ResearchI am a post-doctoral researcher within the systems and modeling research unit of the Montefiore Institute. I work on the GREDOR project funded by the Walloon Region, in the team of Prof. Damien Ernst. This project consists in rethinking the Walloon electrical distribution networks, because of the challenges originating from the increase of distributed and renewable capacity and the evolution of consumers behavior.
More generally, my research interests are in the optimization of electrical power systems. I am particularly interested in electricity generation planning, electricity markets, network operational planning and demand side management. Well, everything that can be cast to a sequential decision making problem under uncertainty and solved with techniques from Mixed Integer Programming, Stochastic Programming, and Machine Learning.
I received my electrical engineering degree in 2006 at the university of Liège. I completed my Ph.D. in 2010 in the systems and modeling research unit of the department of Electrical Engineering and Computer Science under the supervision of Prof. Louis Wehenkel.
I worked for a little more than two years as a consultant and software developer for n-Side. I have contributed to the research and development of the algorithm for coupling the European day-ahead electricity markets.
- Active network management: planning under uncertainty for exploiting load modulation. Quentin Gemine, Efthymios Karangelos, Damien Ernst, Bertrand Cornélusse. 2013 IREP Symposium-Bulk Power System Dynamics and Control -IX (IREP).
- Tree based ensemble models regularization by convex optimization. Bertrand Cornélusse, Pierre Geurts and Louis Wehenkel. 2nd NIPS Workshop on Optimization for Machine Learning, Whistler, Canada, 2009. [Video, Slides, bibtex]
- Supervised learning of intra-daily recourse strategies for generation management under uncertainties. Bertrand Cornélusse, Gérald Vignal, Boris Defourny and Louis Wehenkel. Proc. IEEE Power Tech Conference, Bucharest, July 2009. [bibtex]
- Automatic learning for the classification of primary frequency control behaviour. Bertrand Cornélusse, Claude Wera and Louis Wehenkel. Proc. IEEE Power Tech Conference, Lausanne, July 2007. [bibtex]
My work consisted in devising a new method for solving the online unit commitment problem. In the unit commitment problem, the operator of a set of electrical power plants has to decide which plants must be started or stopped and how to set their generation levels in order to satisfy some load and reserve requirements, for a certain time horizon. This is thus a sequential decision making problem which can be further qualified of stochastic, given the numerous sources of uncertainty that affect it (uncertain load, plant outages,...). The classical approach is to formulate the problem as a Mixed Integer Program, sometimes with an explicit treatment of uncertainty, and to solve it in day-ahead. Then in intra-day the problem is periodically solved with updated information about the exogenous variables, in order to take some recourse actions. For some practical reasons (e.g. the solution time of the algorithm used in day ahead may be very variable), it is desirable to have a method that always quickly provides a good approximation of the recourse actions. In this thesis I investigated the use of machine learning methods for computing offline some recourse policies that can be used to have a fast response in an online setting. This opened many interesting questions, such a how to formulate the learning problem, how to learn a policy that always yields feasible recourses, what is the performance of the learning algorithm, how to take advantage of the prior knowledge of the system in the learning algorithms, what is the relationship with stochastic programming,... Although the computational complexity of the unit commitment problem is large, there is some kind of regularity in the solution space that some supervised learning algorithms were able to identify quite efficiently.