Machine learning of proxies for power systems reliability management

Abstract

Nowadays, electrical power is essential for the functioning of our society. Power system reliability management intends to prevent service interruptions while minimising the socio-economic costs. In order to meet this purpose, it takes decisions based on a reliability criterion which is currently a deterministic criterion. However, several factors such as the ageing of the power systems’ components, the increasing share of renewable energies and the development of the electricity market complicate the work of power system’s operators. All of this calls for a new probabilistic reliability management approach and criterion, that is able to take into account the stochastic nature of a power system. This should allow operators to design and operate their system more economically while ensuring a desired level of reliability. In this context, the European project GARPUR developed a new reliability management approach and criterion. This master’s thesis intends first to study the behaviour of this new reliability management approach and criterion when its use in real-time operation is anticipated in the day-ahead operation planning context, and to build simplified models (called proxies) with machine learning so as to mimic the optimisation tool implementing the criterion. In order to reach these objectives, a database is generated using Monte-Carlo simulations and the optimisation tool developed in the GARPUR project. The dataset is first analysed by using classical exploratory statistical methods and then by computing feature importances derived from tree-based ensemble methods to assess the link between input parameters and target quantities computed by optimising real-time decisions based on the reliability criterion. Finally, several machine learnt proxies are tested to assess their accuracy to predict several target outputs of the decision making program. The analysis demonstrated the importance of the weather on the results. A comparison of the proposed probabilistic reliability criterion with a “pseudo” deterministic criterion based on a predefined list of contingencies has shown that the deterministic criterion is more costly for an identical reliability level. Furthermore, considering the fixed N-1 set of contingencies is typically conservative, but may sometimes be insufficient. The building of proxies also revealed that tree-based ensemble methods seem to be the most accurate estimators. In particular the models based on extremely randomized trees were found to be the most accurate models to predict most target outputs.