention encoder (Transformer encoder), which enables global dependency modeling.

[0043] Since sequences with multiple time steps (such as a 5400s sequence) can be quite long, LSTM alone may struggle to remember early information. Therefore, the Transformer self-attention mechanism is introduced to enhance the learning of global dependencies.

[0044] In this embodiment, a Transformer encoder is added between the LSTM encoder and the LSTM decoder. The Transformer can view the features of the entire sequence in parallel, focusing on important information at different time points through a multi-head attention mechanism. A 2–4 layer Transformer encoder can be used, with each layer containing multi-head attention (e.g., using 8 heads) and a feedforward network. The representation (target hidden state sequence) processed by the Transformer will contain the relationships between global time steps (e.g., the long-term effect of illumination changes on subsequent temperatures).

[0045] The process of describing multi-head self-attention is described by the following equation (all operations below are performed on the sequence in parallel at the same time step): ; ; ; ; in, The representation of the input sequence (initial hidden state sequence) can be the matrix obtained by embedding the original feature sequence; These represent the vectors corresponding to the query, key, and value, respectively. The weight matrix is ​​a linear transformation used to map the input to the query, key, and value space; Indicates the dimension of the key vector; This indicates that a softmax operation is performed row by row to obtain the attention distribution; Indicates the first Heah's attention output; For the first A projection matrix specific to each attention head; The number of heads indicating multi-head attention; This indicates that each head will output. Concatenate along the vector dimension; It is a linear transformation weight matrix for multi-head output.

[0046] Step S104: Input the target hidden state sequence into the long short-term memory decoder, use the long short-term memory decoder to perform temperature prediction based on the target hidden state sequence, and output the first predicted temperature curve corresponding to the target crew cabin.

[0047] The following section explains the Long Short-Term Memory (LSTM) decoder.

[0048] The hidden state sequence (target hidden state sequence) processed by the Transformer encoder provides enhanced representation information for the LSTM decoder. The task of the Transformer decoder is to generate predicted temperature curves for multiple time steps based on this representation information.

[0049] As can be seen, the fusion of LSTM and Transformer can be achieved by first using an LSTM encoder to initially encode the input sequence to obtain the initial hidden state sequence, then using a Transformer to perform global relational modeling on the initial hidden state sequence, and finally providing the improved representation (target hidden state sequence) to the LSTM decoder to generate the predicted temperature curve.

[0050] The Transformer captures long-range dependencies by noting the influence of initial peak temperature and ambient temperature on the final segment of the curve, while LSTM ensures sequence continuity and smooth short-term predictions. The fused model can simultaneously possess the ability to provide detailed short-term characterization and grasp long-term trends. This describes how the Transformer's self-attention mechanism processes the entire sequence simultaneously and focuses on the relationships between different time periods, achieving global capture of long-range dependencies. Combined with LSTM, the Transformer can perform multi-faceted, global modeling of the sequence, thereby improving the performance of long-sequence prediction tasks.

[0051] In this embodiment, a hybrid deep learning architecture combining LSTM and Transformer is employed. This architecture leverages the advantages of LSTM in capturing long-term dependencies in time-series data, along with the powerful capabilities of Transformer in parallel processing and capturing global information. This fusion architecture not only improves the model's understanding and prediction capabilities for complex sequence data but also significantly enhances training efficiency. Furthermore, it strengthens the model's generalization ability and robustness, enabling it to perform exceptionally well in handling long sequences, multi-task learning, and large-scale datasets. It combines flexibility and versatility. Therefore, using a target deep learning model comprising a cascaded long short-term memory encoder, a self-attention encoder, and a long short-term memory decoder for temperature prediction can improve the reliability of temperature prediction.

[0052] In some embodiments, the target deep learning model described above is obtained through training. The training of the target deep learning model may include the following steps: Obtain the first training set.

[0053] The target deep learning model is trained using the first training set.

[0054] The first training set includes experimental temperature training samples obtained from experiments and simulated temperature training samples obtained from simulations. Each temperature training sample includes a second feature vector set of a crew cabin and the temperature curve label corresponding to the second feature vector set.

[0055] If the time step corresponding to the first feature vector set can be denoted as the first time step, and the feature vector corresponding to the first feature vector set can be denoted as the first feature vector, then the second feature vector set includes the second feature vectors corresponding to the crew cabin under multiple second time steps, and each second feature vector includes the temperature influence parameters of the crew cabin at the corresponding second time step.

[0056] During the training of the target deep learning model using the first training set, the LSTM decoding process can employ the "Teacher Forcing" technique, which uses the true value as the input at the current time step to guide the decoder in generating the prediction value at the next time step. This method can effectively improve the stability of long sequence prediction.

[0057] In some embodiments, obtaining the first training set may refer to obtaining a first dataset, which includes multiple temperature training samples. After preprocessing and feature engineering the first dataset, the first training set is obtained by partitioning it. High-quality data processing is fundamental to successful model training. For the dataset, systematic cleaning and feature engineering are necessary to improve model training performance.

[0058] The aforementioned passenger cabin can be the passenger cabin of a mobile device, which can be a means of transportation such as a vehicle or an airplane.

[0059] The crew cabins corresponding to each temperature training sample can be different crew cabins.

[0060] Each temperature training sample can be obtained under different operating conditions, which may include temperature conditions (such as ambient temperature, supply air temperature, target temperature, etc.), humidity conditions (such as relative humidity, etc.), wind speed conditions (such as supply air speed, etc.), and illumination conditions (such as solar radiation intensity, solar altitude angle, etc.).

[0061] For example, suppose we need to build a model to predict the temperature of a vehicle's passenger compartment during operation. The first training set could include temperature training samples from 100 vehicles under 100 different operating conditions.

[0062] The above-mentioned experimental temperature training samples are derived from actual experimental operations, while the above-mentioned simulated temperature training samples are generated through computer simulation. The above-mentioned simulated temperature training samples can correspond to the same working conditions as the experimental temperature training samples, that is, the simulated temperature training samples are obtained through computer simulation under the same working conditions as the experimental temperature training samples.

[0063] For example, we can use the experimental temperature training samples of 100 existing vehicles to simulate the temperature under the same conditions to obtain the corresponding simulated temperature training samples, thus forming experimental-simulation paired data.

[0064] In one implementation, simulated temperature training samples can correspond to different operating conditions as experimental temperature training samples. On the other hand, new operating condition data can be generated beyond the experimental range, such as higher initial temperatures, stronger sunlight, and different air conditioning setting temperatures, to enrich the input distribution. Adding these simulated temperature training samples to the training allows the model to see more "extreme" and "uncovered" scenarios, thus enabling it to make reasonable inferences when encountering new real-world scenarios.

[0065] In some implementations, the crew cabin can be tested multiple times under the same operating conditions to obtain multiple experimental temperature training samples, or the crew cabin can be simulated multiple times under the same operating conditions to obtain multiple simulation temperature training samples. Therefore, the experimental temperature training samples can include multiple experimental temperature training samples of the crew cabin under the same operating conditions, and the simulation temperature training samples can include multiple simulation temperature training samples of the crew cabin under the same operating conditions.

[0066] In one implementation, the simulated temperature training sample may contain more types of temperature-influencing parameters than the experimental temperature training sample. The simulated temperature training sample may include temperature-influencing parameters that cannot be obtained through real experiments, such as flow field and temperature field distribution.

[0067] The aforementioned temperature curve labels can include the temperature at each of the multiple second time steps. These temperature curve labels can be either transient cooling curves or transient heating curves. This temperature prediction method can be applied to scenarios where the passenger cabin is in cooling or heating mode, predicting the transient cooling or heating curves of the passenger cabin. When applied to the passenger cabin in cooling mode, the training model's prediction target is the transient cooling curve of the passenger cabin temperature at multiple time steps; when applied to the passenger cabin in heating mode, the training model's prediction target is the transient heating curve of the passenger cabin temperature at multiple time steps.

[0068] In this embodiment, real experimental data represents high-fidelity information captured under limited real-world conditions, containing all actual physical noise and uncertainties. This ensures the model's fundamental effectiveness in the real world. Physical simulation data can generate massive amounts of extreme boundary condition samples at low cost and high efficiency, such as extremely high temperatures, extremely rapid temperature changes, and rare fault modes. These scenarios are difficult to reproduce in experiments or are prohibitively expensive to obtain. The target deep learning model is trained using the first training set, enabling the model to learn both high-fidelity information from the real world and various possible "extreme situations." This allows it to exhibit superior generalization performance and stability when facing new and not entirely identical operating conditions.

[0069] In some embodiments, obtaining the above-mentioned simulated temperature training samples may include the following steps: Using a thermodynamic model of the crew compartment, temperature curves of the crew compartment under various operating conditions are simulated and generated.

[0070] Based on the operating conditions and the temperature curves of the crew cabin under each operating condition, simulation temperature training samples corresponding to each operating condition are generated.

[0071] The crew cabin thermodynamic model was constructed using a one-dimensional system simulation model and a three-dimensional computational fluid dynamics model. The feature vector set of each simulation temperature training sample includes temperature influence parameters for multiple time steps corresponding to the operating condition, and the temperature curve label of each simulation temperature training sample is the crew cabin temperature curve for the corresponding operating condition.

[0072] In some implementations, the one-dimensional system simulation model described above can be constructed using GT-suit, while the three-dimensional computational fluid dynamics model described above can be constructed using STAR-CCM+.

