Promoteur : Prof. François VALLEE
Our society is currently undergoing a major energy transition, mainly driven by an increased penetration of renewable energy sources in electricity grids (e.g. wind and solar), which are highly volatile and partially unpredictable. Moreover, this evolution takes place in a liberalized environment (characterized by competition at both generation and retail levels) where the objective of the different market players is to maximize their own profit.
Operating electricity networks under this framework is hard, and there is a growing need of flexibility that can be provided by pumped storage hydropower plants (where the energy is stored in the form of potential energy by pumping water in an upper reservoir, which can thereafter be released through turbines to generate power in case of need).
The use of such solutions is therefore fully investigated in the Walloon Region through the Smartwater project that aims at evaluating the feasibility of the rehabilitation of old industrial infrastructures into small to medium-sized pump hydro storage stations (from one to tens of megawatt). However, operating these small to medium-sized units alone is not optimal (due to their limited energy capacity) and is subject to multiple nonlinearities (arising from their complex topology). In this work, we demonstrate that accurately considering these nonlinear effects is essential to extract the full economic potential of such underground stations, and the subsequent objective is thus to integrate these characteristics within a computationally efficient formulation of the day-ahead scheduling of aggregators participating in electricity markets.
The resulting decision process is subject to uncertainties (e.g. regarding renewable generation of future electricity prices) that need to be efficiently characterized so as to properly represent the future state of the stochastic decision environment. The goal is to cross the barrier between power systems analysis and machine learning (a research field specialized in learning, extracting and exploiting the complex patterns that are hidden within historical data) so as to provide state-of-the-art predictive tools. Practically, this work capitalizes on recent breakthroughs in Deep Learning (which are based on the use of neural networks with an improved memory management, similar to those exploited by major technology companies for products such as Google Translate or the speech recognition applications in smartphones) to generate more accurate multi-step ahead probabilistic forecasts.