Research On Time Series Hybrid Prediction Method Based On LSTM

Time series is a kind of data that contains abundant information.Time series prediction is an important data analysis method and has been receiving attention from scholars in various fields.In practical applications,simple and regular stationary sequences are rare,and it is very important to find a method that can effectively deal with various complex nonstationary time series and predict them accurately.This study combines the advantages of Long Short-Term Memory(LSTM)for nonlinear nonstationary time series with strong approximation ability and the characteristics of AutoRegressive Moving-Average(ARMA)for modeling and prediction of linear stationary time series,and proposes a new kind of time series hybrid prediction method based on deep learning.Compared with the existing methods,the difference of the method is:(1)Based on the theory of multiresolution analysis,the time series is decomposed by wavelet to obtain multiple sub-sequences and analyze the stationarity of each sub-sequence.In the field of time series prediction,wavelet analysis is used to denoise the sequence in order to improve the prediction accuracy.In this paper,wavelet decomposition is used to get the stationary sub sequences representing the details and the non-stationary sequences containing periodicity,seasonality and trend information.(2)The ARMA with lower computational complexity is used to model and predict the smooth subsequences representing the fluctuations of the detail.This step improves the fitting and prediction accuracy of the details of the time series.The predecessors method mostly removes the influence of periodicity,seasonality,or trend by means of differentials,etc.,to obtain a sequence that satisfies the prerequisites of the time series prediction method based on mathematical statistics.In this paper,the single-branch reconstruction of sub-sequences after wavelet decomposition is performed,and then ARMA is used for modeling and prediction,which preserves the complete details of the original time series.(3)Unlike the artificial neural network(artificial neural network)for fitting and predicting the whole time series,this paper only uses LSTM to model and predict the non stationary subsequences containing the periodic,seasonal and trend information in the original time series.By this step,the influence of the slowly changing factors in the time series on the time series is retained,the fitting and prediction accuracy of the overall trend of the time series are improved,and the complexity of the network is also reduced.(4)In this paper,we use wavelet transform to reconstruct the original signal perfectly and reconstruct the prediction subsequence of each model to get the final prediction result.Compared with previous methods,the predicted results have neither loss of detail information nor loss of overall trend information.Compared with the Autoregressive Intergrated Moving Average(ARIMA)method,BP neural network and direct LSTM method,this method has obvious advantages in fitting accuracy and prediction accuracy on different types of data sets.Both the forecast of the cycle and the overall trend,as well as the capture of the details of fluctuations,can be balanced.At the same time,the network structure of the deep learning method LSTM has been streamlined,and problems such as difficulty in training the LSTM network have been alleviated.