Rapidly Trainable Large-Scale Probabilistic Heat Pump Load Forecasting: A Kernel Density Estimation Approach

As the electrification of heating accelerates, accurate and scalable probabilistic forecasting of heat pump loads becomes increasingly important for grid operators, utilities, and other practitioners. While state-of-the-art methods such as quantile gradient boosting offer high forecast accuracy, they are often computationally intensive and difficult to scale for large datasets. To address this challenge, we propose a Kernel Density Estimation (KDE)-based approach for probabilistic load forecasting in households with heat pumps. Using real-world data from 1,193 Swiss households, we benchmark KDE against quantile gradient boosting regarding forecast accuracy—measured via pinball loss—and computational efficiency. Our results show that KDE performs slightly better in 8 out of 9 quantiles while reducing training time by 3500×, though at the cost of higher inference time due to its sampling-based quantile estimation.