Long Sequence Time-Series Forecasting via Gated Convolution and Temporal Attention Mechanism

Abstract

In many defacto application, we seek to forecast long time series, such as electricity usage planning. Long sequence time-series forecasting (LSTF) requires a model with high predictive ability, that is, it can effectively capture a more accurate long-range correlation coupling between output and input. Recently, Informer achieved satisfactory performance. However, the complexity and uncertainty of a large amount of time series information limit the accuracy of long-series time series forecasting, and many practical problems put forward new requirements for noise filtering of time series forecasting. To ameliorate these issues, we improve Informer based on gated convolution and temporal attention mechanism, called GCTAM: (i) The proposed temporal gated convolution explicitly leverages temporal information to automatically route the results according to the temporal information. (ii) The proposed temporal attention mechanism filters out low-frequency noise well. Experiments on multiple real datasets empirically demonstrate that our method outperforms Informer.