Abstract
Efficient and effective multivariate time series (MTS) forecasting is critical for real-world applications, such as traffic management and energy dispatching. Most of the current deep learning studies (e.g., Spatio-Temporal Graph Neural Networks and Transformers) fall short in a trade-off between performance and efficiency. Existing MTS forecasting studies have yet to fully and simultaneously address issues such as modelling both temporal and variate dependencies, as well as the temporal locality, hindering broader applications. In this paper, we propose a lightweight model, i.e., Time Series MLP (TSP). TSP is built upon MLP and relies on the PrecMLP with the proposed computationally free Precurrent mechanism to model both the variate dependency and temporal locality, thus being simple, effective and versatile. Extensive experiments show that TSP outperforms state-of-the-art methods on 16 datasets for both Long-term Time-series Forecasting and Traffic Forecasting tasks. Furthermore, it attains a significant reduction of at least 95.97% in practical training speed on the CPU.
