IMPLICIT STACKED AUTOREGRESSIVE MODEL FOR WEATHER FORECASTING
Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)
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: Weather Forecasting, Climate Change
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: We propose a stacked auto-regressive method that combines the auto-regressive method and the non-autoregressive method.
Abstract: As global climate change intensifies, the accuracy and reliability of weather forecasting have become increasingly crucial. Accurate predictions are vital not only for preparing for extreme weather events, but also for understanding the long-term implications of changing climate patterns. To address these issues, data-driven methods have begun to be applied. Three primary methods have been proposed: the autoregressive method, lead time embedding, and the non-autoregressive method. However, the autoregressive method has shown a significant decline in performance as the lead time increases due to the accumulation of errors. While the non-autoregressive method offers high performance, it can only predict at fixed lead times and intervals. Lastly, the lead time embedding method, which does not perform temporal modeling, failed to predict complex patterns. In this paper, we introduce the Implicit Stacked Autoregressive Model for Weather Forecasting (IAM4WF), an implicit video prediction model that employs a stacked autoregressive approach. Similar to non-autoregressive methods, stacked autoregressive methods utilize the same observed frame to forecast all subsequent frames. Yet, they incorporate their predictions as input, much like autoregressive methods. As predictions span over an increasing number of time steps, they are systematically queued in sequence. To validate IAM4WF's efficacy, we conducted tests on three prevalent future frame prediction benchmark datasets and weather and climate prediction datasets. Experimental results show that IAM4WF significantly improves performance on all datasets we evaluated.
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: 5703
Loading