Vision-Enhanced Time Series Forecasting by Decomposed Feature Extraction and Composed Reconstruction

Mingyang Yu, Peng Chen, Xiahui Guo, Zhenkai Li, Yang Shu

Keywords: Time Series Forecasting

Abstract: Time series forecasting plays a crucial role in various domains, such as power and weather forecasting. In recent years, different types of models have achieved promising results in long-term time series forecasting. However, these models often produce predictions that lack consistency with the style of the input, resulting in reduced reliability and trust in the forecasts. To address this issue, we propose the Vision-Enhanced Time Series Forecasting by Decomposed Feature Extraction and Composed Reconstruction (VisiTER), which leverages the rich semantic information provided by the image modality to enhance the realism of the predictions. It consists of two main components: the Decomposed Time Series to Image Generation and the Composed Image to Time Series Generation. In the first component, the Decomposed Time Series Feature Extraction Model extracts periodic and trend information, which is then transformed into images using our proposed time series to vision transformation architecture. After converting the input time series into images, the resulting images are used as style features and concatenated with the previously extracted features. In the second component, we use our proposed TimeIR along with the previously obtained feature set to perform image reconstruction for the prediction part. Due to the rich information provided, the reconstructed images exhibit better consistency with the input images, which are then transformed back into time series. Extensive experiments on seven real-world datasets demonstrate that VisiTER achieves state-of-the-art prediction performance on both traditional metrics and new metrics.

Primary Area: learning on time series and dynamical systems

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: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 3717