Probabilistic Forecasting Using Monte Carlo Dropout Neural Networks

摘要

Using artificial neural networks for forecasting tasks is a popular approach that has proven to be very accurate. When used to estimate prediction intervals, a normal distribution is usually assumed as the data noise uncertainty term, as in MVE networks, while model parameters uncertainty is often ignored. Because of this, prediction intervals estimated by them are narrow in uncertain regions where train data is scarce. To tackle this problem we apply Monte Carlo dropout, which is a model uncertainty representation technique, to the network parameters of a Long Short-Term Memory MVE network, allowing us to construct better prediction intervals in probabilistic forecasting tasks. We compare our proposal with the pure MVE method in four wind speed and one consumer load real forecasting datasets, showing that our method improves results in terms of the Winkler loss in both one step ahead and multi-step ahead probabilistic forecasting.