The shift towards renewable energy sources and increased consumer demand due to electric vehicles and heat pumps has significantly influenced the electricity generation landscape. This shift has also resulted in a grid that is subject to fluctuating inputs, thus necessitating an adaptive power infrastructure. Research suggests that bus switching at the substation can help stabilize the grid to some extent, and the implementation of Deep Reinforcement Learning (DRL) may drastically cut the computational costs involved in this process.
While these expandable topologies may be useful for the next step, they may also lead to less-than-ideal scenarios. Grid operations often do not factor in autonomous substation activities. Instead, they consider switching several substations in phases. However, DRL studies focused on optimizing grids rarely examine these comprehensive topology methods, possibly due to the computational costs of determining combinations or because of the design limitations of the L2RPN Grid2Op environment, which only allows one substation change at a time.
A new study from researchers at Kassel University addresses this issue by examining the entire electric grid’s topology rather than focusing on individual substation switching operations. The study finds that certain topologies (Target Topologies or TTs) are more stable than others, and it’s more advantageous to get close to these TTs if the current topological state isn’t sufficiently secure. TTs can be achieved from nearly any topology configuration, negating the need to understand specific combinations of substation activities.
The researchers present a search strategy for TTs that meet specific criteria, demonstrating that TTs remain stable amidst instability given a set of existing substation activities. The study incorporated a greedy search component with TTs into a previously reported CAgent method, resulting in a Topology Agent (TopoAgent). During testing on the WCCI 2022 L2RPN challenge’s validation grid, TopoAgent scored 10% higher and lasted 25% longer than the benchmark, indicating that this method might be beneficial for optimizing the grid.
Ultimately, the researchers contend that employing TTs as a greedy iteration barely increases the runtime. They argue that the research community should investigate TTs in combination with DRL, as they could yield significant benefits. The researchers emphasize the importance of these findings to the wider community for discussion and exploration.
