Modelica proxy model generation method, device and medium

Abstract

The invention discloses a Modelica proxy model generation method and device and a medium, and relates to the technical field of engineering simulation, and the method comprises the steps: obtaining an initialization parameter of a Modelica target model, and determining a value range and probability distribution of the initialization parameter; constructing a training data set according to the value range and probability distribution of the initialization parameters; constructing an LSTM neural network model, and determining network parameters of the LSTM neural network model after random initialization; performing training optimization on network parameters of the LSTM neural network model by adopting an MAML algorithm according to the training data set to obtain optimized network parameters; according to the optimized network parameters, constructing an optimized meta-model; and performing small sample training on the optimized meta-model to obtain a Modelica proxy model. According to the method, the solution speed of the Modelica model can be effectively improved, and the modeling simulation efficiency is improved.

Description

Technical Field

[0001] The present application relates to the field of engineering simulation technology, and in particular to a Modelica proxy model generation method, device and medium. Background Art

[0002] In the current field of engineering simulation, Modelica (an open, object-oriented, equation-based computer language) is widely used in the modeling and simulation of multi-domain systems as an efficient modeling and simulation language. However, with the increasing complexity of the system, the speed of solving the Modelica model has become one of the key factors restricting its application. In particular, when performing real-time simulation or optimization design, higher requirements are placed on the speed of solving the Modelica model.

[0003] On the other hand, in actual application, Modelica models often contain a large number of parameters and variables, and the interactions between these parameters and variables are complex, which makes the model solution process complicated and time-consuming. At the same time, with the increase in system complexity, traditional numerical solution methods often cannot meet the needs of real-time simulation or optimization design. Especially in the construction of multi-physics field models, it often involves solving high-order differential equations. For example, the Navier-Stokes equation in fluid mechanics, the Schrodinger equation in quantum mechanics, and the wave equation in electromagnetism are all high-order differential equations. When solving these equations numerically, a large number of gradient calculations are involved, and the solution speed is slow and the efficiency is low.

[0004] Therefore, how to improve the solution speed of the Modelica model and thus improve the efficiency of modeling and simulation has become a technical problem that needs to be solved urgently in this field. Summary of the invention

[0005] The purpose of this application is to provide a Modelica proxy model generation method, device and medium, which can effectively improve the solution speed of the Modelica model and improve the efficiency of modeling and simulation.

[0006] To achieve the above objectives, this application provides the following solutions.

[0007] In a first aspect, the present application provides a Modelica proxy model generation method, and the Modelica proxy model generation method includes the following steps.

[0008] The initialization parameters of the Modelica target model are obtained, and the value range and probability distribution of the initialization parameters are determined; the Modelica target model is the original model for which the proxy model needs to be generated, and the initialization parameters include initial conditions, boundary conditions, physical property parameters and penalty coefficients.

[0009] A training data set is constructed according to the value range and probability distribution of the initialization parameters.

[0010] Construct an LSTM neural network model, and determine network parameters of the LSTM neural network model after random initialization.

[0011] According to the training data set, the MAML algorithm is used to train and optimize the network parameters of the LSTM neural network model to obtain optimized network parameters.

[0012] An optimized meta-model is constructed according to the optimized network parameters; the optimized meta-model refers to a model obtained by replacing the randomly initialized network parameters with the optimized network parameters.

[0013] The optimized meta-model is trained with a small sample to obtain a Modelica proxy model corresponding to the Modelica target model; the Modelica proxy model is used for computational solution of simulation tasks, and the solution speed of the Modelica proxy model is greater than the solution speed of the Modelica target model.

[0014] In a second aspect, the present application provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any of the above-described Modelica proxy model generation methods.

[0015] In a third aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described Modelica proxy model generation methods.

