Selective-HeatFlex: Selective Forecasting of Heat Pump Flexibility using Model Confidence

Heat pumps enable demand-side flexibility by adapting energy use to renewable availability and electricity prices while maintaining in-door temperatures within user-defined bounds. Existing heat pump flexibility prediction approaches are deterministic and cannot quantify model confidence, making them less robust to outliers and variations in deployments. This paper proposes Selective-HeatFlex, a novel approach that leverages so-called selective time-series fore-casting to enable selective heat pump temperature and flexibility prediction. Selective-HeatFlex only makes predictions when model confidence is high and the expected prediction error falls within user-defined error bounds. Selective-HeatFlex is built on state-of-the-art selective forecasting methods and combines best-in-class forecasting accuracy with the ability to selectively reject predictions based on their potential error. We evaluated Selective-HeatFlex’s performance using data from 9 real-world households across 3 datasets, comparing it against state-of-the-art flexibility prediction models, as well as deterministic and probabilistic forecasting baselines. Experiments show that Selective-HeatFlex outperforms existing deterministic approaches, achieving up to 83.1% higher temperature forecast accuracy compared to baseline models. Furthermore, by rejecting non-confident predictions based on user-defined error bounds, Selective-HeatFlex reduces prediction error by an additional 31.7% while maintaining 43.4% coverage. Finally, Selective-HeatFlex achieves robust performance on unseen heat pump data distributions, reducing prediction error by up to 32.8% through selective prediction.