Virtual power plants aggregating multiple renewable energy sources such as Photovoltaics and Wind are promising candidates for the provision of balancing ancillary services. A requisite for the provision of these services is that forecasts of aggregated production need to be highly reliable in order to minimize the risk of not providing the service. Yet, a reliability greater than 99% is unattainable for standard forecasting models. This work proposes alternative models for the day-ahead prediction of the lowest quantiles (0.1% to 0.9 %) of renewable Virtual power plant production. The proposed approaches derive conditional quantile forecasts of aggregated Wind/PV/Hydro production, obtained from tailored parametric models and machine learning models, including a Convolutional Neural Network architecture for predicting extremes. Reliability deviation is reduced up to 50 % and probabilistic skill score up to 18% compared to Quantile Regression Forest. Forecasting models are subsequently applied to the provision of downward reserve capacity by a renewable Virtual power plant. Increased forecasting reliability leads to a higher reliability of the reserve capacity, but reduces the average reserve volume offered by the renewable aggregation.