Abstract :
[en] Power systems reliability management aims at taking decisions ahead in time, from several years to a few minutes ahead, so as to facilitate the operation of the system over a future target horizon and ensure a continuous supply of electricity to end-users. As electricity is essential in our modern societies, this activity is critical. However, power systems are currently undergoing major changes. Among others, the increasing penetration of renewable energy, the liberalization of the electricity sector and the aging of the grid result in a large increase in the amount of uncertainties in power systems, complicating the task of the operators and calling for new methods for operation and planning, taking into account these uncertainties.
Focusing on the short-term operation planning context, that aims at taking decisions from several days to several hours ahead to ensure that adequate resources will be available in real-time operation to meet the electricity demand, we propose in this thesis new decision-making tools leveraging machine learning, Monte-Carlo simulations and optimization, to tackle the increasing uncertainties in power systems. In particular, we propose to exploit simplified models, called proxies, of the behavior of the operator in response to realization of uncertainties over the future target horizon considered. These proxies must be fast and yet accurate, in order to replace traditional heavy models of the behavior of the operator and allow one to anticipate the impact of a candidate operation planning decision over a very large number of possible future operating conditions in a short amount of time.
In this thesis, we propose a methodology to build these proxies with machine learning and we illustrate how they can be used for the two aspects of reliability management: reliability assessment (i.e. evaluating the anticipated outcomes of real-time operation) and control (i.e. selecting which decisions to commit).
More specifically, we develop a methodology based on supervised machine learning to build proxies of real-time operation and we show on a case-study that such built proxies have good accuracies and are several orders of magnitude faster than traditional models of real-time operation.
Considering the reliability assessment problem, we combine a variance reduction technique called the control variates approach with our machine learnt proxies in order to speed up the estimation of the induced expected costs of real-time operation for a given candidate operation planning decision. We show that this method yields unbiased estimates of the expected costs of real-time operation while requiring a significantly smaller number of scenarios compared to classical Monte-Carlo techniques for a given target accuracy.
We then generalize this approach to several unseen candidate decisions to further help choosing among operation planning decisions. We show that this approach can be used to rank a list of candidate decisions according to their induced expected costs of real-time operation in order to identify good operation planning decisions.
Motivated by the reliability control problem, we propose to exploit input convex neural networks to build convex approximations of non-convex feasible domains of optimization problems and we demonstrate that the such learnt approximation can be expressed as a set of linear inequalities in a lifted space.
Finally, we review recent works applying machine learning techniques for reliability management, to showcase the potential of these techniques and the progress achieved to date.