[0016] According to the specific embodiments provided in this application, this application discloses the following technical effects: The present application provides a method, device and medium for generating a Modelica proxy model. By combining the MAML algorithm with the LSTM neural network model, the Modelica proxy model can be generated, which can effectively save the steps of compiling and optimizing the Modelica model by the compilation and solution engine, reduce the engine's computational workload, speed up the model's running speed, and thus improve the model's solution speed. In addition, by introducing the MAML algorithm, based on the meta-learning method in the MAML algorithm, the Modelica proxy model can achieve rapid adaptation and rapid solution capabilities for new tasks, reduce the model's dependence on the amount of sample data, reduce the Modelica engine's computational workload for batch simulation, and improve the efficiency of modeling and simulation. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 A flowchart of a method for generating a Modelica proxy model provided in one embodiment of the present application.

[0019] Figure 2 A schematic diagram of the principle of a Modelica proxy model generation method provided in one embodiment of the present application. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0021] like Figure 1 As shown, this embodiment proposes a Modelica proxy model generation method, and the Modelica proxy model generation method specifically includes the following steps.

[0022] Step S1, obtaining the initialization parameters of the Modelica target model, and determining the value range and probability distribution of the initialization parameters.

[0023] In this embodiment, the Modelica target model is the original model for which a proxy model needs to be generated.

[0024] Modelica is an open, object-oriented, equation-based computer language that can model complex physical systems across domains, including mechanical, electronic, electrical, hydraulic, thermal, control, and process-oriented subsystem models.

[0025] From a macroscopic perspective, any model designed based on the Modelica computer language can be called a Modelica model, and at different stages, the Modelica model includes the Modelica target model, the Modelica mechanism model, and the Modelica proxy model, wherein the Modelica target model refers to the original model before data processing, which has not been processed; the Modelica mechanism model refers to the model in the data processing stage, which performs data processing and mechanism learning based on the Modelica target model; the Modelica proxy model refers to the stage after data processing is completed, which generates a corresponding proxy model for the Modelica target model. The ultimate goal of this embodiment is to output a Modelica proxy model corresponding to the Modelica target model.

[0026] In this embodiment, the Modelica target model refers to an original model designed based on the Modelica computer language for which a proxy model needs to be generated, and the initialization parameters include initial conditions, boundary conditions, physical property parameters, penalty coefficients, and the like.

[0027] Step S2: construct a training data set according to the value range and probability distribution of the initialization parameters.

[0028] In this embodiment, step S2 constructs a training data set according to the value range and probability distribution of the initialization parameters, and specifically includes the following steps.

[0029] Step S21: randomly generate a set of basic model parameters according to the value range and probability distribution of the initialization parameters.

[0030] Step S22: construct a number of simulation tasks according to the basic model parameters, and input the several simulation tasks into the Modelica simulation software for batch simulation calculation to obtain solution results.

[0031] Step S23: normalize the solution result to obtain a normalized solution result.

[0032] Step S24: construct a training data set based on the solution results after normalization processing.

[0033] Step S3: construct an LSTM (Long Short-Term Memory) neural network model, and determine the network parameters of the LSTM neural network model after random initialization.

[0034] In this embodiment, the gating unit of the LSTM neural network model is a Sigmoid (S-type) activation function, the state update and hidden state use a Tanh (Tangent hyperbolic) activation function, the optimizer is an Adam function, and the loss function is a Mean Squared Error (MSE) function.

[0035] The LSTM neural network model is a special RNN (Recurrent Neural Network). By introducing network structures such as memory cells, input gates, output gates, and forget gates, it effectively solves the problem of long-term dependencies in time series. Compared with traditional RNNs, LSTM solves the problems of gradient disappearance and gradient explosion, and is more suitable for processing and predicting important events with long intervals in time series. LSTM is mainly used in text generation, speech recognition, machine translation, and time series prediction.

[0036] Step S4: According to the training data set, the MAML (Model-Agnostic Meta-Learning) algorithm is used to train and optimize the network parameters of the LSTM neural network model to obtain the optimized network parameters.

