PANDA: Patch-Aware Graph Network with Dual Alignment for Time Series Forecasting

Abstract

Multivariate time series (MTS) forecasting aims to predict future patterns by extracting features from multivariate history. Predominant methods face challenges in learning spatial dependencies while capturing long-term trends and local details, leading to suboptimal performance in MTS forecasting. To address this problem, we propose a Patch-Aware graph Network with Dual Alignment (PANDA) of features in both time and frequency domain. Specifically, we utilize a patch-aware graph neural network to extract spatio-temporal dependencies. Additionally, dual alignment is employed for multi-perspective optimization. By performing multi-view joint optimization, our model reduces feature redundancy and improves the extraction of structured spatio-temporal patterns. Extensive experiments demonstrate that PANDA achieves superior forecasting accuracy in both long- and short-time series forecasting. Code is available at this repository: https://github.com/lichen0620/PANDA.