Bayesian Approach to Energy Load Forecast with Neural Networks

Short- and long-term forecasts have become increasingly important since the rise of highly competitive electricity markets. The latter point is particularly evident within recently liberalized frameworks, as it has happened, e.g. in some European countries like Germany, and it is going to happen in Italy. Lately, forecasting of possible future loads turned out to be fundamental to build efficient energy management strategies as well as to avoid energy wastage. Such type of challenges are difficult to tackle both from a theoretical and applied point of view. In fact, latter tasks require sophisticated methods to manage multidimensional time series related to stochastic phenomena which are often highly interconnected. During last years, the most promising results related to energy-based time series and their forecasting were obtained using machine learning algorithms, particularly with respect to the realization of ad hoc developed, deep neural networks (NNs) approaches. It is worth to mention that deep NNs architectures as, e.g. convolutional NNs and recurrent NNs have shown their power in handling complex temporal data. Nevertheless, mostly because of the large number of parameters these models are based on, such NNs are often hard to regularize, difficult to manage for out-of-sample data as well as outliers, and, moreover, they often suffer from lack of meaning. In what follows, we propose a novel approach to energy load time series forecasting, which is based on tailored realized combination of deep learning NNs techniques and probabilistic programming. Properly merging such approaches, we can include uncertainty components both in predictions and representations. This leads to efficient regularization procedures, with priors, also allowing for a more powerful way to build complex NNs for forecasting data of interest.