ML-FORMER: Forecasting by Neighborhood and Long-Range Dependencies

As sensors are deployed widely, collected data present features of large quantity and high dimensionality, which pose enormous challenges to multivariate long sequence time-series forecasting (MLTF). Existing methods for MLTF tasks can not efficiently capture neighborhood and long-range dependencies, resulting in low prediction accuracy. In this paper, we propose a novel multivariate long sequence time-series method, called ML-Former, that captures both neighborhood and longrange dependencies to enhance the prediction capacity. Specifically, MLFormer first conducts a time-series embedding that integrates neighborhood dependencies, positions, and timestamps. Then, it captures neighborhood and long-range dependencies by using a time-series encoderdecoder. Furthermore, an innovative loss function is designed to improve the convergence of ML-Former. Experimental results on three real-world datasets show that ML-Former reduces forecasting error by up to 35.4% compared with benchmarking methods.