[0073] For scenarios where the training model is used to predict the transient temperature drop curve of the passenger compartment when the air conditioning is on, a thermodynamic model for simulating the passenger compartment temperature is constructed using a one-dimensional system simulation model and a three-dimensional computational fluid dynamics (CFD) model. This can be achieved by building a one-dimensional thermodynamic model (one-dimensional system simulation model) of the vehicle's air conditioning system in GT-SUITE, including key components such as the compressor, condenser, expansion valve, and evaporator, and setting relevant boundary conditions and system parameters. A three-dimensional geometric model (three-dimensional CFD model) of the vehicle's air conditioning system is created in STAR-CCM+, meshed, and the physical properties and boundary conditions of the fluid and solids are defined. Combining the one-dimensional system simulation model and the three-dimensional CFD model yields the thermodynamic model for simulating the passenger compartment temperature. GT-SUITE and STAR-CCM+ can be connected via a co-simulation interface to achieve data exchange and coupled computation. During the co-simulation, GT-SUITE provides system-level thermodynamic parameters, while STAR-CCM+ provides detailed flow field information, jointly simulating the cooling performance of the vehicle's air conditioning system under different operating conditions.

[0074] In this embodiment, a thermodynamic model for vehicle air conditioning cooling is established using both a one-dimensional system simulation model and a three-dimensional computational fluid dynamics model. This approach fully leverages the advantages of both, achieving efficient and accurate system-level and local detail simulations. Furthermore, effective one-dimensional system simulation models and three-dimensional computational fluid dynamics models can be obtained using physical simulation tools such as GT-SUITE and STAR-CCM+. GT-SUITE provides one-dimensional system-level modeling capabilities, enabling the rapid construction of a thermodynamic model of the entire air conditioning system and facilitating system-level performance analysis and optimization. STAR-CCM+ excels in three-dimensional fluid dynamics simulation, capable of simulating the flow field and heat transfer within components such as the condenser and evaporator in detail, providing accurate local heat flow distribution data. Moreover, this co-simulation method can flexibly handle different operating conditions and boundary conditions, ensuring the air conditioning system operates efficiently under various complex conditions. The data generated from these simulations is used to pre-train a deep learning model, enabling the model to learn fundamental physical laws and improving the effectiveness of model training.

[0075] In some embodiments, training the target deep learning model using the first training set may include the following steps: A temperature training sample from the first training set is input into the target deep learning model. The target deep learning model is then used to predict the temperature of the second feature vector set included in the temperature training sample, and the second predicted temperature curve is output.

[0076] The mean square error between the second predicted temperature curve and the temperature curve labels of the temperature training samples is calculated using the first loss function.

[0077] The product of the target squared value and the preset hyperparameter is calculated using the second loss function.

[0078] The target loss function value of the target deep learning model is determined based on the mean square error and the product.

[0079] If the target loss function value does not meet the training stopping condition, adjust the model parameters of the target deep learning model and return to the previous step. Then, input a temperature training sample from the first training set into the target deep learning model until the target loss function value meets the training stopping condition.

[0080] The target square value is the square of the difference between the final temperature on the second predicted temperature curve and the ideal temperature at the final time.

[0081] The above process involves inputting a temperature training sample from the first training set into the target deep learning model, using the target deep learning model to predict the temperature of the second feature vector set included in the temperature training sample, and outputting a second predicted temperature curve. This process can be referred to in steps S102-S104 above for predicting the first predicted temperature curve.

[0082] In one implementation, the first loss function can be the mean squared error (MSE) to measure the error between the second predicted temperature curve and the temperature curve labels in the temperature training samples. The expression of the first loss function can be shown in the following formula (1): (1); Where MSE represents the mean squared error. This represents the temperature at the i-th second time step on the temperature curve label in the temperature training samples. Let represent the temperature at the i-th second time step on the second predicted temperature curve corresponding to the temperature training sample, and n represent the number of second time steps corresponding to the temperature training sample.

[0083] In one implementation, the second loss function is determined by multiplying the square of the difference between the temperature at the final moment (final time step) on the second predicted temperature curve and the ideal temperature at the final moment with a preset hyperparameter. For example, if multiple second time steps include 5400s, then the final moment is the 5400th moment.

[0084] The second loss function mentioned above can be represented by the following formula (2): (2); in, This represents the temperature at the final moment on the second predicted temperature curve. This represents the ideal temperature at the final moment, where λ>0, and λ is a preset hyperparameter.

[0085] The ideal temperature mentioned above can be preset, and it can be calculated based on the vehicle characteristics and environmental conditions corresponding to the passenger compartment.

[0086] In some embodiments, determining the target loss function value of the target deep learning model based on the mean squared error and the product can be achieved by determining the sum of the mean squared error and the product as the target loss function value of the target deep learning model.

[0087] The calculation of the target loss function value can be shown in the following formula (3): (3).

[0088] The above-mentioned method, which adjusts the model parameters of the target deep learning model when the target loss function value does not meet the training stopping condition, and then returns to the process of inputting a temperature training sample from the first training set into the target deep learning model until the target loss function value meets the training stopping condition, can be implemented by adjusting the model parameters of the target deep learning model, obtaining a new temperature training sample from the first training set, and then returning to the process of inputting a temperature training sample from the first training set into the target deep learning model until the target loss function value meets the training stopping condition.

[0089] In this embodiment, the first loss function (MSE) ensures that the model's predictions at all time steps are as close as possible to the true values, optimizing the overall fitting ability. The second loss function forces the model to output more accurately at key time points (such as the final time point), avoiding error accumulation that could cause the final result to deviate from the ideal state.

[0090] In some other embodiments, determining the target loss function value of the target deep learning model based on the mean squared error and the product may further include the following steps: The first loss function value of the target deep learning model is determined based on the mean square error and the product.

[0091] The target loss function value is obtained by weighted summation of the first and second loss function values.

[0092] The second loss function value is determined by the auxiliary quantity curve labels included in the auxiliary quantity training samples and the predicted auxiliary quantity curve obtained by the target deep learning model based on the auxiliary quantity influence parameters included in the auxiliary quantity training samples.

[0093] In this embodiment, when using temperature training samples to train the target deep learning model for temperature prediction, auxiliary training samples can also be used to train the target deep learning model for auxiliary prediction, thus achieving multi-task training of the target deep learning model. The final output layer of the target deep learning model can branch into multiple heads, each outputting prediction curves for different prediction training tasks.

[0094] In one implementation, the above-mentioned use of a first training set to train the target deep learning model can be achieved by using both a first training set and a third training set. The third training set includes multiple auxiliary quantity training samples from the passenger cabin, each auxiliary quantity training sample including the influence parameters of the auxiliary quantities in the passenger cabin and the corresponding auxiliary quantity curve labels.

[0095] In some embodiments, obtaining the third training set may involve obtaining a second dataset, which includes multiple auxiliary training samples. After preprocessing and feature engineering the second dataset, the third training set is obtained from the second dataset.

[0096] The process of determining the second loss function value from the auxiliary quantity curve labels included in the auxiliary quantity training samples and the predicted auxiliary quantity curve obtained by the target deep learning model based on the auxiliary quantity influence parameters included in the auxiliary quantity training samples can be referred to as the process of determining the first loss function value from the temperature curve labels included in the temperature training samples and the second predicted temperature curve obtained by the target deep learning model based on the temperature influence parameters included in the temperature training samples. It will not be elaborated here.

[0097] The above method of determining the first loss function value of the target deep learning model based on the mean squared error and the product can be achieved by summing the mean squared error and the product to determine the first loss function value of the target deep learning model.

[0098] The above-mentioned weighted calculation based on the first loss function value and the second loss function value to obtain the target loss function value can be achieved by setting a first weight corresponding to the first loss function value and a second weight corresponding to the second loss function value, and then performing a weighted calculation based on the first loss function value, the second loss function value, the first weight, and the second weight to obtain the target loss function value.

[0099] The aforementioned auxiliary quantity training samples can be experimental auxiliary quantity training samples or simulation auxiliary quantity training samples. The aforementioned experimental auxiliary quantity training samples can be obtained through actual experimental measurements, such as through corresponding sensors, while the aforementioned simulation auxiliary quantity training samples can be obtained through computer simulation.

[0100] The aforementioned auxiliary quantities are physical quantities that affect the temperature of the passenger compartment. For example, in a scenario where the training model is used to predict the transient cooling curve of the passenger compartment temperature at multiple time steps when the air conditioning in the passenger compartment is turned on, the aforementioned auxiliary quantities can be the temperature of different positions in the driver's seat or other seats in the passenger compartment, or the power of the air conditioning compressor or the air conditioning cooling capacity, etc.

[0101] The above-mentioned auxiliary quantity influencing parameters are parameters that affect the change of auxiliary quantity. When the auxiliary quantity is the air conditioning cooling capacity, the auxiliary quantity influencing parameters may include at least one of the following: cooling power, indoor and outdoor temperature difference, solar radiation intensity, or humidity.

[0102] The auxiliary quantity training samples mentioned above include auxiliary quantity influence parameters, which may include auxiliary quantity influence parameters at each of the multiple second time steps. The auxiliary quantity curve labels mentioned above may include auxiliary quantities at each of the multiple second time steps.

