A hybrid deep learning network for non-linear time series prediction

Abstract

Non-linear time series prediction is highly significant because most of the practical situations deals with time series which are non-linear in nature. This study suggests a new time series prediction CEEMDAN-SVDLSTM model amalgamating Complete Ensemble EMD with Adaptive Noise, Singular Value Decomposition and Long Short Term Memory network. Non-linear and non-stationary data can be analysed by deploying the above model. CEEMDAN stage, SVD stage and LSTM stage are the main parts of the model. The break down of the data into some IMF components plus a residue is carried out by CEEMDAN in the first stage. Each IMF component and residue so obtained is de-noised by SVD in the second stage. Third stage deployed LSTM to predict all the de-noised IMF components. The foretold values of the actual data is then obtained by adding all the predicted IMF components and residue. We compared the model with other models such as LSTM model, EMD-LSTM model, EEMD-LSTM model, CEEMDAN-LSTM model and EEMD-SVD-LSTM model. The results show that the suggested CEEMDAN-SVD-LSTM model works better than other models in terms of efficiency in predicting future values.