Temporal Attention Gate Network With Temporal Decomposition for Improved Prediction Accuracy of Univariate Time-Series Data

Abstract

Time-series forecasting has widely been addressed in data science and various domains, but many limitations persist in terms of prediction accuracy. We propose a network architecture called temporal attention gate network (TAGNet) to improve the prediction accuracy of time-series prediction. TAGNet integrates new concepts of temporal filter and temporal attention gate. First, the temporal filter learns information embedded in time-series data by decomposing the input data through variational mode decomposition. Second, the temporal attention gate learns the relationship between the decomposed time-series signals and hidden states to learn their relationships. To verify the performance of the proposed TAGNet, a comparative experiment was conducted on three univariate time-series datasets. The results show that the prediction performance improves by 15% on average for short-, medium-, and long-term predictions compared with various deep learning methods.