Research On Parameter Identification Technology Of Modelica Model Based On Reinforcement Learning

Modelica,as a multi-domain unified modeling language,has become the de facto standard in multi-domain modeling.For a Modelica model,the accuracy of the parameters directly determines the effectiveness of the model,so calibration parameters are a key part of the process of creating a model.The existing calibration methods are mainly manual trial and error based on manuals or empirical knowledge.The calibration process is cumbersome and cannot guarantee the accuracy of the parameters;the identification of model parameters using least squares estimation,particle swarm optimization and other algorithms has improved to a certain extent The speed and accuracy of the parameter calibration have been solved,but there are still problems such as easy to fall into local optimum and complicated mathematical form.Therefore,it is necessary to study the parameter identification of the Modelica model.This paper has carried out research on the parameter identification algorithm of the model and the overall identification scheme.The paper improves based on the existing identification algorithm CARLA and proposes the ECARLA algorithm.There are three main improvements.One is to use a normal random number generator instead of an integration operation to increase the convergence speed;the other is to introduce randomness and try to avoid localization.The best;the third is to change the public reward factor to a private reward factor to eliminate the learning bias caused by the difference in sensitivity.The paper expands on the framework of traditional reinforcement learning algorithm QLearning and proposes the EQL algorithm.It designs the reward and punishment function,action selection strategy and Q value table update strategy for the characteristics of the parameter identification problem,and proposes "training cycle-optimization Round-the search cycle " triple loop serial learning process,to avoid meaningless action trial and error,improve the convergence speed of the algorithm.The thesis puts forward the parameter identification scheme of "three-phase identification".The first stage is "sensitivity analysis",using Sobol method and Monte Carlo sampling to analyze the sensitivity of the parameters to be identified;the second stage is "parameter coarse adjustment",using ECARLA or EQL algorithm to carry out preliminary parameter identification of the model;the third stage For "parameter fine-tuning",particle swarm optimization is used to identify the final parameters of the model.On this basis,two sets of model parameter identification schemes of "sensitivity analysis + EQL + PSO" and "ECARLA+ PSO" were formed.Finally,the thesis designs and implements a modular parameter identification system,and the validity and application scenarios of identification algorithms and schemes are verified through case analysis.