Probabilistic electric load forecasting through Bayesian Mixture Density Networks

This work presents a novel approach to address challenging and still unsolved problem of neural network based load forecasting systems, that despite the significant results reached in terms prediction error reduction, lack suitable indications regarding sample-wise trustworthiness their predictions. The present is framed on Bayesian Mixture Density Networks, enhancing mapping capabilities networks by integrated predictive distributions, encompassing both aleatoric epistemic uncertainty sources. An end-to-end training method developed, aimed discover latent functional relation conditioning variables, characterize inherent stochasticity, convey parameters unique framework. To achieve reliable computationally scalable estimators, Mean Field variational inference deep ensembles are integrated. Experiments have been performed short-term tasks at regional fine-grained household scale, investigate heterogeneous operating conditions. Different architectural configurations compared, showing Continuous Ranked Probability Score tests performance improvements achieved integrating flexible patterns multi-modalities posterior space.