[0103] In this embodiment, temperature prediction training can be used as the key training task, and auxiliary quantity prediction training can be used as the auxiliary task. Therefore, the first weight can be higher than the second weight, and the weight relationship between each second loss function value can be determined according to the degree of influence of the corresponding auxiliary quantity on temperature prediction.

[0104] The aforementioned auxiliary quantity prediction task can also be designed as a derived task such as predicting the final equilibrium temperature at the final moment of a preset duration (multiple second time steps), or the time required to reach a preset temperature threshold.

[0105] When the target loss function is used to calculate the value of the first loss function, the first loss function is used to determine the mean square error between the temperature on the temperature curve label in the temperature training sample and the temperature on the corresponding second predicted temperature curve. The second loss function is determined by the product of the square of the difference between the temperature at the final time and the ideal temperature at the final time on the second predicted temperature curve and the preset hyperparameter.

[0106] When the target loss function is used to calculate the value of the second loss function, the first loss function is used to determine the mean square error between the auxiliary quantity on the auxiliary quantity curve label and the corresponding auxiliary quantity on the predicted auxiliary quantity curve in the auxiliary quantity training sample. The second loss function is determined by the product of the square of the difference between the auxiliary quantity at the final time and the ideal auxiliary quantity at the final time on the predicted auxiliary quantity curve and the preset hyperparameter.

[0107] In this embodiment, the target loss function value is obtained by weighting the loss function values ​​corresponding to different prediction tasks. By reasonably setting the weights of the loss function, the importance of different prediction tasks can be balanced, ensuring that the model performs well on key tasks while maintaining good performance on other auxiliary tasks. This helps the model better capture complex relationships in the data and improves its adaptability and robustness in practical applications.

[0108] In one implementation, the above-mentioned training of the target deep learning model using the first training set and the third training set may also include, but is not limited to, the following steps: The temperature influence parameters in the simulated temperature training samples of the mixed training sample combination are input into the target deep learning model. The target deep learning model uses the temperature influence parameters to predict the temperature and obtain the simulated predicted temperature curve. The auxiliary quantity influence parameters in the simulated auxiliary quantity training samples of the mixed training sample combination are input into the target deep learning model. The target deep learning model uses the auxiliary quantity influence parameters to predict the auxiliary quantity and obtain the simulated predicted auxiliary quantity curve.

[0109] The above-mentioned mixed training sample combination includes experimental temperature training samples, simulated temperature training samples, experimental auxiliary quantity training samples, and simulated auxiliary quantity training samples. Each training sample in the above-mentioned mixed training sample combination can correspond to the same working condition.

[0110] For each mixed training sample combination, the third loss function value is determined based on the simulated predicted temperature curve and the temperature curve labels of the experimental temperature training samples in the mixed training sample combination; the fourth loss function value is determined based on the simulated predicted temperature curve and the temperature curve labels of the simulated temperature training samples in the mixed training sample combination; the fifth loss function value is determined based on the simulated predicted auxiliary quantity curve and the auxiliary quantity curve labels of the experimental auxiliary quantity training samples in the mixed training sample combination; and the sixth loss function value is determined based on the simulated predicted auxiliary quantity curve and the auxiliary quantity curve labels of the simulated auxiliary quantity training samples in the mixed training sample combination.

[0111] For each combination of mixed training samples, the target loss function value is obtained by weighting the third, fourth, fifth, and sixth loss function values.

[0112] If the target loss function value does not meet the training stopping condition, adjust the model parameters of the target deep learning model to obtain the updated target deep learning model, and return to input the temperature influence parameters of the simulated temperature training samples in the mixed training sample combination into the target deep learning model until the training stopping condition is met, thus obtaining the trained target deep learning model.

[0113] In this embodiment, the loss function value between the experimental curve and the simulation curve under the corresponding conditions is used as part of the target loss function, so that the model output can both approximate the true value and deviate too much from the physical simulation value. In this way, while pursuing accuracy, the model is constrained by the physical model, and will not produce prediction results that obviously violate physical common sense.

[0114] In some embodiments, after outputting the second predicted temperature curve corresponding to the temperature training sample and before calculating the mean square error between the second predicted temperature curve and the temperature curve labels included in the temperature training sample using a first loss function, the method may further include: smoothing the second predicted temperature curve if the second predicted temperature curve does not satisfy a preset monotonicity.

[0115] For scenarios where temperature prediction is performed in the passenger cabin when the air conditioning is on, the aforementioned second predicted temperature curve must conform to a monotonically decreasing trend. For scenarios where temperature prediction is performed in the passenger cabin when the air conditioning is on, the aforementioned second predicted temperature curve must conform to a monotonically increasing trend.

[0116] In one implementation, for the auxiliary quantity prediction task, if it is necessary to output a predicted auxiliary quantity curve that meets the monotonicity, after outputting the predicted auxiliary quantity curve corresponding to the auxiliary quantity training sample and before calculating the mean square error between the predicted auxiliary quantity curve and the auxiliary quantity curve labels included in the auxiliary quantity training sample through the first loss function, it may further include: smoothing the predicted auxiliary quantity curve if the predicted auxiliary quantity curve does not meet the preset monotonicity.

[0117] In this embodiment, the second predicted temperature curve is smoothed to ensure it meets specific monotonicity requirements. This effectively reduces noise and abnormal fluctuations in the second predicted temperature curve, making it more consistent with actual physical laws and improving the reliability of the prediction results. Furthermore, the smoothed monotonicity curve simplifies subsequent analysis and processing, enhances the model's generalization ability, and makes it more stable when facing new input data, thus better meeting the needs of practical applications.

[0118] In some embodiments, the aforementioned passenger compartment may be the passenger compartment of a vehicle, and the aforementioned temperature training samples may also include vehicle model data.

[0119] After inputting the temperature training samples into the target deep learning model as described above, the following steps may also be included, but are not limited to: The vehicle model embedding vector corresponding to the vehicle model data is embedded into the target deep learning model.

[0120] The vehicle model embedding vectors mentioned above are used to represent vehicle models, and these embedding vectors can be used to represent the characteristics of the corresponding vehicle models.

[0121] Different vehicle models (such as black sedans and white SUVs) can have subtle but potentially significant effects on the surrounding microenvironment temperature due to factors such as color, size, material, and engine thermal efficiency, or their operating state may be highly correlated with temperature. Therefore, "vehicle model" is not a random variable unrelated to temperature, but rather a characteristic that may contain thermodynamic information.

[0122] The vehicle data mentioned above may include at least one of the following: vehicle series, model, body style, body color, vehicle size, and vehicle weight.

[0123] The process of embedding the vehicle model embedding vector corresponding to the vehicle model data into the target deep learning model can include the following steps: A target deep learning model is used to extract features from the model data to obtain the model features corresponding to the model data.

[0124] The model feature is encoded into a model embedding vector using a target deep learning model, and the model embedding vector is then embedded into the embedding vector layer in the target deep learning model.

[0125] To further enhance the model's adaptability under different vehicles and operating conditions, an adaptive learning rate optimizer (Adam, AdamW, etc.) is used during model training, combined with learning rate scheduling strategies (cosine annealing and exponential decay) to ensure loss convergence and generalization. Early training allows for rapid learning of the overall trend, while later training fine-tunes the learning rate to reduce prediction error.

[0126] In this embodiment, by embedding vehicle model embedding vectors corresponding to vehicle model data into the model through transfer learning or meta-learning, richer semantic information and feature representations can be provided for vehicle temperature prediction, enabling the model to adapt to the "vehicle model differences" dimension. Vehicle model embedding vectors can capture the unique attributes of different vehicle models, helping the model better understand the temperature change patterns of different vehicle models under different operating conditions, thereby improving the accuracy and robustness of predictions.

[0127] In some embodiments, the target passenger compartment can be the passenger compartment of a target vehicle. Before inputting the first feature vector set into the long short-term memory encoder, the following steps may also be included: Obtain the second training set.

[0128] The second training set mentioned above includes data on multiple vehicle models associated with the model of the target vehicle.

[0129] The target deep learning model is adjusted using a second training set.

[0130] The above steps are commonly referred to as "fine-tuning" in machine learning. The core idea is to first train a base model using a large, general dataset, and then fine-tune this pre-trained model using a small, specialized dataset to make it perform better on a specific task.

[0131] In this embodiment, after training, the vehicle embedding vector is fine-tuned for new vehicle models using only a few samples, thus quickly adapting to the new models (achieving "sample-based model customization"). This method is equivalent to adding adaptive parameters to the model; when a new vehicle arrives, only a small number of parameters need to be adjusted to optimize prediction accuracy.

[0132] To better understand the above method, the embodiments of this application provide the following complete embodiments for illustration: Currently, in the field of transient cooling prediction for passenger cabin air conditioning, the main traditional methods include simulation methods based on physical models and single deep learning models (such as LSTM or Transformer). However, these methods have the following limitations: Traditional physical models, while providing theoretical guidance based on thermodynamic principles, often oversimplify their assumptions, making it difficult to accurately capture dynamic changes under complex operating conditions. Furthermore, the predictive accuracy of traditional models significantly decreases when facing extreme conditions or novel vehicle designs.

[0133] Single deep learning models, such as those using only LSTM or Transformer, are insufficient for long sequence prediction tasks. While LSTM excels at processing time series data, it has limitations in capturing global temporal correlations; and while Transformer can model global dependencies, it lacks the ability to finely characterize local dynamics when processing continuous time series.

