Advancing Time Series Forecasting: LSTM Networks with Multiple Attention Mechanisms

Recently recurrent neural networks due to their ability to capture time-dependent features have been applied to time series forecasting showing important improvements with respect to previous methods. Simple RNN architectures though suffer from vanishing/exploding gradient problems and cannot discriminate exogenous series in case these are given as input. We examine the effectiveness of a solution that uses LSTM networks and multiple attention mechanisms by comparing it to simpler models, including an encoder architecture. Our analysis aims to showcase the outstanding performance of the proposed approach as its primary objective.