Research Group CoSES:
The Center for Combined Smart Energy Systems (CoSES) is a low-inertia, sector-coupled, low voltage active distribution grid laboratory to investigates smart operation strategies for sustainable energy systems at distribution grid prosumer levels. The research of our young group covers data driven model validation and calibration, real-time and optimal energy management, and dynamic control strategies of electrical and heat prosumers. The lab hosts the energy infrastructure of five realistic buildings, with real and emulated DERs, and a bidirectional interaction with the electrical and heating networks. The lab uses Power Hardware-in-the-loop (PHiL) philosophy to implement developed control and simulation models in real-time over a real long reconfigurable electrical grid and an emulated heating grid.
Lab website - https://www.mep.tum.de/en/mep/coses/
Research profile - https://shorturl.at/B245c
Topic background:
First principle modelling, mostly through ODEs, has long been the dominant way to analyze and control dynamical systems. However real-world systems often contain higher order terms and non-linearities, colloquially known as “Missing Physics”. Machine learning tools provide a reasonable toolset to explore the gap between fundamental physics modelling and real measurements to identify, hitherto undiscovered interactions and sub-systems within a larger dynamical system.
Topic description:
In the recent past, we have developed data-driven models for heat pumps and building energy controllers. However, the training data requirements for a purely data-driven approach makes quite unattractive for real-world equipment and apartments. This project will explore the use of physics-inspired neural networks and ML embedded differential equations. The goal would be to take a generic first principle model of a heat pump and a building and identify the missing components in the ODE description of the system. If the entire non-linear dynamical system is identified, we could use standard plant inversion techniques to design precise controllers for even non-linear systems.
Tasks:
• Familiarization with the existing data-driven code repository at CoSES for heat pump and building energy systems.
• Familiarization with the SciML and Lux Julia library, with minimal working examples.
• Building a simulation environment to test the physics identification tool.
• Connecting the tool with real measurements and online feedback
• Designing a controller based on the identified model (optional)
• Preparation of a report describing the key aspects and findings of the conducted work
• Presenting the results to a scientific audience.
• Maintaining a detailed documentation of the code repository
• Participate as a full member of CoSES in all team activities and laboratory duties during your stay.
References:
• Rackaukas, C. et.al., Universal Differential Equations for Scientific Machine Learning, 2021, https://doi.org/10.48550/arXiv.2001.04385
• Julia SciML, Automatically Discover Missing Physics by Embedding ML into Differential equations, https://docs.sciml.ai/Overview/stable/showcase/missing_physics/
• Thümmel, M. et. al, Nonlinear Inverse Models for Control, 4th International Modelica Conference, 2005, https://elib.dlr.de/12298/1/otter2005-modelica-inverse-models.pdf