[0134] Existing fusion solutions: Previous research has attempted to combine physical simulation and machine learning methods for prediction, but most are limited to simple data fusion or single-task learning frameworks. These methods fail to fully utilize the advantages of combining the time-series modeling capabilities of deep learning models with the prior knowledge of physical simulation.

[0135] In summary, existing technologies still have significant shortcomings in handling long sequence prediction tasks, adaptability to multiple vehicle models, and generalization capabilities under complex operating conditions. Therefore, there is an urgent need for a prediction method that can effectively combine deep learning and physical simulation, and possesses good generalization and adaptability, as illustrated in the following embodiments.

[0136] Based on experimental data from 100 vehicles of different models (SUVs, sedans, and MPVs), this study requires the development of a vehicle interior transient cooling prediction model. This model aims to predict the cooling process over 5400 seconds (approximately 90 minutes) after the vehicle's air conditioning is turned on. Given environmental and operating parameters, it should output a curve showing the temperature decrease over time inside the vehicle, ensuring that the predicted temperature does not deviate from the measured temperature by more than 1°C. This approach utilizes deep learning (combining LSTM and Transformer), integrates physical simulation data, and employs sophisticated data processing and training strategies to achieve high-precision prediction.

[0137] The overall architecture diagram is as follows: Figure 2 As shown, Figure 2 This is a diagram showing the overall architecture for model training provided in this embodiment.

[0138] Input parameters ( Figure 2 Input features): The model input covers key factors affecting vehicle interior cooling (i.e., the temperature influence parameters mentioned above), including but not limited to: Heat load: Initial heat load inside the vehicle (number of occupants and heat dissipation from equipment); Light intensity: Solar radiation intensity (simulating the effect of direct sunlight on the temperature increase inside the vehicle); External temperature: Ambient temperature outside the vehicle (°C); Vehicle speed: The speed at which a vehicle travels (affects external airflow and heat dissipation); Air outlet velocity: The air velocity and air volume at the air conditioner outlet (a measure of the air conditioner's cooling capacity). Other characteristics: initial interior temperature, vehicle volume, and thermal insulation parameters (used to characterize differences between different vehicle models).

[0139] These inputs can be time-varying sequences (studying vehicle speed and illumination variations over time) or constant initial conditions (initial temperature and vehicle characteristics). When using the model, a sequence of operating conditions and key initial parameters for each time point within 5400 seconds are required.

[0140] Output objective: The model outputs a transient cooling curve of the vehicle interior temperature over 5400 seconds. This is typically represented as a 5400-second time series (temperature prediction once per second). The model should capture the rapid temperature drop from the initial value and its gradual stabilization. To meet accuracy requirements, the predicted temperature across the entire curve should closely match the actual experimental curve, ensuring an error within ±1℃.

[0141] To address the challenge of predicting long-term series data while maintaining high accuracy, this approach employs a hybrid deep learning architecture that integrates LSTM and Transformer, incorporating necessary physical priors. The main idea is to leverage LSTM's expertise in time series modeling and Transformer's global dependency capture capabilities to design a model that can learn short-term dynamics while also memorizing long-term trends. The specific architecture and strategy are as follows: To address the challenge of predicting long-term series data while maintaining high accuracy, this approach employs a hybrid deep learning architecture that integrates LSTM and Transformer, incorporating necessary physical priors. The main idea is to leverage LSTM's expertise in time series modeling and Transformer's global dependency capture capabilities to design a model that can learn short-term dynamics while also memorizing long-term trends. The specific architecture and strategy are as follows: 1. LSTM Encoder: Time Series Feature Extraction LSTM (Long Short-Term Memory) networks are well-suited for processing time-series data, capable of capturing short-term dependencies and periodic patterns in temperature variations. This research constructs multiple stacked LSTM layers as the core of the model's time-series modeling: Input feature processing: The aforementioned multi-dimensional input feature sequence (feature vector set, changes in heat load, illumination, vehicle speed, etc. over time) is input into the LSTM at each time step. For features that remain constant (vehicle type and fixed parameters), they can be input as additional features at the initial time step, and can be repeatedly padded into the feature vector at each time step.

[0142] Hidden layer design: Multi-layer LSTM is used to improve modeling capabilities. The study uses a two-layer stacked LSTM, with each layer having a certain number of hidden units (128 units per layer, which can be adjusted according to experiments). The multi-layer structure can extract sequence features layer by layer, and the higher-level LSTM captures longer-term trends.

[0143] For the internal state update mechanism, LSTM can update its internal state using the following equation: ; ; ; ; ; ; in, This represents the input vector at time t (temperature-affected parameters, including dynamic features such as heat load, illumination, and vehicle speed, as well as additional static features). This represents the hidden layer output (hidden state) at time t. This represents the cell state at time t; This represents the activation vector of the input gate, which controls the rate at which the current input updates the cell state. The activation vector represents the forget gate, which controls the proportion of the cell state forgotten from the previous time step. This represents the activation vector of the output gate, used to control the effect of the cell state on the output of the hidden layer; Indicates the candidate cell state. This represents the sigmoid activation function; Represents the hyperbolic tangent activation function; These are the weight matrices for the input gate, forget gate, output gate, and candidate cell states, respectively. These are the bias vectors for the corresponding gate and candidate cell states, respectively; This represents the output of the hidden layer at the previous moment. This indicates the cell state at the previous moment.

[0144] Output: After processing by two LSTM encoder layers, the model will generate a hidden state sequence (initial hidden state sequence) {h1, h2, ..., h...} T (T=5400), this hidden state sequence contains the temporal dependence and dynamic change pattern of the input features.

[0145] 2. Transformer Encoder: Global Dependency Modeling Since a 5400-second sequence is quite long, LSTM alone may struggle to remember early information. Therefore, this solution introduces a Transformer self-attention mechanism to enhance the learning of global dependencies. Transformer Encoder: A Transformer encoding module is added between the LSTM encoder and decoder. The Transformer can examine the features of the entire sequence in parallel, focusing on important information at different time points through a multi-head attention mechanism. Research can use 2–4 layers of Transformer encoders, each layer containing multi-head attention (8 heads in this study) and a feedforward network. The representation processed by the Transformer will include the relationships between global time steps (the long-term effects of illumination changes on subsequent temperatures).

[0146] The process of describing multi-head self-attention is described by the following equation (all operations below are performed on the sequence in parallel at the same time step): ; ; ; ; in, The representation of the input sequence (initial hidden state sequence) can be the matrix obtained by embedding the original feature sequence; These represent the vectors corresponding to the query, key, and value, respectively. The weight matrix is ​​a linear transformation used to map the input to the query, key, and value space; Indicates the dimension of the key vector; This indicates that a softmax operation is performed row by row to obtain the attention distribution; Indicates the first Heah's attention output; For the first A projection matrix specific to each attention head; The number of heads indicating multi-head attention; This indicates that each head will output. Concatenate along the vector dimension; It is a linear transformation weight matrix for multi-head output.

[0147] 3. LSTM Decoder: Temperature Profile Generation The hidden state sequence (target hidden state sequence) processed by the Transformer encoder provides enhanced representation information for the LSTM decoder. The decoder's task is to generate temperature prediction curves over 5400 seconds based on this representation information.

[0148] Integration methods (such as) Figure 2 The feature fusion module in the model first uses LSTM to initially encode the input sequence to obtain the hidden state sequence. Then, a Transformer is used to model the global relationships of the hidden state sequence, and the improved representation is provided to the LSTM decoder to generate predictions. Additionally, during the training phase, the decoding process can employ a "teacher forcing" technique: using the true value as the input at the current time step to guide the decoder in generating the prediction value for the next time step. This method can effectively improve the stability of long sequence predictions.

[0149] Finally, the predicted temperature curve, i.e., the 5400-second temperature prediction curve, is output through the output module.

[0150] 4. Summary of Model Advantages Transformer captures long-range dependencies and can observe the impact of initial peak temperature and ambient temperature on the final segment of the cooling curve; LSTM ensures sequence continuity and smooth short-term predictions. The fused model is expected to possess both the ability to provide detailed short-term characterization and to capture long-term trends.

[0151] The equations above describe how the Transformer's self-attention mechanism processes the entire sequence simultaneously and focuses on the relationships between different time periods, achieving global capture of long-range dependencies. Combined with LSTM, the Transformer can perform multi-angle, global modeling of sequences, thereby improving the performance of long sequence prediction tasks.

[0152] as follows Figure 3 As shown, Figure 3 The flowchart of the preprocessing process before model training provided in the embodiments of this application shows the overall process of preprocessing, feature extraction, data fusion and parameter limitation modules for real experimental data and physical simulation data.

[0153] (1) Integrated physical simulation (GT-suit & Starccm+) data: To enhance the physical credibility of the model and its generalization to new vehicle models, the study considers combining physical priors provided by simulation tools such as GT-suit and Starccm+: Simulation pre-training: A thermodynamic model of vehicle air conditioning cooling was built using GT-suit & Starccm+ to simulate cooling curves under various vehicle models and conditions. The simulation data (i.e.,...) was then used for pre-training. Figure 3 The physical simulation data is used to pre-train the deep learning model, enabling the model to learn basic physical laws (including the approximate shape of the cooling curve and the time constant). The amount of simulation data can be very large; training with it first can improve the model's robustness in predicting extreme conditions.

