Research On Nonlinear Time Series Prediction Methods Based On Deep Learning

Nonlinearities exits everywhere,such as network traffic sequences under local area networks,missile trajectories in the sky,meteorological data,molecular operation laws and solar activity are all data generated by nonlinear systems,and the predictive value of the nonlinear data can be used to make reasonable planning measures in the future,which can generate immeasurable value.In conventional nonlinear prediction,the hyperparameters of nonlinear prediction networks are generally determined by grid optimization or multiple trial-and-error methods,which have great uncertainties and high trial-and-error costs.In addition,the phase space reconstruction of chaotic time series data in nonlinear data is too chaotic,and its phase space structure lacks adaptability for time-series prediction networks.In addition,the data prediction are trained on the data using the time window method,and the length of the time window size is also experimented out several times.These problems bring a lot of uncertainty troubles to nonlinear data prediction.Based on this,the following improvement works are carried out in this paper for the hyperparametric optimization of network parameters and the reconstruction of the training data set of chaotic time series.Firstly,to address the problem that hyperparameters in nonlinear time-series prediction networks are difficult to determine,this paper proposes an improved differential evolution-based optimization algorithm based on the use of differential evolution algorithm for parameter finding,which can find parameters with greater global diversity,so that the final found parameters can achieve better accuracy for target neural network prediction,and the algorithm outperforms the general genetic algorithm for prediction networks by verifying one-dimensional nonlinear data and three-dimensional nonlinear data.Secondly,to optimize the dimensional validity and training time window size of the training data for chaotic time series prediction,a phase space parameter estimation algorithm based on the incremental attention mechanism is proposed,and the dimensional weight criterion corresponding to the attention weights from low to high dimensions proposed in this paper is used to traverse and verify the phase space dimensions generated by different parameters,and the parameter that best meets the dimensional weight criterion is selected as the parameter for phase space reconstruction,and the experiments show that the phase space reconstruction method proposed in this paper has higher adaptability to the traditional phase space reconstruction method compared with the traditional phase space reconstruction method,which makes the prediction more accurate.Finally,by embedding the algorithm of reconstructing the phase space by incremental attention mechanism into the input layer of the long short-term-convolutional neural network,a network model more suitable for chaotic time series prediction is proposed.By comparing with the four networks,this network reflects higher accuracy of nonlinear chaotic prediction,while calculating the training window size of traditional nonlinear data by using incremental attention mechanism,it is found that it also improves the prediction accuracy.