[0037] In this embodiment, the MAML algorithm mainly includes an inner loop optimization phase and an outer loop optimization phase, which correspond to the inner loop optimization step and the outer loop optimization step respectively. The inner loop optimization phase processes the basic model, focusing on the rapid adaptation of a single basic model. The specific steps include task sampling, gradient calculation, and parameter update. The outer loop optimization phase processes the meta-model, focusing on improving the initial parameters of the meta-model through learning multiple tasks. , in order to improve the generalization ability of the model on new tasks. The specific steps include echelon aggregation and parameter updating.

[0038] Meta-learning refers to the process of improving learning algorithms in multiple learning stages, that is, using past knowledge and experience to guide the learning of new tasks, and having the ability to learn to learn. Its purpose is to learn a learning method or meta-knowledge from existing tasks, so as to improve the learning efficiency of new tasks. The core idea of ​​MAML is to optimize the model initialization parameters so that the model can quickly learn and adapt when faced with new tasks. The goal of MAML is to find the optimal initialization parameters. For new tasks, a small number of samples and a small number of gradient update steps can achieve better solution performance.

[0039] In this embodiment, step S4 uses the MAML algorithm to train and optimize the network parameters of the LSTM neural network model according to the training data set to obtain the optimized network parameters, which specifically includes the following steps.

[0040] Step S41, according to the training data set, calculate the loss function of the basic model, and use the gradient descent method to update the network parameters of the basic model; the basic model is the model corresponding to the LSTM neural network model in the inner loop optimization stage of the MAML algorithm.

[0041] Step S42: sum the loss functions of all basic models, and use the gradient descent method to update the network parameters of the meta-model to obtain the optimized network parameters. The meta-model is the model corresponding to the LSTM neural network model in the outer loop optimization stage of the MAML algorithm.

[0042] Step S5: construct an optimized meta-model according to the optimized network parameters, wherein the optimized meta-model refers to a model obtained by replacing the randomly initialized network parameters with the optimized network parameters.

[0043] In this embodiment, step S5 constructs an optimized meta-model according to the optimized network parameters, which specifically includes the following steps.

[0044] The randomly initialized network parameters of the LSTM neural network model are replaced with the optimized network parameters to obtain the optimized meta-model.

[0045] Step S6: Perform small sample training on the optimized meta-model to obtain a Modelica proxy model corresponding to the Modelica target model.

[0046] In this embodiment, the Modelica proxy model is used for computational solutions of various types of simulation tasks, and the solution speed of the Modelica proxy model is greater than the solution speed of the Modelica target model, which can significantly improve the solution speed and efficiency of the model.

[0047] In this embodiment, step S6 performs small sample training on the optimized meta-model to obtain a Modelica proxy model corresponding to the Modelica target model, which specifically includes the following steps.

[0048] Step S61: numerically calculate and solve the Modelica target model to construct a small sample training data set.

[0049] Step S62: Use the small sample training data set to perform small sample training on the optimized meta-model to obtain a Modelica proxy model.

[0050] In order to make the technical solution of this embodiment clearer, the specific implementation process of the technical solution of this embodiment is described in detail below in the form of examples. Figure 2 As shown, the specific implementation steps include the following.

[0051] Step 1: Identify and extract the initialization parameters (such as the initial conditions, boundary conditions, physical parameters, and penalty coefficients of the system) of the Modelica target model for which the proxy model needs to be generated, and confirm the value range and probability distribution of each initialization parameter .

[0052] For example, for a vehicle dynamics model, the initialization parameters of the model may include parameters such as vehicle speed and body steering angle, whose value range can be considered to be between 0-120km / h and between 30° and 50°, and the probability distribution can be considered to obey uniform distribution or normal distribution.

[0053] Step 2: Based on the value range and probability distribution , randomly generate a set (e.g. Basic model parameters , using this set of basic model parameters Construct a batch of simulation tasks, perform batch simulation calculations in the Modelica simulation software, obtain solution results, normalize the solution results, and construct a training data set.

