In the context of the energy transition, the power grid must cope with more and more uncertainties and volatility on the power injections and withdrawals. It must adapt in real time to rapidly changing conditions while ensuring an acceptable level of reliability.
To maintain the system in its secured operation domain, the operator has some levers that can be activated preventively or after the occurrence of a disturbance: one consists of changing the power injections/withdrawals – which has an impact on the producers and consumers and is therefore costly; the other consists of re-configuring the grid by acting on the switching devices which has the huge benefit of being costless.
TSO are interested in new tools that would support the operator’s support in finding adequate grid configuration (topology) that:
• ensure a defined level of reliability over a short-term time horizon (from real time to few hours),
• respect some constraints (limited number of simultaneous modifications, minimum time interval between two modifications of the same device. . .),
• provide a set of simple conditional rules that could be applied by the operator after a disturbance to quickly find a grid configuration.
However, this problem is a huge and difficult optimisation problem which is still hard to solve. In this regard, this thesis would investigate possible approaches mixing optimization and Machine Learning, such as:
1. Before the loop Machine Learning: Machine Learning used to reduce the problem complexity and size then solving sub problems with classical optimization methods which could be a sequence of optimization problems.
2. In the loop Machine Learning: Machine Learning included in the optimization engine: oracle of binding constraints, good decomposition, learning best internal heuristics from previous run, etc.
3. After the loop Machine Learning: Machine Learning use the optimal solutions found in step and try to find a better solution using Reinforcement Learning.