Short-term load forecasting is a fundamental task in reliable and secure power system operation, particularly in the current landscape marked by increased integration of renewable energy sources and electric vehicles, which introduces stochasticity and raises uncertainty. To express uncertainty in load predictions in the form of a probabilistic forecast, prediction intervals are generated. The variability in load values exhibits higher volatility during the day due to increased human activities, contrasting with lower variability at night. Classic methods for constructing prediction intervals cannot correctly model the variability in uncertainty leading to overly conservative prediction intervals. In this paper, we propose a novel approach – conformalized quantile regression – to create more informative, variable-length prediction intervals. Experimental results, based on a real load dataset from the Croatian Transmission System, showcase the method’s superior performance in capturing adaptive-length prediction intervals. This translates to achieving higher coverage with shorter prediction intervals compared to conventional methods.
Keywords probabilistic load forecasting, short term load forecasting, conformalized quantile regression, prediction interval, random forest