Probabilistic Energy Forecasting Based on Self-organizing Inductive Modeling

Abstract

Self-organizing inductive modeling represented by the Group Method of Data Handling (GMDH) as an early implementation of Deep Learning is a proven and powerful data-driven modeling technology for solving ill-posed modeling problems as found in energy forecasting and other complex systems. It develops optimal complex predictive models, systematically, from sets of high-dimensional noisy input data. The paper describes the implementation of a rolling twelve weeks self-organizing modeling and probabilistic ex ante forecasting, exemplarily, for the Global Energy Forecasting Competition 2014 (GEFCom 2014) electricity price and wind power generation forecasting tracks using the KnowledgeMiner INSIGHTS inductive modeling tool out-of-the-box. The self-organized non-linear models are available analytically in explicit notation and can be exported to Excel, Python, or Objective-C source code for further analysis or model deployment. Based on the pinball loss function they show an overall performance gain of 67.3% for electricity price forecasting and 47.5% for wind power generation forecasting relative to corresponding benchmark measures.