Multi-step probabilistic forecasting model using deep learning parametrized distributions

When deep learning models are used to predict the probability distribution of future values, a task called probabilistic forecasting, they need to handle the epistemic and the aleatoric uncertainties. For the former, some are based on Monte Carlo dropout and have focused on prediction interval estimation assuming a normal distribution for the aleatoric uncertainty, without looking into general probabilistic forecasting tasks such as quantile regression or scenario generation. Time series in practice are usually not normally distributed, usually having a non-symmetrical distribution. So, these models need to be adapted with pre- and post-processing ad-hoc transformations to recover a normal distribution. We propose a supervised deep model based on Monte Carlo dropout that handles both sources of uncertainty and can be applied for multi-step probabilistic forecasting, including prediction interval estimation, quantile regression, and scenario generation. For the aleatoric uncertainty, our model allows changing the normality assumption easily to other families of distributions with the same number of parameters, such as Gamma, Weibull, and log-normal. For multi-step prediction, we use a procedure of re-injection of intermediate samples. To validate our proposal, we examine the behavior of the model on wind speed, wind power and electrical load forecasting, important tasks for the energy sector, as electrical grids present more uncertainty due to the penetration of variable renewable sources. We show that generating scenarios with our recurrent approach works better than directly estimating the distribution of all future values, also that considering the epistemic uncertainty makes the model more robust, and that changing the output distribution of the model is a property that may improve the metrics for specific datasets.