A Deep Learning Framework for Univariate Time Series Prediction Using Convolutional LSTM Stacked Autoencoders
Aniekan Emmanuel Essien, Cinzia Giannetti · 2019
This paper proposes a deep learning framework where wavelet transforms (WT), 2-dimensional Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) stacked autoencoders (SAE) are combined towards single-step time series prediction. Within the framework, the input dataset is denoised using wavelet decomposition, before learning in an unsupervised manner using SAEs comprising bidirectional Convolutional LSTM (ConvLSTM) layers to predict a single-step ahead value. To evaluate our proposed framework, we compared its performance to two (2) state-of-the-art deep learning predictive models using three open-source univariate time series datasets. The experimental results support the value of the approach when applied to univariate time series prediction.