Adéchola KOUANDE

Every day, EDF solves optimization problems to optimally plan the operation of its power generation fleet, ensuring the balance between electricity supply and demand while adhering to constraints. Although the data changes daily, the general structure of the problems remains largely the same. These problems are typically solved using Mixed Integer Linear Programming (MILP), particularly via the Branch-and-Bound method. However, the solvers used rely on generic heuristics, without leveraging the similarity between successive instances. The goal of my current project is precisely to introduce machine learning techniques to speed up the resolution of future instances by capitalizing on data from past resolutions. - Development of heuristics to solve constrained optimization problems related to the management and scheduling of electricity generation resources. - 25% improvement in solving performance compared to the SCIP solver. - Training of artificial intelligence models (GCNN) to accelerate the resolution of optimization problems, achieving a 10% reduction in solving time.

Internship at Télécom Paris applying Deep Learning to image processing. Developed Plug-and-Play algorithms for hyperspectral image unmixing, where a trained denoiser replaces the proximal operator in the proximal gradient descent algorithm. The denoiser was trained on synthetic data generated using Gaussian field methods.