[0154] Simulation data fusion (e.g.) Figure 2 as well as Figure 3 The data fusion module integrates real experimental data and physical simulation data: In actual training, simulation curves are added to the training set as a form of data augmentation. The research utilizes multi-task learning to allow the model to simultaneously predict real experimental curves and simulation curves under corresponding conditions (i.e., the output should approximate the true value without deviating too much from the physical simulation value). In this way, while pursuing accuracy, the model is constrained by the physical model, preventing predictions that significantly violate common physical principles.

[0155] Physical Prior Constraints: Traditional loss functions typically only focus on the difference between the model output and the true value. In this study, to make the prediction results more consistent with actual physical laws (such as the interior temperature of the vehicle should gradually decrease and tend to stabilize), a basic mean squared error (MSE) and a final temperature deviation penalty term L were designed. final The comprehensive loss function L total : The basic mean square error (MSE) measures the error in temperature prediction over all times in the entire time series. ; in, This represents the predicted temperature at second t. This represents the actual temperature at second t.

[0156] The final temperature deviation penalty term encourages the model to predict a final temperature close to the ideal in-vehicle temperature T. ideal : ; in, This indicates the predicted temperature at the 5400th second. It is the ideal final temperature calculated based on vehicle characteristics and environmental conditions. λ>0 is a hyperparameter used to control the importance of this penalty term.

[0157] The total loss function combines the two parts mentioned above: L total =L MSE +L final ; During training, the optimizer will adjust the total loss function L. total To adjust the model parameters. By introducing L final The model not only needs to minimize the overall prediction error (i.e., fit the real cooling curve as accurately as possible), but also needs to ensure that the temperature at the final moment is close to the ideal value. This dual optimization objective helps to avoid the model focusing too much on intermediate moments and ignoring long-term trends. Meanwhile, simulations show that the cooling curve should decrease monotonically and have fixed convexity, allowing for the smoothing of non-physical oscillations during post-prediction processing.

[0158] (2) Multi-task learning strategies (such as...) Figure 2 Multi-task learning in China): To improve the generalization ability of the model, research can introduce a multi-task learning architecture, which allows the model to optimize multiple related objectives simultaneously during training.

[0159] Auxiliary Task Selection: In addition to the primary task of predicting the average in-vehicle temperature, relevant physical quantities can be selected as auxiliary tasks for prediction. This includes researching and predicting temperature curves at different locations in the driver's seat and rear seats (if multiple sensors are used in the experiment), and predicting the power / cooling capacity of the air conditioning compressor over time. Data for auxiliary tasks can come from simulations and experiments. If no additional measurements are available, derivative tasks such as predicting the final equilibrium temperature at 5400 seconds and the time required to reach a certain temperature threshold can also be designed.

[0160] Joint training: The model's final output layer can branch into multiple heads, each outputting predictions for different tasks. During training, the loss is calculated for each task, and the network is updated using a weighted sum. Weights are appropriately set to prioritize the accuracy of the primary task (temperature curve), while auxiliary tasks are used to train and enhance the physical meaning and robustness of internal features.

[0161] Results: Multi-task learning forces the model to extract more general features, avoiding the bias of only memorizing a single task, thus enabling it to provide reliable predictions even when encountering new vehicle models and slightly different operating conditions. Simultaneous prediction of temperatures at different locations allows the model to learn the heat distribution patterns inside the vehicle more comprehensively, improving its adaptability to vehicles with new layouts.

[0162] (3) Adaptive optimization strategies (such as...) Figure 2 (Adaptive vehicle model) To address the cooling characteristics of different vehicle models, this method introduces vehicle model embedding and adaptive optimization strategies, such as... Figure 4 As shown, Figure 4 This diagram illustrates the vehicle model embedding and adaptive optimization strategy provided in this application, showcasing the overall process from feature extraction of different vehicle model data to vehicle model vector embedding, and then to adaptive training using transfer learning / meta-learning. Figure 4 and Figure 3 The initial model in the model is a deep learning model (or simply model).

[0163] To further enhance the model's adaptability under different vehicles and operating conditions, the study considers employing adaptive optimization techniques during training and deployment: Adaptive training: During model training, an adaptive learning rate optimizer (Adam, AdamW, etc.) is used, combined with learning rate scheduling strategies (cosine annealing and exponential decay) to ensure loss convergence and generalization. Early training allows for rapid learning of the overall trend, while later training fine-tunes the learning rate to reduce prediction error.

[0164] Meta-learning / transfer learning: Using data from 100 vehicles, model transfer and meta-learning strategies can be employed to enable the model to adapt to the dimension of "vehicle model differences." The research first trains a model containing vehicle model embedding vectors, with each vehicle corresponding to an embedding representing its characteristics. After training, for new vehicle models, the embedding vectors can be fine-tuned with a small number of samples, thus quickly adapting to the new vehicle model (achieving "sample-based model customization"). Figure 4 (The adaptive optimization module in the model). This method is equivalent to adding adaptive parameters to the model; when a new vehicle arrives, only a small number of parameters need to be adjusted to optimize the prediction accuracy.

[0165] Online calibration (e.g.) Figure 4 Online Feedback and Correction: An online learning mechanism can be enabled after deployment. If the vehicle-mounted terminal can obtain real-time temperature feedback, the model can be corrected online by comparing the predicted and measured errors. The study employs local online training (briefly training and updating the model on the vehicle-mounted terminal with a small amount of new data) and a lighter correction method (adjustment of correction coefficients is required). To avoid overfitting and safety issues, the online update process must be cautious; the study only adjusts the bias of the last layer of the model to eliminate systematic bias.

[0166] (4) Data processing and feature engineering (e.g.) Figure 3 (Data preprocessing in the process) High-quality data processing is fundamental to model success. For experimental data from 100 different vehicle models, the study requires systematic data cleaning and feature engineering to improve model training performance. Data integration and labeling: First, aggregate data from 100 vehicles, ensuring each experimental record includes the input feature sequence and the corresponding temperature curve. If data sources have different formats, they need to be converted to a unified format. The study aligns timestamps to a uniform interval (1 second and other fixed steps) and marks missing records.

[0167] Outlier Handling: Inspect experimental data for outliers and errors. These may include temperature spikes due to sensor malfunctions, abnormal peaks in illumination values, etc. Outliers can be removed or interpolated: small outliers can be smoothly replaced with values ​​from preceding and following times; obviously erroneous experimental segments may need to be deleted or processed separately.

[0168] Noise reduction and smoothing: While in-vehicle temperature changes are theoretically relatively smooth, sensor readings may contain high-frequency noise. Filtering methods can be used to reduce noise in temperature sequences and continuous input features (real-time solar radiation intensity curves). Research is being conducted on moving average filtering and low-pass filtering to filter out minor, meaningless fluctuations while preserving the overall trend. This helps the model learn the main patterns more easily without being disturbed by noise.

[0169] Feature standardization: This involves standardizing / normalizing the input features. Since features have different dimensions (light intensity might range from hundreds to thousands, temperature from tens, and vehicle speed from tens), direct input can make it difficult for the network to balance signals of different magnitudes. A common method is to subtract the training set mean from each feature, divide by the standard deviation (standardization), and scale to the [0,1] interval (Min-Max normalization). Note: The standardization parameters (mean / variance) must be calculated on the training set and applied to validation and testing to prevent information leakage.

[0170] Feature engineering and selection: Based on an understanding of physics, new features can be derived to aid the model. For example: Calculating cumulative solar heat: Solar radiation intensity has a cumulative effect on temperature rise; the cumulative heat load at each moment can be obtained by integrating the solar radiation intensity as a feature. Initial temperature difference: The difference between the initial temperature inside the vehicle and the target temperature (outside temperature and air conditioning set temperature) serves as the temperature difference that cooling needs to overcome, helping the model understand the initial cooling rate. Vehicle parameters: If some inherent vehicle parameters (interior volume, glass area, air conditioning power, etc.) are known, they can be added as static features. This helps the model distinguish the thermal inertia differences between SUVs and sedans. If not explicitly provided, vehicle type (categorical variable) can be considered as a feature using embedding vectors.

[0171] Dataset partitioning: The cleaned data is divided into training, validation, and test sets. To evaluate the model's predictive ability for "new car models," a vehicle-based partitioning method is recommended: 80 vehicles are selected for training, and 20 vehicles are reserved entirely for testing (the model does not see data from these 20 vehicles during training). This partitioning ensures that every curve in the test set comes from a car model the model has not seen before, simulating the situation of encountering a new car model in real-world applications. Furthermore, a portion of curves is extracted from the training set as a validation set for parameter tuning and early stopping. The distribution of various car models and operating conditions within each set should be as similar as possible to prevent training bias.

[0172] Through the above processing, a clean, rich, and structurally consistent training dataset is obtained, which is ready for model training.

[0173] (5) Data augmentation strategies: Considering that the amount of data from only 100 vehicles might be limited, the study employed various data augmentation methods to expand the data coverage, including introducing physical simulation data and traditional data augmentation techniques: Physical simulation data expansion: Additional cooling process data under different conditions is generated using simulation tools such as GT-suit and Starccm+. On one hand, simulations can be run under identical conditions on an existing dataset of 100 vehicles to obtain corresponding simulation curves, forming real-simulation paired data for model calibration (as mentioned in the multi-task training). On the other hand, new operating condition data can be generated beyond the experimental scope to study conditions such as higher initial temperatures, stronger sunlight, and different air conditioning set temperatures, enriching the input distribution. Adding this simulation data to training allows the model to encounter more "extreme" and "uncovered" scenarios, enabling it to make reasonable inferences when encountering new real-world scenarios.

