PowerGPT: Foundation Model for Power Systems

Shihao Tu, Yupeng Zhang, Yan Huajiang, Fangbin Ye, An Wen, Zhendong Fu, Yang Yang

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: power systems, foundation model, self-supervised learning, time series

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: We propose a foundation model, namely PowerGPT, to model electricity time series (ETS) data, which learns generic representations of load and electricity consumption data by pre-training, providing a large-scale, off-the-shelf model for power systems. PowerGPT is the largest model in the field of power systems and is pre-trained on a large-scale ETS data including load and electricity consumption data. The design of PowerGPT is to capture long-term temporal dependency and hierarchical correlation from massive ETS data, providing information that spans from the fine-grained to coarse-grained scales. As a foundation model, PowerGPT achieves SOTA performance on various downstream tasks in power systems (i.e. forecasting, missing value imputation, and anomaly detection), showing the generalization ability to a wide range of tasks. The low-resource label analysis further illustrates the effectiveness of our pre-training strategy. In addition, we explore the effect of model size to show that a larger-scale model with a higher capacity can lead to performance improvements.

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: 6875