Research On Prediction Method Of Combined Model Based On Time Series Decomposition

Data prediction is considered to be one of the most challenging problems in the field of data mining.As an important part of data mining,time series data has been widely concerned by scholars in various fields.With the advent of the Internet era,the availability of time series data prediction technology is increasing day by day.Time series widely exists in finance,transportation,medical treatment,energy and other fields.Accurate prediction of time series data will bring great help to all walks of life.It is very important to explore the method of modeling and forecasting complex nonlinear and non-stationary time series.This paper discusses the research status of time series analysis methods at home and abroad,and compares the characteristics and applicability of various methods.In view of a specific problem in the field of macroeconomics,this paper discusses how to establish the corresponding model based on the analysis of time series and the advantages of the evaluation model in the prediction process.The specific research contents are as follows:This paper first tests the distribution of time series data samples,improves the homogeneity and normal distribution of variance of time series samples by box Cox transformation,reduces the correlation between the prediction variables and the unobservable errors,and is different from the method of eliminating the influence of non-stationary information in time series by means of difference to meet the preconditions of statistical prediction methods.This paper focuses on the non-stationary information The stable time series is decomposed and analyzed from three aspects: trend,season and random term.The trend information is modeled by exponential smoothing method,and the weight is determined according to the influence of historical data to increase the correction of trend information,which can accurately model the long-term trend factors;the long-term and short-term memory network is used to model the random items containing non trend and non season information in the time series to fit the non-linear information in the time series,and accurately to the short-term fluctuations Finally,the combination model is used to reconstruct the prediction results of each model,and the final prediction results of the original time series are obtained by synthesizing the characteristics of different models.The experimental results show that: the algorithm retains the overall trend information,seasonal change information and random fluctuation information in the time series,and has better results in the fitting accuracy and prediction accuracy in the training process;compared with the deep multi neuron complex neural network structure modeling,this paper uses a simpler long-term and short-term memory network structure,which is slow to a certain extent The difficulty of training optimization and over fitting of complex neural network model are solved.