[0174] Interpolation and Synthesis: Interpolating experimental data can construct additional samples. For example, if two experimental curves correspond to light intensities of 600 W / m² and 800 W / m² respectively, the study can interpolate to obtain an approximate cooling curve under approximately 700 W / m² conditions as a training sample (assuming a relatively linear thermal effect). Interpolation can be performed in the feature space (between input parameters) or by making small perturbations to the time series: slightly shifting and scaling the temperature curve of a car to simulate another possible outcome under similar conditions. This synthesis based on existing data increases sample diversity.

[0175] Noise perturbation: Adding slight random noise to the input data during training is also a common data augmentation method. Studies have shown that multiplying the light intensity curve by a small random factor of (1±5%) and slightly fluctuating the ambient temperature by 0.5℃ are effective methods. The output curve can also be supplemented with noise to simulate measurement errors as a training target. This prevents the model from overfitting the precise training curve and enhances its robustness to measurement errors and environmental randomness.

[0176] Time axis transformation: To adapt the model to different sampling times and process speeds, the cooling curve can be stretched / compressed in the time direction as data augmentation. In this study, a 5400-second process in an experiment was simulated by interpolation / thinning to take 4800 seconds and 6000 seconds (equivalent to slightly faster and slower processes), while simultaneously adjusting the input feature (the rate of change of ambient temperature) accordingly. This teaches the model to predict changes in the cooling rate, preventing it from adapting only to a fixed rhythm.

[0177] Ensuring physical plausibility: Data augmentation must be handled with care to ensure it does not violate fundamental physical laws. The temperature curve should not rise midway, and the combination of input features must be reasonable (avoid generating contradictory samples where sunlight is zero but the cabin temperature is rising). It is crucial to use simulation models and physical experience to verify the plausibility of the synthesized data. Only high-quality augmented data can truly improve model generalization without being misleading.

[0178] With the above enhancements, the training set will contain richer scenarios, which will help the model learn more comprehensive mapping relationships and improve the prediction accuracy for new vehicle models and new operating conditions.

[0179] (6) Model training and validation (e.g.) Figure 3 (Parameter constraint module and initial model parameters in the model) With the architecture and data in place, the training process needs careful design to ensure the model converges to a high-precision solution that meets the error requirements. Below are the main steps and technology choices for training and validation: Model Implementation and Hyperparameters: The above-mentioned fusion model architecture was implemented using popular deep learning frameworks (TensorFlow / Keras and PyTorch). Appropriate hyperparameters were set, including: the number of LSTM hidden units (128), the number of layers (two-layer LSTM), the number of Transformer heads and layers (8 heads, 2 self-attention layers), and the dropout ratio (to prevent overfitting, set to 0.1~0.3). He normal initialization and other initialization methods were used to initially stabilize the network weights.

[0180] Loss function design: The mean squared error (MSE) and mean absolute error (MAE) are primarily used as the regression loss for the temperature series to measure the difference between the predicted and actual curves. Furthermore, to strictly control the maximum error, a peak error loss can be added to the loss: this involves calculating the maximum absolute value of the difference between the predicted and actual temperatures and penalizing any deviations exceeding 1°C. This will guide the model to specifically avoid large biases. If multi-task learning is used, the total loss is a weighted sum of the main task loss and the auxiliary task loss; the hyperparameters need to be adjusted and balanced.

[0181] Training Strategy: Segmented sequence training improves efficiency and stability. Since the 5400-step sequence is relatively long, truncated backpropagation (TBPTT) can be used: the sequence is divided into smaller segments (equivalent to 300 seconds of data per 5 minutes input to the model, with hidden states running between segments). This ensures that the LSTM can learn long-range relationships while reducing the computational cost per run and preventing gradient vanishing and exploding. For Seq2Seq models, teacher-forced prediction can also be tried: the predictions for the first n seconds during training are replaced with true values ​​to reduce the accumulation of long prediction errors. As training progresses, the proportion of teacher-forced prediction can be gradually reduced, allowing the model to learn to autonomously predict long sequences.

[0182] Algorithm optimization: The Adam adaptive optimization algorithm is used for gradient descent iterations to accelerate convergence. Learning rate scheduling is incorporated: the initial learning rate is set to 1e-3. If the validation set error no longer decreases after several epochs, the learning rate is multiplied by 0.1 to reduce the learning rate and training continues. An Early Stopping mechanism is implemented: the validation set error is monitored; if it does not improve after several consecutive epochs, training is stopped and the optimal model parameters are selected to prevent overfitting.

[0183] Integrated training process: If simulation pre-training is introduced, the main parts of the model are first trained on a large batch of simulation data (parameters such as vehicle model embedding can be frozen to train common parts first), allowing the model to initially grasp physical trends. Then, training is switched to real data, gradually unfreezing all parameters and fine-tuning with a small learning rate, so that the model does not forget physical laws while fitting real curves. During training, real data and simulation data can be mixed and fed in. For example, 80% of a batch consists of real curves and 20% consists of simulation curves to maintain physical constraints.

[0184] Validation and Evaluation: Evaluate model performance on a validation set (containing vehicle data not previously seen by the model). The primary metric is the temperature error curve: calculate the error at each time point, focusing on whether the error remains within 1°C throughout the entire 5400 seconds. Summary metrics such as MAE and RMSE can also be calculated. Special attention is paid to prediction errors in the initial few minutes (the period of fastest cooling) and the later stages approaching ambient temperature. If errors are found to be excessive in certain stages, the model structure and loss weights can be adjusted accordingly. Research is being conducted on adding Transformer layers to improve long-term memory and increasing loss weights for the initial stages to allow the model to more accurately fit the initial rapid temperature changes.

[0185] Test Set Validation: After the final model is determined, a final evaluation is performed on an independent test set (new vehicle model data that was not used in training). This verifies whether the model's prediction performance for the new vehicle model meets the target. The expected result is that the error of the entire cooling curve for most samples is within ±1℃, especially with accurate predictions at key time points (temperatures at 5 minutes, 30 minutes, and 90 minutes). Any errors exceeding 1℃ are recorded, and the reasons are analyzed (whether the vehicle model is unique, whether the input feature range exceeds the training distribution, etc.). If the test results do not meet the requirements, adjustments need to be made during the training phase (investigating increasing the data weights for this type of vehicle and retraining the model with a small amount of data from this vehicle model).

[0186] Through the above training and validation process, the study will obtain a model that is fully trained in terms of both temporal patterns and physical laws. This model not only fits well for known vehicle models in the training, but also has good predictive ability for unknown vehicle models, meeting the error requirements.

[0187] In some embodiments, after model training is completed, the deployment method for practical applications needs to be considered. As required, both server-side and vehicle-side deployments should be supported, and cloud-vehicle collaborative optimization should be possible to achieve continuous model improvement.

[0188] Deployment Options: If server-side deployment is chosen, the model can be encapsulated as a service (research REST API and RPC interfaces). Vehicles and users only need to upload the required input parameters (vehicle type and environmental conditions), and the server immediately returns the calculated 5400-second temperature prediction curve. Servers typically have strong computing power (CPU / GPU), capable of supporting more complex model structures, facilitating subsequent updates and maintenance.

[0189] For in-vehicle deployment, the performance and resources of the in-vehicle computing unit need to be considered. Edge device deployment is generally adopted, where the trained model is converted to a lightweight format (ONNX, TensorRT engine, and TensorFlow Lite model) using tools, and inference is run on the embedded hardware in the vehicle. It is necessary to ensure that the model size is moderate (the number of parameters is controlled within millions), and model compression techniques may be used to prune unimportant neurons and 8-bit quantized weights to reduce computation and storage requirements. The advantage of in-vehicle deployment is high real-time performance and no network dependency; however, the model must be sufficiently efficient, and a complete inference curve should be completed within seconds.

[0190] Input Interface: Regardless of the deployment, the interface should be simple and user-friendly. It should allow users to input parameter combinations for different vehicle models. For vehicle models, there can be a model identifier and a set of vehicle characteristic parameters; for operating conditions, outside temperature, sunlight intensity, and initial interior temperature can be input. When deployed on the vehicle side, these parameters can be provided in real time by in-vehicle sensors and systems (the air conditioning controller knows the set temperature and airflow, environmental sensors provide outside temperature, sunlight intensity can be estimated by sun position and brightness sensors, and vehicle speed is provided by the speedometer). On the server side, these parameters may be manually entered by the user or transmitted from other systems.

[0191] Output Demonstration: The 5400-second temperature curve output by the model can be transmitted to the in-vehicle infotainment system and cloud applications for display and further utilization. Research suggests plotting the predicted cooling curve on the in-vehicle central control screen to inform occupants of the approximate cabin temperature at various time points and to estimate in advance when a comfortable temperature will be reached. The output data can also be used as a reference for the air conditioning control algorithm to optimize airflow and compressor control strategies (this is a further application expansion).

[0192] Cloud-vehicle collaborative optimization: Deployment is not the end. Due to the possibility of encountering new situations not covered by training during actual use, continuous model optimization is necessary. Through cloud-vehicle collaboration, online model learning can be achieved (e.g., Figure 2 Feedback and online updates): Data feedback: During actual vehicle operation, the recorded real-world cooling curves and model predictions can be uploaded to the cloud (with passenger permission and under a privacy agreement). This is especially valuable when new models are first launched. The cloud collects new data from multiple vehicles and various environments, continuously enriching the database.

