Modele uczenia głębokiego dla regresji probabilistycznej - przykłady zastosowań 
w prognozowaniu pogody i na rynku energii
[Deep learning models for distributional regression, with applications in weather 
and energy forecasting]

Sebastian Lerch (Statistical Methods and Econometrics, Institute of Economics, 
Karlsruher Institut für Technologie - KIT, Germany)


Nowadays, a transition from deterministic to probabilistic forecasting can be 
observed in a variety of application domains. In this talk, I will discuss various 
aspects of designing, estimating and evaluating distributional regression models 
utilizing recent advances from machine learning. In particular modern deep 
learning methods offer various advantages over standard approaches. For example, 
distributional regression models based on neural networks allow for flexibly 
modeling nonlinear relations between arbitrary predictor variables and forecast 
distribution parameters that are automatically learned in a data-driven way rather 
than requiring prespecified link functions.

The focus of the talk will be on applications in weather predictions, and their 
subsequent use in energy forecasting models. Weather forecasts today are based on 
ensemble simulations of numerical weather prediction models. To obtain reliable 
and accurate probabilistic forecasts, these ensemble weather predictions require 
statistical post-processing to correct systematic errors such as biases or an 
underestimation of the true forecast uncertainty. Traditionally, this is 
accomplished with parametric distributional regression models in which the 
parameters of a predictive distribution are estimated from training data. I will 
present recent examples and case studies where modern deep learning methods were 
able to achieve substantial improvements in predictive performance, and will 
illustrate how those methods can contribute to improving renewable energy 
forecasting.