Gaussian process-based Bayesian optimization for data-driven unit commitment

Abstract

Global efforts aiming to shift towards de-carbonization give rise to remarkable challenges for power systems and their operators. Modern power systems need to deal with the uncertain and volatile behavior of their components (especially, renewable energy generation); this necessitates the use of increased operating reserves. To ameliorate this expensive requirement, intelligent systems for determining appropriate unit commitment schedules have risen as a promising solution. This is especially the case for weak power systems with low dispatching flexibility and high dependency on imported fossil fuels. In this work, we introduce a radically new paradigm for addressing the optimal unit commitment problem, that is capable of accounting for the largely unaddressed challenge of the uncertain and volatile behavior of modern power systems. Our solution leverages widely adopted developments in the field of uncertainty-aware machine learning models, namely Bayesian optimization. This framework enables the effective discovery of the best possible configuration of a volatile system with uncertain and unknown dynamics, without the need of introducing restrictive prior assumptions. Based on appropriately selected acquisition function and Gaussian process regression, it constitutes a radically different from existing approaches, which heavily rely on heuristic approximations and do not allow to account for volatile behavioral patterns. On the contrary, it guarantees global optimum solutions in non-convex optimization tasks in the least possible number of trials. The demonstrated results show better performance in terms of total production cost and number of function evaluations, inspiring system operators to better schedule their power networks in the forthcoming, de-carbonized grids.