To enhance fault resilience in microgrid systems at the energy management level, this paper introduces a novel proactive scheduling algorithm, based on uncertainty modelling thanks to a specifically designed neural network. The algorithm is trained and deployed online and it estimates uncertainties in predicting future load demands and other relevant profiles. We integrate the novel learning algorithm with a stochastic model predictive control, enabling the microgrid to store sufficient energy to adaptively deal with possible faults. Experimental results show that a reliable estimation of the unknown profiles’ mean and variance is obtained, improving the robustness of proactive scheduling strategies against uncertainties.