Abstract: Long-term time series forecasting (LTSF) is still very challenging in many real-world applications. A fundamental difficulty is in efficiently modeling both the short-term temporal patterns and long-term dependencies. in this paper, we introduce a novel two-stage attention-based LTSF model called Memory Attention for Time-Series forecasting (MATS). In stage I, short-term temporal patterns are extracted to a memory bank such that the input time series is represented by a much shorter sequence of memory attentions. In stage II, a sequence-to-sequence predictor is trained to discover long-term dependencies in the memory attention sequence, and forecast memory attentions corresponding to the time series in the future. The use of attention allows a flexible representation, and its shorter sequence length enables the model to more easily learn long-term dependencies. Extensive experiments on a number of multivariate and univariate benchmark datasets demonstrate that MATS outperforms SOTA LTSF methods almost all the time.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics
Submission Guidelines: Yes
Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)