Simula@BI seminar: PhD Candidate Pietro Manzoni, Politecnico di Milano
Title: "RNN(p) models for Probabilistic Forecasting: Theory and Applications."
A Recurrent Neural Network that operates on several time lags, called an RNN(p), is the natural generalisation of an Autoregressive ARX(p) model. It is a powerful forecasting tool for time series that exhibit multiple-scale dependencies, as it happens e.g. in the energy sector -- where hourly, daily, weekly and yearly interactions coexist. We present a RNN(p)-based methodology for probabilistic forecasting, addressing two issues that often characterise these models: the lack of computational efficiency and the unreliability of the predictions. In detail: first, we propose a new fast learning algorithm capable of outperforming the ones which are commonly used in the industry; second, we introduce two novel loss functions designed to tackle “overconfidence”, i.e. the underestimation of the predictive uncertainty. We apply the proposed modelling approach to forecast the hourly electricity demand. Experimental results show that the obtained probabilistic forecasts are extremely accurate and reliable; moreover, the analysis also demonstrates the computational efficiency of the new learning algorithm.
A joint work with Roberto Baviera, Politecnico di Milano.