[0054] In this embodiment, taking the above vehicle dynamics model as an example, it is assumed that the value range of the vehicle speed is , steering angle , and obeys uniform distribution, then sampling N=5 sets of initialization parameters, the sampling results may be , ,Based on the above parameters, a batch of initialization parameter combinations for simulation tasks can be constructed, such as , , , , ,Each parameter combination corresponds to a basic model and is also a simulation task.

[0055] Step 3: Build an LSTM neural network model and randomly initialize network parameters Among them, the network parameters The parameters may include the learning rate, batch size, number of hidden layers, number of neurons in each hidden layer, feature dimensions of input / output vectors, number of stacked layers of the hidden layer network, and weight coefficient matrix and bias value of the hidden layer. It mainly refers to the weight coefficient matrix and bias value of the neural network of the LSTM neural network model.

[0056] Step 4: Use the training data set constructed in step 2 to calculate the loss function of the basic model and use the gradient descent method to update the network parameters of the basic model, as shown in formula (1).

[0057] (1).

[0058] in, Represents the basic model After optimizing the network parameters, Represents the network parameters after random initialization, that is, the initial value of the network parameters, As a basic model The learning rate is usually , As a basic model The corresponding loss function is, The network parameters are LSTM neural network model, Represents the basic model The loss function Relative network parameters The gradient calculation result of .

[0059] This step corresponds to the inner loop optimization step of the MAML algorithm, which is used to optimize the performance of the model on the current task set and ensure the local adaptability of the model on the training data set.

[0060] This embodiment introduces the MAML algorithm, the main purpose of which is to be able to quickly adapt to the new model generated after modifying the parameters, and to quickly generate a proxy model that adapts to the new model after training with a small sample. The numerical calculation and solution process of the Modelica model is the main task of the Modelica compilation and solution engine, which mainly includes three steps: compilation, optimization and solution of the Modelica model. Among them, the compilation step is mainly responsible for completing the processing of checking, translating and flattening the Modelica syntax; the optimization step is mainly responsible for completing the processing of reducing the order, optimizing and sorting the Modelica equations; the solution step is mainly responsible for completing the numerical calculation.

[0061] Step 5: Sum the loss functions of all basic models and use the gradient descent method to update the meta-model network parameters, as shown in Equation (2).

[0062] (2).

[0063] in, represents the network parameters after meta-model optimization, represents the network parameters after random initialization, is the learning rate of the meta-model, usually , The network parameters are The meta-model of Represents the loss function of the meta-model relative to the network parameters When updating the meta-model network parameters, the loss calculation is performed by updating the meta-model loss function parameters, thereby updating the meta-model network parameters and finally outputting the optimized network parameters of the meta-model.

[0064] This step corresponds to the outer loop optimization step of the MAML algorithm. Its function is to optimize the performance of the meta-model in later tasks, that is, to ensure the global adaptability of the model within the range of parameter values ​​and improve the generalization ability of the model.

[0065] Step 6: Use the optimized network parameters to build the optimized meta-model At this point, the MAML algorithm is completed, the meta-learning training is completed, and the optimized meta-model is finally output Compared with the metamodel in the outer loop optimization phase of the MAML algorithm, the optimized metamodel is a new model with replaced parameters.

[0066] Step 7: Refer to the process of building a training data set in step 2, perform numerical calculations on the Modelica target model, collect a small amount of sample data, and build a small sample training data set.

[0067] Step 8: Use the small sample training dataset constructed in step 7 to optimize the meta-model Fine-tuning is to fine-tune the optimized meta-model. Perform small sample learning to obtain the Modelica agent model.

