Probabilistic day-ahead load forecast using quantile regression forests

Load forecast is one of the most important tasks in modern and smart grids. With the integration of renewable intermittent sources and the adoption of demand response strategies, an accurate short-term prediction becomes mandatory. Modern forecast approaches do not merely estimate future values, but provide also confidence intervals with different widths and probabilities. Therefore, this paper proposes a probabilistic day-ahead load forecast approach based on quantile regression forests. Quantile regression forests are extensions to random forests that provide confidence intervals instead of single points. The forecaster inputs are chosen according to measures of correlation and importance, profile analysis and wavelet decomposition of load curves. Several tests are performed using real data sets from the Ontario market. The results reflect the accuracy and the effectiveness of the proposed model under different circumstances.