Method for predicting transient working condition of electromechanical composite transmission system by fusing physical information and Attention-LSTM (Long Short Term Memory) model
By integrating physical information with the Attention-LSTM model for prediction, the problems of model accuracy and interpretability under transient conditions of electromechanical composite transmission systems are solved, achieving high-precision and high-generalization prediction results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV
- Filing Date
- 2025-12-12
- Publication Date
- 2026-04-17
AI Technical Summary
In the prediction of transient operating conditions of electromechanical composite transmission systems, existing technologies suffer from modeling errors in physical information models, weak generalization ability of data-driven models under operating conditions outside of training data, poor model interpretability, and difficulty in ensuring consistency with physical laws.
A prediction method integrating physical information and an Attention-LSTM model is constructed. The model weights are dynamically adjusted through an adaptive fusion layer. An Attention-LSTM network is built by combining the torque balance equation and the power conservation law. Key features are selected by using mutual information and the maximum relevant information coefficient, thereby achieving the synergistic optimization of physical information and data-driven models.
It improves the prediction accuracy and dynamic trend capture capability under transient conditions, reduces model complexity, enhances the sensitivity and extraction capability of system instability precursor features, and realizes the strong generalization advantage of physical information models and the high precision advantage of data-driven models.
Smart Images
Figure CN121880768A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of state prediction technology for electromechanical composite transmission systems, and relates to a method for predicting transient operating conditions of electromechanical composite transmission systems by integrating physical information and Attention-LSTM model. Background Technology
[0002] As the core power transmission device for heavy equipment such as special vehicles, the electromechanical composite transmission system is a complex electromechanical system composed of deeply coupled multiple energy components such as engine-generator set, drive motor, energy storage unit and power converter. Under transient high-power conditions such as rapid acceleration, climbing, and turning, this system is prone to dynamic response mismatch between the front and rear power chains, which can lead to problems such as DC bus voltage oscillation and engine speed instability, and in severe cases, even loss of system function.
[0003] Currently, there are two main types of methods for predicting transient operating conditions in transmission systems: The first is a modeling method based on physical information, which uses physical laws such as energy conservation and torque balance to construct analytical models. While this type of method has clear physical definitions and interpretability, it often suffers from model simplification and limited prediction accuracy under strongly nonlinear, multi-timescale coupled transient operating conditions. The second type is a data-driven modeling method, such as using network models like Long Short-Term Memory (LSTM) networks, which can extract dynamic features from historical time-series data. However, this type of method has weak generalization ability under operating conditions outside the training data, poor model interpretability, and difficulty in ensuring consistency with physical laws.
[0004] Therefore, there is an urgent need for a prediction method that can integrate physical information with data-driven models. Summary of the Invention
[0005] In view of this, the purpose of this invention is to provide a method for predicting transient operating conditions of electromechanical composite transmission systems by integrating physical information and Attention-LSTM models.
[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting transient operating conditions of an electromechanical hybrid transmission system by integrating physical information and an Attention-LSTM model includes the following steps: S1: Collect multi-source time-series data during the operation of the electromechanical composite transmission system, preprocess and select key features to construct a time-series prediction dataset containing transient high-power conditions; S2: Construct a transient condition prediction model that integrates physical information and an Attention-LSTM network, including a Long Short-Term Memory (LSTM) network, an attention mechanism layer, a physical information model, and an adaptive fusion layer; the physical information model is based on the torque balance equation and power conservation law of the electromechanical composite transmission system to calculate the engine speed change trend; the adaptive fusion layer dynamically adjusts the fusion weights of the physical information model and the neural network model based on the engine speed deviation and speed change rate. S3: Train and test the transient condition prediction model using the time-series prediction dataset; S4: Use the trained transient condition prediction model to predict the transient conditions of the electromechanical composite transmission system.
[0007] Furthermore, the multi-source timing data during the operation of the electromechanical composite transmission system described in step S1 includes bus voltage, bus current, engine speed, and engine throttle opening.