[0193] Model retraining: The model is periodically retrained and fine-tuned using accumulated new data in the cloud. For example, incremental training is triggered whenever a certain amount of new vehicle model data is collected, merging the new data into the training set to retrain the model. To avoid forgetting existing performance, incremental training and fine-tuning with a few epochs can be used to update model parameters, and validation on the old test set is performed to ensure no regression.

[0194] OTA Updates: Once the improved model is received in the cloud, it is deployed to the vehicle via OTA (Over-the-Air) technology. The vehicle receives the update and replaces the old model during idle periods and restarts. Version control and redundancy mechanisms ensure the update process is safe and reliable. If an update is unsatisfactory, it can be rolled back to the previous stable version.

[0195] Federated Learning (Optional): If direct data uploads raise privacy concerns, a federated learning approach can be adopted. Each vehicle trains a local update of its model using the new data, sending only the model weight updates (gradients) to the cloud for aggregation. The cloud aggregates the updates from all vehicles to improve the global model, then distributes the global model back to each vehicle. In this way, the original data doesn't leave the vehicles, but all vehicles collectively improve model performance.

[0196] Through the aforementioned collaborative mechanism, the model becomes increasingly intelligent with use: as more vehicle models and environmental data are added, its prediction accuracy will further improve, consistently remaining within 1°C of error, even when facing new vehicle models and conditions in the future. This cloud-edge collaboration also ensures that the model on each vehicle is always up-to-date and optimal, providing users with the best experience.

[0197] Based on the above method, the following can be achieved: Define the prediction task and performance metrics (e.g., a temperature curve over 5400 seconds with an error ≤1℃), and develop the model architecture and technical roadmap (LSTM+Transformer hybrid model, combining simulation, multi-task learning, etc.). Collect and integrate training samples from multiple (e.g., 100) different vehicle models. Organize the input features and output curves, ensure data quality, and divide the data into training / validation / test sets to prepare for subsequent training.

[0198] Data cleaning and feature engineering are performed, including handling missing and outlier values, smoothing noise, calculating derived features (cumulative heat load), and standardizing features and outputs. Appropriate data preprocessing improves model training effectiveness and stability. Furthermore, physical simulations (GT-suit & starccm+) are used to generate more samples, expanding the training set coverage. Simultaneously, interpolation and perturbation are applied to existing data to increase sample diversity, ensuring the physical validity of the augmented data and providing the model with richer learning materials.

[0199] Furthermore, a deep learning model was constructed according to the design architecture, defining the input layer (processing multi-source feature sequences), LSTM encoding / decoding layer, Transformer attention layer, and multi-task output layer within the framework. Training process code was written, and loss functions (mainly regression loss based on temperature curves, with additional physical constraint loss) and evaluation metrics were set.

[0200] For model training, a phased training approach is adopted: pre-training on simulated data (this strategy exists) followed by training on real data. The Adam optimizer and learning rate scheduling are used, along with Early Stopping to ensure optimal learning points are not missed. During training, the validation set error is continuously monitored, with particular attention paid to whether the maximum error drops below 1℃. Hyperparameters and training strategies are adjusted as needed (increasing the number of attention layers, adjusting loss weights, etc.) to meet accuracy requirements. Finally, model evaluation is performed on an independent test set, comparing the predicted curves with the actual curves and calculating error statistics. Verification is conducted to ensure that the error requirements are met across all vehicle models. Visualization is used to check the overlap between the predicted and measured temperature curves for several vehicles, ensuring no systematic bias or abnormal inaccuracies. Speed ​​and resource consumption are evaluated to confirm that the model can run efficiently in a real-world environment.

[0201] Model deployment is also conducted, with the deployment method selected based on application requirements. For server deployment, a backend service is built, the model is encapsulated into an API for invocation, and the input / output interface formats are designed. For vehicle deployment, the model is exported to an efficient format and integrated into the vehicle system's software architecture. Deployment testing is performed to simulate actual calls and verify prediction time and result accuracy. Edge computing scenarios are tested on the vehicle side to study long-term stability.

[0202] The model will be put into trial operation, and its predictive performance will be continuously monitored. Feedback data from real-world use will be collected, including the differences between the actual temperature curves and the model predictions. If errors are found to be close to or exceed 1°C in certain new models or under special environments, the reasons will be recorded and analyzed promptly.

[0203] Finally, online learning optimization is implemented, and a cloud-based collaborative mechanism is initiated to regularly retrain the model with new data. The improved model is updated to the server and vehicles via OTA (Over-The-Air) updates and release mechanisms. A cycle is maintained: data → training → deployment → data collection, continuously iterating to improve model performance and adapt to changes (such as improvements to the air conditioning system, differences between new energy vehicles and traditional vehicles). Each update is thoroughly tested and verified to ensure that the prediction error consistently meets or exceeds requirements.

[0204] Following the steps above, an innovative and feasible in-vehicle temperature prediction model will be developed. This model, based on a deep learning-physics fusion approach, features clear logic and a complete set of steps. In practical applications, it can predict temperature changes in different vehicle models with high accuracy and efficiency, providing strong support for optimizing automotive air conditioning systems and enhancing the passenger experience.

[0205] Figure 5 A schematic diagram of the temperature prediction device provided in an embodiment of this application is shown. The temperature prediction device is applied to an electronic device, which deploys a target deep learning model. The target deep learning model includes a cascaded long short-term memory encoder, a self-attention encoder, and a long short-term memory decoder, such as... Figure 5 As shown, the temperature prediction device 500 includes: The first acquisition module 501 is used to acquire a first feature vector set of the target crew cabin; wherein, the first feature vector set includes feature vectors corresponding to multiple time steps, and each feature vector includes temperature influence parameters of the target crew cabin at the corresponding time step; The second acquisition module 502 is used to input the first feature vector set into the long short-term memory encoder, and use the long short-term memory encoder to extract time series features from the temperature influence parameters included in the first feature vector set, and output the initial hidden state sequence; the initial hidden state sequence contains the time dependence and dynamic change pattern of the temperature influence parameters. The third acquisition module 503 is used to input the initial hidden state sequence into the self-attention encoder, use the self-attention encoder to perform global relation modeling on the initial hidden state sequence, and output the target hidden state sequence; the target hidden state sequence contains the long-term effect of temperature influence parameters on temperature; The fourth acquisition module 504 is used to input the target hidden state sequence into the long short-term memory decoder, use the long short-term memory decoder to perform temperature prediction based on the target hidden state sequence, and output the first predicted temperature curve corresponding to the target crew cabin.

[0206] In some embodiments, the temperature prediction device 500 further includes a model training module 505, specifically used for: acquiring a first training set; wherein the first training set includes experimental temperature training samples obtained based on experiments and simulated temperature training samples obtained based on simulations; each temperature training sample includes a second feature vector set of a crew cabin and a temperature curve label corresponding to the second feature vector set; and using the first training set to train the target deep learning model.

[0207] In some embodiments, the model training module 505 is specifically used to: generate crew cabin temperature curves for each operating condition using a crew cabin thermodynamic model; wherein the crew cabin thermodynamic model is constructed using a one-dimensional system simulation model and a three-dimensional computational fluid dynamics model; generate simulation temperature training samples corresponding to each operating condition based on each operating condition and the crew cabin temperature curves for each operating condition; wherein the feature vector of each simulation temperature training sample includes a temperature influence parameter for one time step of the corresponding operating condition, and the temperature curve label of each simulation temperature training sample is the crew cabin temperature curve for the corresponding operating condition.

[0208] In some embodiments, the model training module 505 is specifically configured to: input a temperature training sample from the first training set into the target deep learning model; use the target deep learning model to predict the temperature of the second feature vector set included in the temperature training sample, and output a second predicted temperature curve; calculate the mean square error between the second predicted temperature curve and the temperature curve label of the temperature training sample using a first loss function; calculate the product of the target squared value and the preset hyperparameters using a second loss function; the target squared value is the square of the difference between the temperature at the final moment and the ideal temperature at the final moment on the second predicted temperature curve; determine the target loss function value of the target deep learning model based on the mean square error and the product; if the target loss function value does not meet the training stopping condition, adjust the model parameters of the target deep learning model, and return to input a temperature training sample from the first training set into the target deep learning model until the target loss function value meets the training stopping condition.

[0209] In some embodiments, the model training module 505 is specifically used to: determine the first loss function value of the target deep learning model based on the mean square error and the product; and perform a weighted summation of the first loss function value and the second loss function value to obtain the target loss function value; wherein the second loss function value is determined by the auxiliary quantity curve labels included in the auxiliary quantity training samples and the predicted auxiliary quantity curve obtained by the target deep learning model based on the auxiliary quantity influence parameters included in the auxiliary quantity training samples, and the auxiliary quantity is a physical quantity that has an impact on the temperature of the crew cabin.

[0210] In some embodiments, the passenger compartment is the passenger compartment of a vehicle, and the temperature training samples also include vehicle model data; the model training module 505 is specifically used to: after inputting the temperature training samples into the target deep learning model, embed the vehicle model embedding vector corresponding to the vehicle model data into the target deep learning model; the vehicle model embedding vector is used to represent the vehicle model.

