A Deep Learning Framework for High-Dimensional Time Series Forecasting

Abstract

Accurate long-term time series forecasting remains a challenging task in various real-world applications, requiring the consideration of high-dimensional features. Traditional forecasting methods often rely on manual feature selection, which can be time-consuming and subjective and may not capture the intricate relationships among variables. While advanced deep learning methods have significantly enhanced our understanding of complex data dynamics, it is important to note that time series possess distinct properties such as trend, seasonality, and autocorrelation, which differentiate them from other data sequences. These specific characteristics should be carefully taken into account in the modeling process. To address these challenges, we propose a novel architecture that combines a graph neural network block with a Long Short-Term Memory (LSTM) encoder-decoder block for end-to-end feature learning and forecasting in multivariate time series forecasting. This architecture learns complex data relationships while considering time series properties, resulting in accurate and reliable predictions. Our evaluation of this framework on ETDataset benchmark datasets for sequence forecasting tasks shows forecasting accuracy and reliability comparable to prior state-of-the-art models. This framework helps decision-makers make informed decisions with deep learning models based on the best practices in time series analysis.