Sequence Attention for Multivariate Time Series Forecasting

Many real-world applications require the prediction of multivariate time series, including traffic route planning and electricity consumption organizing. Temporal data that arise in these real-world applications often involves mixed periodic temporal patterns and the dependencies of different series, which complicate this task. The typical transformer reviews the information at each time step and selects relevant information to help generate the outputs. However, it fails to capture complex dynamic interdependencies between different series. In this paper, we propose a novel attention mechanism to select relevant time series for multivariate time series forecasting. Then we propose a flexible multi-output strategy to avoid error accumulation in the inference phase. Finally, we apply the proposed model to several real-world tasks and achieve state-of-the-art performance in almost all cases.