[0008] Furthermore, step S1, which involves preprocessing the multi-source time-series data and then filtering out key features, includes: S11: Use a linear interpolation function to fill in the missing values in the original time series data and remove outliers based on power conservation constraints;
[0009] S12: A hybrid strategy based on mutual information and maximum relevant information coefficient is adopted to screen key features. The nonlinear dependency between each collected feature and the prediction target is quantified by calculating mutual information, and the correlation strength is evaluated by using the maximum relevant information coefficient. Finally, key features that contribute significantly to the prediction model are screened based on the set joint threshold. S13: Divide the data into training and testing sets, and normalize all key features.
[0010] Furthermore, the transient condition prediction model described in step S2 adopts a two-layer stacked LSTM structure. The first layer of LSTM extracts local temporal features, the second layer of LSTM captures long-term dependencies, and the attention mechanism layer calculates the attention weights of the features at each time step through a fully connected layer, outputting the temporal prediction result.
[0011] Furthermore, the physical information model is constructed based on the torque balance equation and power conservation law of the electromechanical composite transmission system, and the physical calculation relationship of the engine angular acceleration is as follows:
[0012] in, The fitting coefficients for engine torque versus throttle opening are... This is the generator torque-bus power correlation coefficient. The total moment of inertia coefficient, This refers to the engine throttle opening. Bus voltage Bus current, This refers to the actual engine speed. To predict the rotational speed using physics, For engine angular acceleration, To predict the time step.
[0013] Furthermore, the adaptive fusion layer uses a balance coefficient to comprehensively calculate the fusion weights of the speed deviation factor and the speed change rate factor, as shown in the following formula:
[0014] in, For speed deviation factor, The rotational speed change rate factor, the fusion weight σ is obtained through the balance coefficient. calculate, This is the model's overall prediction output. This is the adjustment coefficient for the speed deviation factor. This is the offset of the speed deviation factor. This is the adjustment coefficient for the rate of change of rotational speed. This represents the offset of the rotational speed change rate factor.
[0015] Furthermore, the steps for training the transient condition prediction model in step S3 are as follows: S31: Based on the training set, the physical information model is pre-trained, and the physical parameters are fitted by the least squares method to ensure that the output rotation speed change trend conforms to the actual operation law. S32: Fix the pre-trained physical parameters, train the Attention-LSTM model, and optimize the model weights by adjusting the hyperparameters with the goal of minimizing the temporal prediction loss function; S33: Remove the fixed constraints on the physical parameters and Attention-LSTM parameters, and achieve collaborative optimization of the physical information model and the Attention-LSTM model through adaptive weight adjustment and composite loss function design, thus completing the training of the fusion model.
[0016] Furthermore, step S31 specifically includes the following steps: Will , , , As input features for the physical information model; The least squares method is used to minimize the sum of squared residuals between the physical model's predicted and actual values, and the physical parameters are then fitted. When the angular acceleration prediction residual of the physical model When the residual change is less than 0.001 over 10 consecutive training rounds, pre-training is stopped and the value is fixed. , , parameter.
[0017] Furthermore, step S32 specifically includes the following steps: Set the hyperparameters of the Attention-LSTM model as follows: the number of hidden layers is set to 2, the number of hidden units is 128 and 64 respectively, the number of iterations is set to 200-300, the initial learning rate is set to 0.001, and it decays to 0.8 of the previous round every 50 rounds. Loss prediction using time series To optimize the objective, a loss function is designed; The training termination condition is when the training set loss 20 consecutive rounds of decline At 0.001, stop training and save the weight matrix of the Attention-LSTM model. With bias .
[0018] Furthermore, step S33 specifically includes the following steps: To achieve a collaborative optimization objective, balancing temporal fitting accuracy, physical consistency, and model generalization, a composite loss function for the fusion model is constructed. The calculation formula is:
[0019] in, The temporal prediction loss of the Attention-LSTM model. For time series prediction loss weights, The relative error loss of the physical information model. Weights for physical information loss. For regularization loss, For regularization weights; During training, the weights are dynamically adjusted, and the adaptive fusion layer is based on real-time... , , Update σ; for transient conditions, σ approaches 1, so prioritize the physical information model; for steady-state conditions, σ approaches 0, so prioritize the Attention-LSTM model. Simultaneous fine-tuning of physical parameters using gradient descent method , , With Attention-LSTM parameters 、 until 20 consecutive rounds of decline 0.001.
