BR3: Automated and Interpretable Modelling of the Energy Market Using Machine Learning Methods

Green hydrogen is widely regarded as one of the central building blocks of a sustainable energy transition. It should be taken into account that hydrogen is only produced sustainably or "green" if the electricity used for water electrolysis is exclusively generated from non-fossil (i.e. renewable or regenerative) energy sources such as photovoltaics, wind, or hydropower. To ensure the long-term economic viability of water electrolysis, primarily "surplus" energy should be used for hydrogen production. This means that hydrogen should be produced in particular at times when the energy demand (over a longer period of time) is lower than the available energy potential or supply.

In order to accurately forecast this "surplus" energy and thus use it as efficiently as possible, precise modelling of the respective energy supply and demand is required. When modelling the supply, the geographical location of the energy source and the sometimes very strong weather-related fluctuations must be taken into account. Conversely, energy demand is influenced in particular by demographic and industrial factors. Due to the high volume of data and the volatility of various time- and space-dependent influencing variables, the energy balance can so far - if at all - only be described with very specialised, highly complex models. At the same time, due to the high complexity of these models, a meaningful, i.e. comprehensible for humans, analysis of the algorithmically made decisions is almost impossible, so that no statements can be made about the reliability of the models.

This is where the project proposal comes in. Through a suitable combination, adaptation, and extension of modern methods of artificial intelligence (AI), or more precisely, machine learning (ML), various algorithms and software products (with the primary goal of modelling energy supply and demand) are to be designed and implemented. First of all, an algorithm is to be developed with the help of which the enormous quantities of, in particular, time- and space-dependent data can be reduced completely automatically to the minimum possible quantity of informative influencing variables. Subsequently, this algorithm is to be integrated into a user-friendly AI system (also to be developed) called AutoGREEN, which - following the idea of so-called Automated Machine Learning - automatically selects the most suitable algorithm possible from a set of promising ML methods, with the help of which the energy supply and the associated energy demand (at a given location) are to be modelled. Due to the focus on models that were determined on the basis of small, but as informative as possible influencing variables, this also provides the opportunity to specifically analyse the reliability of the generated models.

Doctoral Student: Markus Leyser

First (Main) Supervisor: Prof. Dr. Pascal Kerschke

Second Supervisor: N. N.