Deep Quantile Transformer for Probabilistic Net Load Forecasting with Extreme Solar Penetration

Net load forecasting has become increasingly challenging due to the growing penetration of solar photovoltaic systems, which introduce significant uncertainties and variability in power system operations. This paper presents a day-ahead probabilistic net load forecasting technique utilizing deep quantile Transformers (DQT). The DQT approach harnesses the unsupervised learning capabilities of Transformer encoders to capture temporal correlations in weather and net load data, while optimizing uncertainty quantification in net load forecasts through the pinball loss. The developed DQT method is applied in the American-Made Net Load Forecasting Prize competition to forecast the net load of four substations with extreme levels of solar penetration. The DQT method achieved a 7th place among over 100 competing methods, with an average 0.0549 continuous ranked probability score and an average 0.1927 forecast skill score.