Loris Di Natale

 

Empa/EPFL Ph.D. Student

Physics-inspired black-box methods for building control

Swiss Federal Laboratories for Materials Science and Technology (Empa)
Überlandstrasse 129
CH-8600 Dübendorf

Email:

GoogleScholar

Linkedin

 

About

I am a Ph.D. student from Ecole Polytechnique Férédale de Lausanne (EPFL) working at Empa since February 2020 and funded by the NCCR automation. I previously obtained a B.Sc. in Mathematics and a M.Sc. in Energy Management and Sustainability, both from EPFL.
My main research interests include physics-inspired (Deep) Machine Learning and Reinforcement Learning (DRL), with an application to the building control problem, to sustain the current energy transition and tackle climate change. Advanced algorithms are indeed bound to support the energy systems of the future through automation, and I want to bridge the gap between the control and Machine Learning communities to design the next generation of algorithms.
More generally, I'm always looking for potential applications of AI in projects related to sustainability to bring the power of machines to ongoing projects where it can help.

 

Ongoing research

Pipeline

Given historical building data (typically from the NEST building, at Empa), we want to build a black-box pipeline to firstly fit a black-box model of the building, then train a DRL agent to control it, and finally apply the learned control policy on the building.

Introducing physics

We typically want to use Neural Networks (NNs) at the core of the building models and DRL agents because of their great expressiveness and accuracy. However, classical NNs suffer from underspecification: they learn everything from scratch without any information on the underlying dynamics. They thus require large amounts of training data to perform well, and, in particular, we cannot guarantee that they respect the laws of physics. This might lead to wrong models (e.g. heating leading to colder temperatures) or to DRL agents that take obviously wrong decisions.
One possibility to tackle this issue is to introduce physical knowledge in the NNs, leading to physics-inspired NNs (PiNNs). We are particularly interested in one category of PiNNs where the idea is to tweak the structure of the network (such as in Convolutional NNs (CNNs) or Recurrent ones (RNNs)).

 

Publications

Physically consistent neural network building models: theory and analysis.
L. Di Natale, B. Svetozarevic, P. Heer and C.N. Jones (2021).
Manuscript in preparation.

Deep Reinforcement Learning for room temperature control: a black-box pipeline from data to policies
L. Di Natale, B. Svetozarevic, P. Heer and C.N. Jones (2021).
J. Phys.: Conf. Ser. 2042 012004. https://doi.org/10.1088/1742-6596/2042/1/012004
Data-driven control of room temperature and bidirectional EV charging using deep reinforcement learning: Simulations and experiments.
B. Svetozarevic, C. Baumann, S. Muntwiler, L. Di Natale, P. Heer and M. Zeilinger (2021).
Applied Energy, 118127. https://doi.org/10.1016/j.apenergy.2021.118127

The Potential of Vehicle-to-Grid to Support the Energy Transition: A Case Study on Switzerland.
L. Di Natale, L. Funk, M. Rüdisüli, B. Svetozarevic, G. Pareschi, P. Heer and G. Sansavini (2021).
Energies. 14(16):4812. https://doi.org/10.3390/en14164812.

 

Student projects supervision

2021 - 2022

  • Master Thesis of Simon Schnellmann on Black-box Model Predictive Control with Physically Consistent Neural Networks
    Professor: Prof. John Lygeros (ETH Zürich)

2020 - 2021

  • Master Thesis of Luca Funk on Switzerland in the Energy Transition: The Role of Hydro Power and Electric Vehicles
    Co-supervisors: Dr. Martin Rüdisüli (Empa), Dr. Bratislav Svetozarevic (Empa)
    Professor: Prof. Giovanni Sansavini (ETH Zürich)

Teaching assistant activities

2021- 2022

  • Model Predictive Control at EFPL, Prof. Colin Jones.

2020 - 2021

  • Model Predictive Control at EFPL, Prof. Colin Jones.

 

Contact me for anything: