Busbar splitting allows transmission system operators (TSOs) to reconfigure substation connectivity and redistribute power flows at zero redispatch cost, making it one of the most efficient tools for congestion management in transmission grids. Despite its operational importance, systematic optimization of busbar configurations across large transmission grids remains an open research problem.

Existing optimization approaches including MILP-based heuristics, exact enumeration methods, and learning-based candidate selection predominantly evaluate candidate topologies using linearized DC power flow, sacrificing voltage accuracy and reactive power modeling. This approximation is particularly consequential in voltage-sensitive operating conditions, which often represent the scenarios where topology optimization yields high operational value.

Moreover, a limitation of standard evolutionary methods is that they treat switching actions as structurally independent, disregarding the electrical coupling between actions.

We develop a family of extended genetic algorithms that evaluate candidate topologies using full AC power flow, providing exact N-1 security assessment and voltage constraint satisfaction without DC simplification. The framework simultaneously accounts for multiple operational objectives relevant to TSO practice - including congestion relief, topological depth relative to the reference configuration, total number of switching actions. To contain the cost of AC evaluations, the framework incorporates sensitivity-based local search exploiting Bus Split Distribution and structure-aware crossover operators that preserve groups of switching actions identified from the evolving population. We investigate operator designs at three levels of granularity: individual circuit breaker operations, full substation reconfigurations, and pairs of electrically coupled substations that jointly govern individual transmission lines. We compare their effectiveness against greedy and other evolutionary baselines from relevant literature.

Applied to historical congestion scenarios from a real grid model the extended GA achieves greater congestion relief than the baselines within the same AC power flow evaluation budget. The results demonstrate that exploiting electrical coupling structure at the substation-pair level yields the strongest improvements, particularly in multi-corridor congestion scenarios where independent-action assumptions break down. These findings provide practical guidance for TSOs seeking to deploy automated topology optimization at scale without sacrificing physical accuracy.