With rising penetration of intermittent renewable energy sources, management of electricity grids is becoming increasingly challenging. Conventional methods for generation scheduling (AKA the unit commitment problem) use deterministic reserve constraints to account for deviations in demand and other contingencies. However, it has been shown that these methods are inadequate in comparison with stochastic optimisation methods which rigorously account for uncertainties. Unfortunately, stochastic methods are computationally expensive, which renders them unusable for high frequency dispatch on large power systems.

Following recent successes of machine learning, there is considerable interest in these methods for a range of tasks in the energy sector. In particular, reinforcement learning methods such as those seen in the highly successful AlphaGo algorithm are promising for planning tasks in complex and uncertain environment such as the unit commitment problem.

In this project, I am developing a reinforcement learning algorithm for solving the unit commitment problem in the GB power system. This algorithm combines a deterministic planning algorithm (Monte Carlo Tree Search) with machine learning to intelligently explore the enormous space of possible schedules. Repeated simulation on a model of the GB power system is used to learn a generalisable policy that can quickly produce commitment schedules on unseen demand profiles. The aim is to achieve the quality of stochastic optimisation methods while remaining computationally tractable. I will explore the resilience of the algorithm in systems with high renewables penetration and limited transmission capacity.