Fast grid topology reconfiguration
AATO – Advanced Analytics for Topology Optimisation researches the application of reinforcement learning for the complex grid topology optimisation problem.
Today only highly experienced system operators, backed by trial-and-error simulations, are able to perform fast topology reconfigurations. However mainly based on what is known to work, potentially better actions are left untested.
The research into analytical tools has been ongoing since the 1980’s, but approaches using approximations and heuristic methods have yielded no practical results. More recent ML techniques are however showing promise.
Optimising the topology of the grid is a very complex problem. So in a first step, we simplify the problem and focus on creating the reinforcement learning model.
Once the model has optimised solving the simple problem, we raise the bar and introduce more complex grids with the goal to improve the RL-model. In a final step we export real-life situations from the operation process and apply them to our tool.
We are supported by N-Side, AI application experts and colleagues from the business of the Elia Group. Additionally we noticed promising research done with pypownet.
Reinforcement Learning is a Machine Learning method based on the natural learning behaviour of humans. As a reaction to our actions we receive feedback from our environment, abstractly presented in the form of a reward or punishment. That is how we learn to behave appropriately.
The AI in operAtIon
The application to a real process is very challenging. Therefore we start training and testing on generic first test cases, with the goal of extending the tool functionalities later on.
Business use case
Every implementation needs diligent adaptions to the individual process needs. We aim at the day ahead congestion forecast, providing us realistic scenarios.
Humans working with AI
The tool is not intended as a stand-alone solution. The goal is to give secure possibilities to human operators who then decide which action should be undertaken. Strict time constraints need to be taken into account to allow the operators enough decision time.