Reinforcement Learning for Automated Power Grid Operation: Can a machine be trained to operate a power grid?
Master thesis
Permanent lenke
https://hdl.handle.net/11250/2787114Utgivelsesdato
2021Metadata
Vis full innførselSamlinger
- Studentoppgaver (TN-IER) [147]
Beskrivelse
Full text not available
Sammendrag
The increasingly high demand and amount of renewable energy sources in the power grid have added more complexity to power grid operations. The electrical power system must ensure that the generated power and the consumed power are balanced and the former can be transported to final consumers. Otherwise, there is a risk of a power outage.
When the grid is overloaded, the transport of electric power to all consumers can be compromised due to bottlenecks or line disconnections. One alternative for alleviating the overload is called re-dispatch, which refers to a short-term remedial action plan to deal with overloading by altering the generator's output. The basic assumption is that by lowering one or more generator's output while increasing one or more generator's output, the total power remains the same, but the congestion is alleviated or removed.
This thesis aims to use reinforcement learning to manage the congestion in a power grid through re-dispatch. The reinforcement learning algorithm can modify the generator's output according to the generator's ramp rate. We calculate the number of discrete time-steps the algorithm can eliminate the congestion and avoid a power outage. We trained two reinforcement learning algorithms to perform re-dispatches on the power grid. The best-trained reinforcement learning algorithm managed congestion in 33.5\% of the 200 scenarios evaluated.