Research Of Prediction Technique For Time Series

In almost every area of life, the prediction of data is ubiquitous. Therefore, the researchof the prediction technique for time series has important practical significance. In this article,the prediction issue of simple linear, nonlinear and complex nonlinear time series isrespectively studied.The existing time series forecasting model and related technologies is analysed.Firstly,the characteristics of time series data is tested, including test of linear, nonlinear andmemory.On the basis of these, the prediction model and methods of different characteristics isstudied, including the model of short memory and long memory and the method of complexnonlinear time series forecasting.On the basis of rolling time series forecasting, the method of the optimal estimationoptimize rolling time series forecasting is studied. Particle filter (PF) optimizes rolling timeseries forecasting is proposed.First, data is rolling extracted. Characteristics of data isanalysed.Secondly, model is adaptable constructed.Particle set is geted by the parameterdisturbances.According to models, prediction is put in practice.Based the deviation ofobservation and verity value,the weight of particle is worked out.Thus, the precious ofsingle-step and multi-step predictive is improved.Futher, Kernel particle filter (KPF)optimizes rolling time series forecasting is proposed.Compered with rolling time seriesforecasting,experimental shows that the method of the optimal estimation optimize rollingtime series forecasting make predict more accurate.On the basis of time series forecast based on neural network, the method of complexnonlinear time series'forecast based on KII neural networks is proposed. Network structure istraversal searched by using test value method and the structure is determined according toOccam's razor principle. After that, parameters of network are trained by using gradientdescent. With classical nonlinear data of sunspot number, effectiveness of the algorithm isverified in the experimental.In this article, the research has certain theoretical meaning and practical application ofvalue for increasing time series forecasting precision.