[0020] The beneficial effects of this invention are as follows: 1) This invention constructs a fusion prediction framework of physical information model and attention-LSTM model. By dynamically balancing the output of the two models through an adaptive fusion layer, it not only utilizes the strong generalization and interpretability of physical information model under transient conditions, but also leverages the high accuracy advantage of data-driven model in complex nonlinear mapping.
[0021] 2) By introducing a key feature screening method based on mutual information and maximum relevant information coefficient, this invention effectively eliminates redundant features with weak correlation to the prediction target, reduces model complexity, and enhances the model's sensitivity and extraction ability to system instability precursor features.
[0022] 3) This invention designs a phased collaborative training strategy. Through adaptive weight adjustment and composite loss function design, it realizes the collaborative optimization of the physical information model and the Attention-LSTM model, thereby improving the steady-state prediction accuracy and dynamic trend capture capability of the fusion model under transient conditions.
[0023] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0024] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a schematic diagram of the fusion prediction framework of the physical information model and the Attention-LSTM model. Figure 2 This is a flowchart for preprocessing and key feature selection of multi-source time series data. Detailed Implementation
[0025] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.
[0026] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.
[0027] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.
[0028] Example 1: Please see Figures 1-2 This invention constructs a fusion prediction framework combining a Physical Information Model (PIMMA) and an Attention-LSTM model. By dynamically balancing the outputs of the two models through an adaptive fusion layer, it leverages both the strong generalization and interpretability of the PIMMA under transient conditions and the high accuracy of the data-driven model in complex nonlinear mappings. By introducing a key feature selection method based on mutual information and maximum relevant correlation coefficients, redundant features with weak correlation to the prediction target are effectively eliminated, reducing model complexity and enhancing the model's sensitivity and extraction ability for system instability precursors. A phased collaborative training strategy is designed, achieving collaborative optimization of the PIMMA and Attention-LSTM model through adaptive weight adjustment and composite loss function design, thereby improving the fusion model's steady-state prediction accuracy and dynamic trend capture capability under transient conditions.
[0029] The transient condition prediction method for electromechanical composite transmission systems that integrates physical information and the Attention-LSTM model, as described in this invention, comprises the following steps: S1: Collect multi-source time-series data during the operation of the electromechanical composite transmission system, including bus voltage, bus current, engine speed, engine throttle opening, etc. Preprocess the raw time-series data, use a method based on mutual information and maximum relevant information coefficient to screen key features, construct a time-series prediction dataset containing transient high-power conditions, and divide the dataset into training and testing sets; specifically including the following steps: S11: Use a linear interpolation function to complete the missing values in the original time series data, based on the power conservation constraint (engine output power). With bus power deviation 15%) outliers were removed;
[0030] S12: A hybrid strategy based on mutual information and maximum relevant information coefficient is adopted to screen key features. The nonlinear dependence between each collected feature and the prediction target is quantified by calculating mutual information, and the correlation strength is further evaluated by using the maximum relevant information coefficient. Finally, key features that contribute significantly to the prediction model are screened based on the set joint threshold. S13: Divide the training set and the test set into an 8:2 ratio and normalize all key features.
[0031] S2: Using a Long Short-Term Memory (LSTM) network as the core, a physical information model and an adaptive fusion layer are introduced to construct a prediction model that integrates physical information and Attention-LSTM. The physical information model is based on the torque balance equation and power conservation law of the electromechanical composite transmission system to calculate the engine speed change trend. The adaptive fusion layer dynamically adjusts the fusion weights of the physical information model and the neural network model based on the engine speed deviation and speed change rate. The specific steps are as follows: S21: Introduce an attention mechanism layer to construct an Attention-LSTM model. A two-layer stacked LSTM structure is adopted. The first LSTM layer extracts local temporal features, and the second LSTM layer captures long-term dependencies. The Attention layer calculates the attention weights of features at each time step through a fully connected layer and outputs the temporal prediction results.
