MCNet: Multivariate long-term time series forecasting with local and global context modeling

Abstract

Time series data typically exhibit various intra-sequence and inter-sequence correlations, resulting in intricate, intertwined dependencies, which pose challenges for accurately predicting future long-term trends. Previous studies have not fully considered the two correlations, and they also still face challenges of excessive time and memory complexity when dealing with long-term predictions. To address these challenges and establish high-precision prediction models, we propose MCNet that consists of a local branch and a global branch. The local branch aims at capturing short-term variations of intra-sequences, as well as capturing inter-sequence correlations. The global branch models long-term dependencies within sequences. Specifically, the local branch consists only of MLP module, which effectively captures short-term variations of intra-sequences by independently modeling the temporal information within and between patches of the most recent time series. Subsequently, inter-sequence dependencies are captured through the channel interaction module, which further explores more key information to improve the performance of MCNet. Meanwhile, global branch models long-term dependencies within the time series through structured global convolution. Experimental results on multiple popular long-term time series forecasting benchmarks demonstrate that MCNet outperforms state-of-the-art methods, yielding a relative improvement of 12% for multivariate time series while also being more efficient.