[0068] The purpose of this embodiment is to generate a corresponding Modelica proxy model for the Modelica target model. After the Modelica proxy model is constructed, its key features have been learned from the Modelica mechanism model. While reducing the computational complexity, it can also maintain a sufficient approximation to the target system. In the subsequent simulation calculation and model building process, it can be used as a substitute model of the original model for further model construction. Specifically, after obtaining the Modelica proxy model, the Modelica proxy model can be simulated as a separate model, or the Modelica proxy model can be used as a component to perform joint modeling with other models, so as to build a larger and more powerful model. For example, for a vehicle dynamics model, after the vehicle dynamics model is constructed, the vehicle dynamics model can be simulated alone, or it can be jointly modeled with a driver model, so as to construct a larger and more complete vehicle model, which can be used for vehicle dynamics testing. For another example, for a fuel cell model, after the fuel cell model is constructed, the fuel cell model can be simulated alone, or it can be jointly modeled with a vehicle transmission model of a vehicle transmission system or a vehicle air conditioning model of a vehicle air conditioning system to obtain a corresponding joint model for more complete thermal management or energy management simulation.

[0069] After the Modelica proxy model is generated in this embodiment, the Modelica proxy model can be used as a proxy model of the original model in the subsequent simulation calculation and model building process, and used in the further model building process. A corresponding proxy model can be generated for various models such as vehicle dynamics model or fuel cell model, and then jointly modeled with other models. Among them, the proxy model (Modelica proxy model) has the same interface as the original model (Modelica target model), and the input / output mapping relationship is basically the same. After replacing the original model, the proxy model can greatly improve the overall solution speed of the model and improve the efficiency of modeling and simulation.

[0070] In this embodiment, the key advantage of the Modelica proxy model is that it reduces a large amount of numerical solution and physical calculation workload by avoiding complex calculation tasks such as high-order differential terms, and simplifies the calculation process. By learning the input-output correspondence of the mechanism model, the prediction results are quickly mapped and generated. The advantages of processing nonlinear systems and complex systems are very obvious, and it is particularly suitable for tasks such as large-scale optimization, design exploration, and parameter sensitivity analysis. At the same time, through regular fine-tuning and verification, it can be ensured that its accuracy can meet actual needs.

[0071] This embodiment uses the MAML algorithm to learn the solution rules of different Modelica models, and then combines the LSTM network to build a Modelica proxy model for the Modelica mechanism model to quickly obtain the solution results. By introducing the MAML algorithm, the Modelica proxy model can be quickly adapted to new tasks and quickly solve the problem, which reduces the model's dependence on the amount of sample data and reduces the computational workload of the Modelica engine for batch simulation, thereby increasing the speed of solving the Modelica model and helping to improve the efficiency of modeling and simulation.

[0072] A Modelica proxy model generation method provided in this embodiment can directly convert a Modelica model into a Modelica proxy model in C language format, thereby saving the steps of compiling and optimizing the Modelica model by the compilation and solution engine, reducing the computing workload of the engine, speeding up the running speed of the model, and then improving the solution speed of the Modelica model. The Modelica proxy model does not change the interface specification of the original model by fitting the input and output data relationship of the Modelica model, can achieve seamless integration with the existing Modelica engine, and will not have any impact on the pre-processing and post-processing of the Modelica model, that is, the Modelica proxy model can simplify the Modelica mechanism model under the premise of keeping the interface and usage method unchanged. However, since complex calculations such as high-order differentials are no longer required when solving, the solution speed of the model can be greatly improved. For application scenarios where the mechanism is relatively fixed and some initialization parameters need to be changed frequently, such as commercial model libraries, the advantages of small sample learning and fast solution can be fully utilized.

[0073] In an exemplary embodiment, a computer device is provided, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any one of the above-described Modelica proxy model generation methods.

[0074] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the method for generating a Modelica proxy model is implemented.

