A Generalized Additive Decomposition Based Deep Neural Network Architecture for Time Series Forecasting Problems

Multivariate time series forecasting is a significant area of research across many domains such as finance, sales, crime prediction, electricity consumption, road occupancy prediction etc. The availability of large multivariate data along with the existence of spatial and temporal short-term and long-term patterns in them make this task a difficult one. In this study, we propose a new idea of a general framework having a multi modular architecture which is built only using deep learning algorithms to perform multi-step forecasting of multivariate time series data. We preprocessed time series data by decomposing it into seasonal, trend and residual components using additive decomposition technique. Then we separately fed each component into independent modules that utilize recurrent deep learning methods such as Stacked GRU, Self attention based LSTM and LSTM encoder decoder architecture. The final result is the summation of the forecast results of all three components. The framework is tested on three benchmark dataset on which it performed encouragingly well.