Signal Decomposition and Relational Inference For Time Series Forecasting

Abstract

Time series forecasting analyzes time series data and predicts the future based on past time series data. In multivariate time series, several time series signals interact with each other in a complex system (more than one- time series to be considered). But until now, multivariate time series forecasting studies have used only their own time series signal without considering the interaction of others. In this paper, we propose a deep learning model called IC-PN-BEATS. This model can be divided into two parts. The first part is the relational inference part that learns the relational information between various time series signals in multivariate time series settings, and the second part is the forecasting part that predicts the future using the learned relational information. The relation inference part allows us to use it on most real-world datasets with no structural, relational information. In addition, in the forecasting part, time series signals were decomposed step by step by applying the doubly residual stack architecture and additionally decomposed to trends and seasonality by a pooling layer. With this architecture, we can decompose complex time series signals into simple trends and seasonality signals which makes forecasting easy. We had an experiment on the multivariate time series benchmark dataset and compared to other transformer-base models, our model performed 18% higher.