[0075] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0076] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A Modelica proxy model generation method, characterized in that: The Modelica proxy model generation method comprises: Obtaining initialization parameters of a Modelica target model, and determining a value range and probability distribution of the initialization parameters; the Modelica target model is an original model for which a proxy model needs to be generated, and the initialization parameters include initial conditions, boundary conditions, physical property parameters, and penalty coefficients; Constructing a training data set according to the value range and probability distribution of the initialization parameters; Constructing an LSTM neural network model, and determining network parameters of the LSTM neural network model after random initialization; According to the training data set, the MAML algorithm is used to train and optimize the network parameters of the LSTM neural network model to obtain optimized network parameters; According to the optimized network parameters, construct an optimized meta-model; the optimized meta-model refers to a model obtained by replacing the randomly initialized network parameters with the optimized network parameters; The optimized meta-model is trained with a small sample to obtain a Modelica proxy model corresponding to the Modelica target model; the Modelica proxy model is used for computational solution of simulation tasks, and the solution speed of the Modelica proxy model is greater than the solution speed of the Modelica target model.

2. The Modelica proxy model generation method according to claim 1, characterized in that: According to the value range and probability distribution of the initialization parameters, a training data set is constructed, specifically including: Randomly generate a set of basic model parameters according to the value range and probability distribution of the initialization parameters; Constructing a plurality of simulation tasks according to the basic model parameters, and inputting the plurality of simulation tasks into the Modelica simulation software for batch simulation calculation to obtain solution results; Normalizing the solution result to obtain a normalized solution result; According to the solution results after the normalization process, a training data set is constructed.

3. The Modelica proxy model generation method according to claim 1, characterized in that: According to the training data set, the MAML algorithm is used to train and optimize the network parameters of the LSTM neural network model to obtain the optimized network parameters, which specifically include: According to the training data set, the loss function of the basic model is calculated, and the network parameters of the basic model are updated by using the gradient descent method; the basic model is the model corresponding to the LSTM neural network model in the inner loop optimization stage of the MAML algorithm; The loss functions of all the basic models are summed, and the network parameters of the meta-model are updated by the gradient descent method to obtain the optimized network parameters; the meta-model is the model corresponding to the LSTM neural network model in the outer loop optimization stage of the MAML algorithm.

4. The Modelica proxy model generation method according to claim 3, characterized in that: The following formula is used to calculate the loss function of the basic model based on the training data set, and the network parameters of the basic model are updated using the gradient descent method: ； in, represents the network parameters after optimization of the basic model, represents the network parameters after random initialization, is the learning rate of the base model, is the loss function corresponding to the basic model, The network parameters are LSTM neural network model, Represents the loss function of the basic model relative to the network parameters The gradient calculation result of .

5. The Modelica proxy model generation method according to claim 3, characterized in that: The following formula is used to represent the sum of the loss functions of all the basic models, and the network parameters of the meta-model are updated using the gradient descent method to obtain the optimized network parameters: ； in, represents the network parameters after meta-model optimization, represents the network parameters after random initialization, is the learning rate of the meta-model, The network parameters are The meta-model of Represents the loss function of the meta-model relative to the network parameters The gradient calculation result of .

6. The Modelica proxy model generation method according to claim 1, characterized in that: According to the optimized network parameters, an optimized meta-model is constructed, which specifically includes: The randomly initialized network parameters of the LSTM neural network model are replaced with the optimized network parameters to obtain the optimized meta-model.

7. The Modelica proxy model generation method according to claim 1, characterized in that: The optimized meta-model is trained with a small sample to obtain a Modelica proxy model corresponding to the Modelica target model, specifically including: Performing numerical calculations on the Modelica target model to construct a small sample training data set; The optimized meta-model is trained with a small sample using the small sample training data set to obtain the Modelica proxy model.

8. The Modelica proxy model generation method according to claim 1, characterized in that: The gating unit of the LSTM neural network model is the Sigmoid activation function, the state update and hidden state use the Tanh activation function, the optimizer is the Adam function, and the loss function is the mean square error function.

9. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the Modelica proxy model generation method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the Modelica proxy model generation method according to any one of claims 1 to 8 is implemented.

Images