[0032] S22: Based on the torque balance equation and power conservation law of the electromechanical composite transmission system, a physical information model is constructed. The physical calculation relationship for engine angular acceleration is established, and the calculation formula is as follows:
[0033] in, The fitting coefficients for engine torque versus throttle opening are... This is the generator torque-bus power correlation coefficient. The total moment of inertia coefficient, This refers to the engine throttle opening. This refers to the actual engine speed. To predict the rotational speed using physics, For engine angular acceleration, To predict the time step.
[0034] S23: Adaptive Fusion Layer: A speed deviation factor and a speed change rate factor are designed separately. A balance coefficient is introduced to comprehensively calculate the fusion weight of the two factors. The specific calculation formula is as follows:
[0035] in, For speed deviation factor, The rotational speed change rate factor, the fusion weight σ is obtained through the balance coefficient. calculate, This is the model's overall prediction output. This is the adjustment coefficient for the speed deviation factor. This is the offset of the speed deviation factor. This is the adjustment coefficient for the rate of change of rotational speed. This represents the offset of the rotational speed change rate factor.
[0036] S3: Based on the training set constructed in step S1, the physical information model is pre-trained, and the physical parameters are fitted using the least squares method to ensure that the output rotational speed change trend conforms to the actual operating law; the specific steps are as follows: S31: Select the option filtered in step S12. , , , As input features for the physical information model; S32: The least squares method is used to minimize the sum of squared residuals between the predicted and actual values of the physical model, and the physical parameters are fitted.
[0037] S32: When the angular acceleration prediction residual of the physical model When the residual change is less than 0.001 over 10 consecutive training rounds, pre-training is stopped and the value is fixed. , , parameter.
[0038] S4: Fix the physical parameters obtained from pre-training in step S3, and train the Attention-LSTM model. Optimize the model weights by adjusting hyperparameters such as the number of hidden layers, hidden units, iterations, and learning rate, with the goal of minimizing the temporal prediction loss function. This includes the following steps: S41: Set the hyperparameters of the Attention-LSTM model: 2 hidden layers with 128 and 64 hidden units respectively, 200-300 iterations, and an initial learning rate of 0.001, decaying to 0.8 every 50 iterations. S42: Predicting Losses Using Time Series To optimize the objective, a loss function is designed.
[0039] S43: Training termination condition: When the training set is lost... 20 consecutive rounds of decline At 0.001, stop training and save the weight matrix of the Attention-LSTM model. With bias .
[0040] S5: Remove the fixed constraints on the physical parameters and Attention-LSTM parameters, and achieve collaborative optimization of the physical information model and the Attention-LSTM model through adaptive weight adjustment and composite loss function design, thus completing the training of the fusion model; specifically including the following steps: S51: Co-optimization objective, balancing time series fitting accuracy, physical consistency, and model generalization, constructing a composite loss function for the fusion model. The calculation formula is:
[0041] in, The temporal prediction loss of the Attention-LSTM model. For time series prediction loss weights, The relative error loss of the physical information model. Weights for physical information loss. For regularization loss, For regularization weights.
[0042] S52: During training, weights are dynamically adjusted, and the adaptive fusion layer is based on real-time... , , Update σ. For transient conditions, σ approaches 1, so prioritize the physical information model; for steady-state conditions, σ approaches 0, so prioritize the Attention-LSTM model.
[0043] S53: Simultaneous fine-tuning of physical parameters using gradient descent method ( , , ) and Attention-LSTM parameters ( 、 ), until 20 consecutive rounds of decline 0.001.
[0044] S6: Verify the performance of the fusion model based on the test set constructed in step S1. Evaluate the model's steady-state prediction accuracy and dynamic trend capture capability under transient conditions using mean absolute error, short-time root mean square error, and long-time root mean square error. The specific steps are as follows: S61: Based on the test set constructed in step S1, the performance of the fusion model is quantitatively evaluated using three key indicators: mean absolute error, short-time root mean square error, and long-time root mean square error.
