SONNET: Solar-disaggregation-based Day-ahead Probabilistic Net Load Forecasting with Transformers
Keywords: Net load forecasting, Probabilistic Modeling, Transformers, Solar Disaggregation, Data Augmentation
TL;DR: We introduce a SOTA probabilistic method for real-world day-ahead net load forecasting challenges
Abstract: The global transition towards sustainable energy sources has positioned solar power as a cornerstone of modern electricity systems, underscoring the critical need for advanced forecasting techniques in grid management. Accurate net load forecasting is crucial for efficient and reliable power grid operations, especially with the rapid deployment of behind-the-meter (BTM) renewable energy sources such as rooftop solar. Notably, BTM solar generation is neither controlled nor monitored by utilities and hence only net load data are observed. Different from load forecasting, net load forecasting faces new challenges because BTM solar, a major component of net load, behaves very differently from and is much more variable than loads. To exploit the distinct natures of solar generation and load and unlock their predictive potentials, we propose ${\bf SONNET}$, which stands for ${\bf SO}$lar-disaggregatio${\bf N}$-based ${\bf NE}$t load forecasting with ${\bf T}$ransformers. It is a novel probabilistic net load forecasting method based on disaggregating net loads into solar generation and loads and feeding both into the predictors. The method further features a) an enhanced Transformer architecture that integrates both historical and future input data, employing a combination of self-attention and cross-attention mechanisms, and b) a data augmentation method that enhances the robustness of net load forecasts against weather forecast errors. Extensive experiments are conducted based on the comprehensive real-world data set from a recent net load forecasting competition organized by the U.S. Department of Energy (DOE). It is demonstrated that our proposed method both improves the accuracy and reduces the uncertainty of net load forecasts. Notably, our proposed method significantly outperforms the state-of-the-art. The proposed techniques also have broad applications for energy and/or general forecasting-related problems.
Supplementary Material: zip
Primary Area: other topics in machine learning (i.e., none of the above)
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: 7475
Loading