Contextual ES-adRNN with Attention Mechanisms for Forecasting

Abstract

In this study, we propose a hybrid contextual forecasting model with attention mechanisms for generating context information. The model combines exponential smoothing and recurrent neural network to extract and synthesize information at both the individual series and collective dataset levels. The model is composed of two simultaneously trained tracks: context track and main track. The main track generates forecasts and predictive intervals, while the context track generates additional inputs for the main track based on representative time series. Attention mechanisms are integrated into the model in six different variations to adjust the context information to the forecasted series and so increase the predictive power of the model.