[0211] In some embodiments, the model training module 505 is specifically used to: after outputting the second predicted temperature curve corresponding to the temperature training sample, and before calculating the mean square error between the second predicted temperature curve and the temperature curve labels included in the temperature training sample through the first loss function, smooth the second predicted temperature curve if the second predicted temperature curve does not satisfy the preset monotonicity.

[0212] In some embodiments, the target passenger compartment is the passenger compartment of the target vehicle, and the model training module 505 is specifically used to: obtain a second training set before inputting the first feature vector set into the long short-term memory encoder, the second training set including multiple vehicle model data associated with the model of the target vehicle; and adjust the target deep learning model using the second training set.

[0213] The temperature prediction device 500 provided in this application embodiment can execute the technical solution shown in the above method embodiment. Its implementation principle and beneficial effects are similar, and will not be described again here.

[0214] Figure 6 A schematic diagram of the hardware structure of the electronic device provided in an embodiment of this application is shown.

[0215] An electronic device may include a processor 601 and a memory 602 storing computer program instructions.

[0216] Specifically, the processor 601 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0217] Memory 602 may include mass storage for data or instructions. For example, and not limitingly, memory 602 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 602 may include removable or non-removable (or fixed) media. Where appropriate, memory 602 may be internal or external to the integrated gateway disaster recovery device. In a particular embodiment, memory 602 is non-volatile solid-state memory.

[0218] In some embodiments, memory 602 may include read-only memory (ROM), random access memory (RAM), disk storage media device, optical storage media device, flash memory device, electrical, optical, or other physical / tangible memory storage device. Thus, generally, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the method according to one aspect of this disclosure.

[0219] If the electronic device can be the model training device described above, then the processor 601 can implement any of the model training methods in the above embodiments by reading and executing the computer program instructions stored in the memory 602.

[0220] If the electronic device can be the temperature prediction device described above, then the processor 601 can implement any of the temperature prediction methods in the above embodiments by reading and executing the computer program instructions stored in the memory 602.

[0221] In one example, the electronic device may also include a communication interface 603 and a bus 610. For example, Figure 6 As shown, the processor 601, memory 602, and communication interface 603 are connected through bus 610 and complete communication with each other.

[0222] The communication interface 603 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0223] Bus 610 includes hardware, software, or both, that couples components of an electronic device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 610 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, this application contemplates any suitable bus or interconnect.

[0224] The electronic device can execute the temperature prediction method in the embodiments of this application, thereby achieving a combination Figure 1 The temperature prediction method and apparatus are described in Figure 7.

[0225] In addition, this application embodiment also provides a computer storage medium for implementation. The computer storage medium stores computer program instructions; if the computer storage medium is located in an electronic device, the computer program instructions, when executed by a processor, implement the temperature prediction method described in the above embodiment.

[0226] This application also provides a computer program product, which, when executed by the processor of an electronic device, enables the electronic device to implement the temperature prediction method described above.

[0227] In conjunction with the electronic devices in the above embodiments, this application also provides a vehicle that includes the electronic devices in the above embodiments.

[0228] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0229] The functional blocks shown in the above-described structural diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0230] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0231] The aspects of this disclosure have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by special-purpose hardware performing the specified functions or actions, or can be implemented by a combination of special-purpose hardware and computer instructions.

[0232] The above description is merely a specific implementation of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A temperature prediction method, characterized in that, The method is applied to an electronic device, wherein the electronic device is equipped with a target deep learning model; the target deep learning model includes a cascaded long short-term memory encoder, a self-attention encoder, and a long short-term memory decoder; the method includes: Obtain a first feature vector set of the target passenger cabin; wherein, the first feature vector set includes feature vectors corresponding to multiple time steps, and each feature vector includes the temperature influence parameter of the target passenger cabin at the corresponding time step; The first feature vector set is input into the long short-term memory encoder, and the long short-term memory encoder is used to extract time series features from the temperature influence parameters included in the first feature vector set, and outputs an initial hidden state sequence; the initial hidden state sequence contains the time dependence and dynamic change pattern of the temperature influence parameters. The initial hidden state sequence is input into the self-attention encoder, and the self-attention encoder is used to perform global relation modeling on the initial hidden state sequence to output the target hidden state sequence; the target hidden state sequence contains the long-term effect of the temperature influence parameter on temperature; The target hidden state sequence is input into the long short-term memory decoder, and the long short-term memory decoder is used to predict the temperature based on the target hidden state sequence, and outputs the first predicted temperature curve corresponding to the target crew cabin.

2. The method according to claim 1, characterized in that, The training of the target deep learning model includes: Obtain a first training set; wherein the first training set includes experimental temperature training samples obtained based on experiments and simulated temperature training samples obtained based on simulation; each temperature training sample includes a second feature vector set of a crew cabin and a temperature curve label corresponding to the second feature vector set; The target deep learning model is trained using the first training set.

3. The method according to claim 2, characterized in that, The acquisition of the simulated temperature training samples includes: The crew cabin temperature curves under various operating conditions are generated by simulating a thermodynamic model of the crew cabin. The thermodynamic model of the crew cabin is constructed using a one-dimensional system simulation model and a three-dimensional computational fluid dynamics model. Based on each operating condition and the passenger compartment temperature curve for each operating condition, simulated temperature training samples corresponding to each operating condition are generated; wherein, the feature vector set of each simulated temperature training sample includes temperature influence parameters for multiple time steps of the corresponding operating condition, and the temperature curve label of each simulated temperature training sample is the passenger compartment temperature curve of the corresponding operating condition.

4. The method according to claim 2, characterized in that, The step of training the target deep learning model using the first training set until the target loss function value of the target deep learning model satisfies the training stopping condition includes: A temperature training sample from the first training set is input into the target deep learning model. The target deep learning model is used to predict the temperature of the second feature vector set included in the temperature training sample, and outputs a second predicted temperature curve. The mean squared error between the second predicted temperature curve and the temperature curve labels included in the temperature training samples is calculated using a first loss function. The product of the target squared value and the preset hyperparameter is calculated using the second loss function; the target squared value is the square of the difference between the temperature at the final moment on the second predicted temperature curve and the ideal temperature at the final moment. The target loss function value of the target deep learning model is determined based on the mean square error and the product. If the target loss function value does not meet the training stopping condition, adjust the model parameters of the target deep learning model to be trained, and return to the step of inputting a temperature training sample from the first training set into the target deep learning model until the target loss function value meets the training stopping condition.

5. The method according to claim 4, characterized in that, Determining the target loss function value of the target deep learning model based on the mean square error and the product includes: The first loss function value of the target deep learning model is determined based on the mean square error and the product. The first loss function value and the second loss function value are weighted and summed to obtain the target loss function value; wherein, the second loss function value is determined by the auxiliary quantity curve labels included in the auxiliary quantity training samples, and the predicted auxiliary quantity curve obtained by the target deep learning model based on the auxiliary quantity influence parameters included in the auxiliary quantity training samples, wherein the auxiliary quantity is a physical quantity that has an impact on the temperature of the crew cabin.

6. The method according to claim 4, characterized in that, The passenger compartment is the passenger compartment of a vehicle, and the temperature training samples also include vehicle model data. After inputting the temperature training samples into the target deep learning model, the method further includes: The vehicle model embedding vector corresponding to the vehicle model data is embedded into the target deep learning model; the vehicle model embedding vector is used to represent the vehicle model.

7. The method according to claim 4, characterized in that, After outputting the second predicted temperature curve corresponding to the temperature training sample, and before calculating the mean squared error between the second predicted temperature curve and the temperature curve labels included in the temperature training sample using the first loss function, the method further includes: If the predicted temperature curve does not meet the preset monotonicity, the predicted temperature curve is smoothed.

8. The method according to claim 1, characterized in that, The target passenger compartment is the passenger compartment of the target vehicle; Before inputting the first feature vector set into the long short-term memory encoder, the method further includes: Obtain a second training set, which includes data on multiple vehicle models associated with the model of the target vehicle; The target deep learning model is adjusted using the second training set.

9. A temperature prediction device, characterized in that, An application to an electronic device, wherein the electronic device is equipped with a target deep learning model; the target deep learning model includes a cascaded long short-term memory encoder, a self-attention encoder, and a long short-term memory decoder; the device includes: The first acquisition module is used to acquire a first feature vector set of the target crew cabin; wherein, the first feature vector set includes feature vectors corresponding to multiple time steps, and each feature vector includes the temperature influence parameter of the target crew cabin at the corresponding time step; The second acquisition module is used to input the first feature vector set into the long short-term memory encoder, and use the long short-term memory encoder to extract time series features from the temperature influence parameters included in the first feature vector set, and output an initial hidden state sequence; the initial hidden state sequence contains the time dependence and dynamic change pattern of the temperature influence parameters. The third acquisition module is used to input the initial hidden state sequence into the self-attention encoder, use the self-attention encoder to perform global relation modeling on the initial hidden state sequence, and output the target hidden state sequence; the target hidden state sequence includes the long-term effect of the temperature influence parameter on temperature; The fourth acquisition module is used to input the target hidden state sequence into the long short-term memory decoder, use the long short-term memory decoder to perform temperature prediction based on the target hidden state sequence, and output the first predicted temperature curve corresponding to the target crew cabin.

10. An electronic device, characterized in that, The electronic device includes: a processor and a memory storing computer program instructions; the processor, when executing the computer program instructions, implements the temperature prediction method as described in any one of claims 1-8.

Images