Research On Dynamic System Modeling Based On Recurrent Neural Network

With the rapid development of artificial intelligence and machine learning,the research of dynamic system modeling based on deep networks is currently a research hotspot in the AI field.Due to the inherent strong nonlinearity,time-varying,and randomness of the dynamic system,how to build the model of the dynamic system has been the focus of people’s research.The recurrent neural network is often used in the modeling and analysis of this type of object due to its basic structural characteristics and time relevance,which has achieved certain results.The research of dynamic systems begins with the study of time series forecasting.Traditional time series forecasting methods have made some progress in the study of dynamic system modeling,but there are still some deficiencies,such as the weak correlation of system states and model training and tuning difficulties.With the rapid development of deep learning technology,dynamic neural networks such as RNN,LSTM,and GRU have been profound research and some applications.Compared with traditional dynamic models,dynamic neural networks have greatly improved in terms of dynamic performance,predictive ability,memory ability,and timely relevance.However,dynamic neural networks still have drawbacks,such as network structure,long-term dependence,computational complexity,and its related research and modeling still need deep analysis and research.This article focuses on the analysis and research on the structure of LSTM,proposes a model optimization algorithm(WAE),and designs a time-skip(TS)recurrent neural network.The experiments show that it has great performance.The main work of this paper:(1)Given the complex structure and parameter redundancy problems of LSTM and other recurrent neural networks,related research,and analysis on the structure of recurrent neural networks have been done.A weight activity evaluation algorithm(WAE)is proposed,which is based on the basics of the network.The unit carries out weight activity evaluation and structure screening to improve the structural rationality of the recurrent neural network and reduce the calculation amount of network parameters.Experimental results show that the algorithm can better optimize the network structure,reduce the redundancy of network parameters,and increase the calculation time.(2)Aiming at the problems of insufficient memory ability and difficulty in backpropagation of gradients in recurrent neural networks,a new recurrent neural network structure is proposed,namely,the time-skip(TS)recurrent neural network model,which strengthens the memory capacity,helps the gradient to propagate backward more easily,and the data transmission in the time direction is more stable and effective.Taking epilepsy data and arrhythmia data as the research objects,the dynamic performance of the model under different parameters is analyzed.The experimental results show that the model can strengthen the memory ability of the recurrent neural network.The problem of long-term dependence have been alleviated.