Transmission System Operators (TSOs) expect Renewable Energy Sources (RES) to participate to the provision of Ancillary Services (AS), in order to substitute conventional dispatchable power plants. A promising solution to ensure a sufficiently reliable provision of AS by weather-dependent RES is to aggregate dispersed plants into a Virtual Power Plant (VPP), so that the total production shows reduced uncertainty. Most AS markets operate at short-term horizon, typically for the day-ahead, and require that the service shall be provided with a minimal frequency of underfulfilment (i.e. reliability close to 100%). Therefore a Wind-PV-based AS offer must be based on an accurate forecast of the production uncertainty.
A probabilistic forecasting model of the aggregated Wind-PV production based on machine learning has been developed in [1] and proved reliable down to the 1%-quantile of the production distribution. However for rare events (quantiles below 1%), the reliability of state-of-the-art machine learning models is known to deteriorate, mostly because of their lack of generalization on unobserved data [2]. We propose here two models specifically designed for a better forecast of the extremes of the distribution: the first model is based on the Extreme Value Theory (EVT) [3], and the second model is based on a quantile regression by a Deep Neural Network (DNN).