Probabilistic day-ahead energy price forecast by a Mixture Density Recurrent Neural Network

Probabilistic electricity price forecast (EPF) systemsrepresent a fundamental tool to achieve robust productionscheduling and day-ahead bidding strategies. However, mostEPF methods, including recently proposed deep learning basedtechniques, are still targeting point predictions, following thecommon Gaussian assumption. In this work, we propose anovel probabilistic EPF approach based on the integration ofa Gaussian Mixture layer, parametrized by a Recurrent NeuralNetwork with Gated Recurrent Units, including an L1-normbased feature selection mechanisms. The network is conceivedto approximate general conditional price distributions throughlearning. Moreover, we developed a multi-hours prediction ap-proach exploiting correlations and patters both in hourly andcross-hour contexts. Experiments have been performed on theItalian market dataset, showing the capability of the proposedmethod to achieve accurate out-of-sample predictions whileproviding explicit uncertainty indications supporting enhanceddecision making.