Beyond the Hype - Two Precise Applications of Machine Learning in Active Distribution Grids

/medias/photo/seminairesmktro_1674209450397-jpg

Abstract 

Machine learning is everywhere in energy research—but where does it make a difference? As distribution grids become increasingly active through the integration of photovoltaics, electric vehicles, heat pumps, batteries, and other distributed energy resources, grid operators and planners face two very different challenges: planning future grids at scale and controlling real grids in real time. In this talk, Anurag Mohapatra, Research Group Leader at CoSES, Technical University of Munich, presents two concrete examples of how machine learning can address these challenges—without treating ML as a black-box solution looking for a problem.

1. ML surrogates for large-scale grid expansion planning

Can a machine-learning model reproduce the relevant outputs of computationally expensive MILP-based energy-system models? We explore how deep-learning surrogates can dramatically accelerate the evaluation of future low-voltage grids, opening the door to bottom-up planning studies across hundreds of thousands of distribution networks.

2. Grid-agnostic voltage control on a real LV grid

Can we control voltage without requiring an accurate model of the grid itself? We show how Physics-Informed Symbolic Regression (PISR) can learn an interpretable surrogate of grid behaviour and embed it directly into predictive voltage control—culminating in experimental validation on a physical low-voltage distribution grid. trol on a low-voltage grid,