From global to local: A lightweight CNN approach for long-term time series forecasting

Abstract

In the context of the artificial intelligence revolution, the demand for long-term time series forecasting (LTSF) across various applications continues to rise. Contemporary deep learning models such as Transformer-based and MLP-based models have shown promise. However, these state-of-the-art (SOTA) approaches encounter notable limitations: Transformer-based models suffer from low computational efficiency and the inherent restrictions of point-wise attention mechanisms, while MLP-based models struggle to effectively capture local temporal dependencies. To overcome these challenges, this paper introduces a novel lightweight architecture centered around CNN-based models with an inherent receptive field, GLCN, explicitly designed to capture and discern intricate relationships in time series. The architecture features a key component, the global–local block, which initially segments the time series into subseries levels to preserve the underlying semantic information of temporal variations and subsequently captures both inter- and intra-patch inherent global and local temporal dynamics. In particular, GLCN utilizes a lightweight CNN-based architecture for prediction to significantly enhance training speed by 65.1% and 86.0% on the Weather and ETTh1 datasets, respectively, while reducing parameters by 94.8% and 94.4%. Comprehensive experiments on seven real-world datasets demonstrate that GLCN reduces contemporary SOTA approaches by 1.6% and 1.8% in Mean Squared Error and Mean Absolute Error.