Scalable Power Grid Control with Reinforcement Learning Why

We are in the middle of a climate crisis, renewable energies must readily be implemented in the power grids. However, solar and wind power integration in the grid is challenging, their production depends on the weather, and our overall storage capacity is insufficient. We think artificial intelligence can assist grid operators in tackling these new challenges.

What

I developed a hierarchical multi-agent reinforcement learning system. Each agent in the system is a reinforcement learning agent and perceives its environment through a graph neural network. The environment is a power grid modelled as a graph.

How

The whole system is implemented in Python and the reinforcement learning agents are implemented from scratch with Pytorch. I scaled the system to almost 40 agents by running them concurrently with Ray.