Most of the developed countries are committed to an energy transition which consists in the substitution of conventional energy sources by new renewable energy sources (RES), in particular wind turbines and solar panels. As opposed to conventional energy sources, new RES are distributed, non-dispatchable, fluctuating and inertialess. This change in electricity generation might increase the vulnerability of power systems if their dynamics can’t be predicted accurately and rapid protective actions undertaken. It is therefore paramount to develop efficient methods and models that, for instance, detect the occurrence of a fault in the system and are able to locate it. For online applications, it is furthermore mandatory that these methods can be computed and updated rapidly. We introduce and study a hierarchy of power system models starting from physics-ignorant standard Machine Learning methods and ending with physics-informed methods in which some information about the system (e.g. physical laws) has been encoded. We compare these methods on performing two tasks: detecting the location of a fault in the system and discovering the equations governing power system dynamics. In particular, we analyze the amount of data needed to train them and how well they generalize.
