Forecasting Multivariate Time-Series Data Using LSTM and Mini-Batches

Abstract

Multivariate time-series data forecasting is a challenging task due to nonlinear interdependencies in complex industrial systems. It is crucial to model these dependencies automatically using the ability of neural networks to learn features by extraction of spatial relationships. In this paper, we converted non-spatial multivariate time-series data into a time-space format and used Recurrent Neural Networks (RNNs) which are building blocks of Long Short-Term Memory (LSTM) networks for sequential analysis of multi-attribute industrial data for future predictions. We compared the effect of mini-batch length and attribute numbers on prediction accuracy and found the importance of spatio-temporal locality for detecting patterns using LSTM.