[0045] S62: The fusion model is compared with the Long Short-Term Memory Network (LSTM) model and the Autoregressive Differential Moving Average (ARIMA) model. Under the same test set and the same evaluation index, the improvement effect of the fusion model in dynamic trend capture and steady-state accuracy under transient conditions is verified.
[0046] Example 2: An electronic device, comprising a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the method described in Embodiment 1 when executing the computer program.
[0047] Example 3: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in Embodiment 1.
[0048] Example 4: A computer program product includes a computer program that, when executed by a processor, implements the method described in Example 1.
[0049] In the above embodiments, the reference to "this embodiment" in the specification indicates that a specific feature, structure, or characteristic described in connection with the embodiment is included in at least some embodiments, but not necessarily all embodiments. Multiple appearances of "this embodiment" do not necessarily all refer to the same embodiment.
[0050] In the above embodiments, although the invention has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory structures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed. The embodiments of the invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims.
[0051] As will be understood by those skilled in the art, the computer-readable storage medium described in this embodiment allows for the implementation of all or part of the steps in the above method embodiments by computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.
[0052] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication between them. The memory is used to store computer programs, the communication interface is used to perform communication, and the processor and the transceiver are used to run the computer programs, so that the electronic terminal performs the steps of the above method.
[0053] In this embodiment, the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.
[0054] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.
[0055] This invention can be used in a wide range of general-purpose or special-purpose computing system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.
[0056] This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.
[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for predicting transient operating conditions of an electromechanical composite transmission system by integrating physical information and an Attention-LSTM model, characterized in that: Includes the following steps: S1: Collect multi-source time-series data during the operation of the electromechanical composite transmission system, preprocess and select key features to construct a time-series prediction dataset containing transient high-power conditions; S2: Construct a transient condition prediction model that integrates physical information and an Attention-LSTM network, including a Long Short-Term Memory (LSTM) network, an attention mechanism layer, a physical information model, and an adaptive fusion layer; the physical information model is based on the torque balance equation and power conservation law of the electromechanical composite transmission system to calculate the engine speed change trend; the adaptive fusion layer dynamically adjusts the fusion weights of the physical information model and the neural network model based on the engine speed deviation and speed change rate. S3: Train and test the transient condition prediction model using the time-series prediction dataset; S4: Use the trained transient condition prediction model to predict the transient conditions of the electromechanical composite transmission system.
2. The method for predicting transient operating conditions of an electromechanical composite transmission system by fusing physical information and an Attention-LSTM model according to claim 1, characterized in that: The multi-source timing data during the operation of the electromechanical composite transmission system described in step S1 includes bus voltage, bus current, engine speed, and engine throttle opening.
3. The method for predicting transient operating conditions of an electromechanical composite transmission system by fusing physical information and an Attention-LSTM model according to claim 1, characterized in that: Step S1 involves preprocessing multi-source time-series data and then filtering out key features, including: S11: Use a linear interpolation function to fill in the missing values in the original time series data and remove outliers based on power conservation constraints; S12: A hybrid strategy based on mutual information and maximum relevant information coefficient is adopted to screen key features. The nonlinear dependency between each collected feature and the prediction target is quantified by calculating mutual information, and the correlation strength is evaluated by using the maximum relevant information coefficient. Finally, key features that contribute significantly to the prediction model are screened based on the set joint threshold. S13: Divide the data into training and testing sets, and normalize all key features.
4. The method for predicting transient operating conditions of an electromechanical composite transmission system by fusing physical information and an Attention-LSTM model according to claim 1, characterized in that: The transient condition prediction model described in step S2 adopts a two-layer stacked LSTM structure. The first layer of LSTM extracts local temporal features, the second layer of LSTM captures long-term dependencies, and the attention mechanism layer calculates the attention weights of features at each time step through a fully connected layer, and outputs the temporal prediction results.
