DDNet: Exploring Dual Dependencies for Long-Term Time Series Forecasting

Abstract

Recent Transformer-based methods have advanced multivariate time series forecasting by focusing primarily on temporal dependencies (cross-time dependencies). However, these methods often overlook crucial multivariate correlations (cross-channel dependencies), leading to suboptimal performance. In this paper, we propose a novel Dual Dependencies modeling Network (DDNet) to model both cross-time and cross-channel dependencies effectively. DDNet employs aggregation tokens in the Aggregation Stage to capture cross-channel relationships and utilizes these tokens in the Broadcast Stage to incorporate cross-time dependencies. Additionally, we propose a fine-grained Patch-Wise Normalization technique to address the overfitting challenges when modeling cross-channel dependencies. Extensive experiments on benchmark datasets demonstrate that DDNet consistently surpasses state-of-the-art methods while maintaining low computational complexity.