PeriodNet:Lightweight And Efficient Time Series Prediction Model Based On Periodic Characteristics
Primary Area: representation learning for computer vision, audio, language, and other modalities
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.
Keywords: Time Series Analysis, Multivariate Timeseries Forecasting, Local And Global Context
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: A novel periodic feature extraction method and a concise local feature and global feature fusion module are proposed. Compared with the state-of-the-art results, significant improvements have been achieved.
Abstract: The task of multivariate time series prediction has always been a challenging task. In this field, various related methods emerge in endlessly, whether based on fully connected, convolutional neural networks or attention-based models, all have achieved remarkable results. However, current long-term prediction tasks mainly rely on complex attention mechanisms or causal convolutions, which result in huge computational costs and are not suitable for edge devices or scenarios with limited computing resources. Therefore, our research focuses on lightweight time series prediction model exploration. Our main work focuses on the analysis of time series data, focusing on the importance of periodic features and the fusion of local features and global features. Based on the mathematical idea of Fourier series, we designed a simple and lightweight module for extracting periodic features; and designed a lightweight module that can effectively fuse local information and global information, thereby enhancing Feature representation and prediction performance. By comparing with the current state-of-the-art results, we verified the effectiveness of the module we designed. On 7 benchmark data sets including etth1, etth2 and ili etc., our model achieved significant performance improvements compared to the state-of-the-art results. The specific code of our research results can be found at https://github.com/sep21Be/periodNet.
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.
Submission Number: 5686
Loading