Nonlinear System Identification Based On LSTM Neural Networ

Nonlinear system identification has become one of the fundamental issues in control science and engineering applications.Its applications involve the fields of machinery,biology,chemistry,and economics.In order to analyze the operating state of a system,it is usually necessary to accurately obtain the mathematical model of the system.However,in the absence of prior knowledge,it is often difficult to obtain a mathematical model of the system.Therefore,the identification of nonlinear systems is a challenging task.The system identification method proposed in this paper is based on neural networks and approximates the mapping relationship of nonlinear systems through learning,thereby achieving the identification of nonlinear systems.The main work of this article is as follows:(1)Aiming at the problems of gradient loss and slow convergence in the time based back propagation training algorithm of recurrent neural networks,an improved method for nonlinear system identification based on a one-way long-short term memory network model is proposed.By separating the long-short term memory network elements into forward and recursive models,a training design scheme with faster backpropagation over time is proposed.Deep long-term and long-term memory networks are implemented by combining deep cyclic neural networks with multi-layer perceptron.For deep cyclic neural networks and training,a training method similar to back propagation is proposed to prove the stability of the algorithm.Experiments were conducted on Wiener Hammerstein systems,general nonlinear systems,and aerodynamic models.The effects of factors such as the input dimension of the system,noise,the number of hidden layers and nodes in the network,training time,and other factors on the experiment were considered.The experimental results showed that the fast training method for long-short term memory networks was superior to other methods.(2)Aiming at the problem that information in traditional one-way long-term and short-term memory networks can only flow in one direction,and the errors in the model will accumulate over time,a nonlinear system identification method based on multi-layer superimposed twoway long-term and long-term memory networks is proposed,which can simultaneously memorize past and future information,reduce error accumulation,and more accurately identify the system.It is extended to a bidirectional long-term and short-term memory network,which trains input data through both forward and reverse training,thereby mining the correlation between data.In experiments on power load distribution systems and highway flow and speed systems,considering factors such as robustness and generalization of different training data samples to the model,the results show that the bidirectional long-short term memory network is superior to traditional methods.Due to the constraints of encoder and decoder architectures on long-short term memory networks,all inputs are forced to be encoded as fixed length vectors for representation,limiting the performance of long-short term memory networks.By adding attention mechanisms,the encoder and decoder architectures are freed from fixed length vectors,thereby improving the performance of the model.By comparing the results of the proposed method with those of previous methods in a GPS trajectory system,the experimental results show that the method can extract dynamic characteristics from data,and has strong memory functions,improve the performance of the model,and increase the accuracy of identification and prediction.