GAF-CNN-LSTM for Multivariate Time- Series Images Forecasting

Abstract

Forecasting multivariate time series is challenging for a whole host of reasons not limited to problem features such as having multiple input variables, time series preparation, and the need to perform the same type of prediction for multiple physical sites. Although the literature on time series forecasting is focused on 1D signals. We use the Gramian Angular Fields (GAFs) to encode time series into 2D texture images, later take advantage of the deep CNN-LSTM architecture where LSTM uses a CNN as front end. Thus, we propose a novel unified framework for forecasting multivariate time series using a way to encode time series as images. Preliminary experimental results on the UEA multivariate time series forecasting archive, demonstrate competitive forecast accuracy (RMSE and MAPE) of the proposed approach, compared to the existing deep approaches as LSTM, CRNN, 1D-MTCNN.