Machine Learning for Energy Systems

The future energy system requires balancing energy supply and energy demand through sector coupling, where energy-consuming sectors are interconnected with the power-generating sector to address electricity storage challenges by adding flexibility to the power grid. Yet, sector coupling involves solving many decentralized optimization problems that, in turn, rely on accurate forecasts of local electricity demand and supply. Manually designing such accurate time series forecasting models for a specific use case is time-consuming and challenging because selecting appropriate methods is iterative and requires expert knowledge. This expert knowledge is particularly crucial, as the forecast accuracy is sensitive to various design decisions. It is also foreseeable that the number of knowledgeable data scientists cannot handle the ever-growing demand for time series forecasts in the future. Therefore, this research project aims to automate forecasting models' design and operation to reduce human effort and satisfy the increasing demand for time series forecasts in the future energy system.

To address different scopes of automation, novel automaton levels for time series forecasting are defined that fuse automated model design and operation. The introduced automation levels are not rigid – interim levels are possible. Also, a high level of automation is not practical for every application or is not yet technically feasible. Therefore, this research project spans the different Automation levels 1, 2, 3, and 4.


The five levels of automated forecasting, inspired by the SAE-J3016 standard for autonomous driving of vehicles.