Scientific Machine Learning in Local Energy Systems

The focus on the development of control strategies for Local Energy Systems is a hot topic of research for the inclusion of distributed renewable energy resources into the energy ecosystem. Although there has been significant research on the application of pure data-driven Machine Learning methods to develop these control strategies, the paradigm of combining scientific computing with the data-driven approaches through a multi-model method is still unexplored. This paper provides one of the first multi-model approaches of scientific machine learning in a real demonstrator site. The architecture proposed connects Agent-Based Modeling(Anylogic-Java) for stakeholder simulation with Pandapower(Python) for power flow analysis, facilitated by state-of-the-art transformer-based forecasting (Chronos 2) and Model Predictive Control (MPC)(Python). The results demonstrate a reduction of peak load and equivalent battery cycles and better performance in comparison to pure data-driven forecasting.