5. The method for predicting transient operating conditions of an electromechanical composite transmission system by fusing physical information and an Attention-LSTM model according to claim 1, characterized in that: The physical information model is constructed based on the torque balance equation and power conservation law of the electromechanical composite transmission system. The physical calculation relationship of the engine angular acceleration is as follows: in, The fitting coefficients for engine torque versus throttle opening are... This is the correlation coefficient between generator torque and bus power. The total moment of inertia coefficient, This refers to the engine throttle opening. Bus voltage Bus current, This refers to the actual engine speed. To predict the rotational speed using physics, For engine angular acceleration, To predict the time step.
6. The method for predicting transient operating conditions of an electromechanical composite transmission system by fusing physical information and an Attention-LSTM model according to claim 1, characterized in that: The adaptive fusion layer uses a balance coefficient to comprehensively calculate the fusion weights of the speed deviation factor and the speed change rate factor, as shown in the following formula: in, For speed deviation factor, The rotational speed change rate factor, the fusion weight σ is obtained through the balance coefficient. calculate, This is the model's overall prediction output. This is the adjustment coefficient for the speed deviation factor. This is the offset of the speed deviation factor. This is the adjustment coefficient for the rate of change of rotational speed. This represents the offset of the rotational speed change rate factor.
7. The method for predicting transient operating conditions of an electromechanical composite transmission system by fusing physical information and an Attention-LSTM model according to claim 1, characterized in that: The steps for training the transient condition prediction model in step S3 are as follows: S31: Based on the training set, the physical information model is pre-trained, and the physical parameters are fitted by the least squares method to ensure that the output rotation speed change trend conforms to the actual operation law. S32: Fix the pre-trained physical parameters, train the Attention-LSTM model, and optimize the model weights by adjusting the hyperparameters with the goal of minimizing the temporal prediction loss function; S33: Remove the fixed constraints on the physical parameters and Attention-LSTM parameters, and achieve collaborative optimization of the physical information model and the Attention-LSTM model through adaptive weight adjustment and composite loss function design, thus completing the training of the fusion model.
8. The method for predicting transient operating conditions of an electromechanical composite transmission system by fusing physical information and an Attention-LSTM model according to claim 7, characterized in that: Step S31 specifically includes the following steps: Will , , , As input features for the physical information model; The least squares method is used to minimize the sum of squared residuals between the physical model's predicted and actual values, and the physical parameters are then fitted. When the angular acceleration prediction residual of the physical model When the residual change is less than 0.001 over 10 consecutive training rounds, pre-training is stopped and the value is fixed. , , parameter.
9. The method for predicting transient operating conditions of an electromechanical composite transmission system by fusing physical information and an Attention-LSTM model according to claim 8, characterized in that: Step S32 specifically includes the following steps: Set the hyperparameters of the Attention-LSTM model as follows: number of hidden layers: 2; number of hidden units: 128 and 64 respectively; number of iterations: 200-300; initial learning rate: 0.001; decay to 0.8 of the previous learning rate every 50 iterations. Loss prediction using time series To optimize the objective, a loss function is designed; The training termination condition is when the training set loss 20 consecutive rounds of decline At 0.001, stop training and save the weight matrix of the Attention-LSTM model. With bias .
10. The method for predicting transient operating conditions of an electromechanical composite transmission system by fusing physical information and an Attention-LSTM model according to claim 9, characterized in that: Step S33 specifically includes the following steps: To achieve a collaborative optimization objective, balancing temporal fitting accuracy, physical consistency, and model generalization, a composite loss function for the fusion model is constructed. The calculation formula is: in, The temporal prediction loss of the Attention-LSTM model. For time series prediction loss weights, The relative error loss of the physical information model. Weights for physical information loss. For regularization loss, For regularization weights; During training, the weights are dynamically adjusted, and the adaptive fusion layer is based on real-time... , , Update σ; for transient conditions, σ approaches 1, so prioritize the physical information model; for steady-state conditions, σ approaches 0, so prioritize the Attention-LSTM model. Simultaneous fine-tuning of physical parameters using gradient descent method , , With Attention-LSTM parameters 、 until 20 consecutive rounds of decline 0.001.