The integration of renewable energy into the power system requires rethinking the operating paradigms. In the future, the operations will be much more volatile and dynamic than in the past. Novel operating approaches are needed that consider these new dynamics, otherwise, investments in redundant grid infrastructure to maintain the security of supply become necessary. In this thesis, machine-learning is investigated in approaches that can operate the power system while assessing and controlling the security of the energy supply.

An approach is proposed that allows learning optimal security rules for operations. The proposed security rules are interpretable, accurate, take low data quality into account and are very fast when applied. Case studies show that security rules are many orders of magnitude faster than current security assessments.

Then, a probabilistic approach is proposed to use machine-learning in combination with current security assessments. This approach allows managing the risks when considering machine-learning in operations. Case studies show that this can reduce computations by 95%. On the French power system, the practicality and scalability toward multiple contingencies are validated.

Subsequently, security rules are inferred for a control approach. The control approach minimises risk and operating costs at the same time. A case study shows that risks are kept low while operating costs slightly increase in comparison to an oracle.

In the future, four key areas need attention to move these ideas forward. The first area refers to make use of the full information available (physics and operating data) to make these approaches applicable and the second area surrounds the generation of the data required to train these machine-learning-based approaches. The third area is to generalise for changes in the system such as multiple grid topological structures and the final area is to advance the data-driven control approaches.