KAN-Enhanced Graph Learning for Active Voltage Control in Dynamic Power Systems

The large-scale integration of distributed energy resources has significantly increased the complexity of industrial power dispatch. While existing multi-agent reinforcement learning (MARL) methods leverage graph neural networks for topology-aware voltage control, their ability to capture evolving grid topologies remains limited. Therefore, we propose GKAN-MA, a dual-enhanced MARL framework specifically designed to maintain voltage stability in power systems with highly dynamic topologies and strongly nonlinear voltage-power dynamics. It achieves robust voltage regulation despite frequent grid reconfigurations, while precisely modeling complex relationships between reactive power and nodal voltages. Through persistent topology awareness and accurate nonlinear function approximation, GKAN-MA ensures consistent performance during network changes. Experimental results on IEEE 33-bus and 141-bus systems demonstrate superior controllability and operational efficiency, validating its adaptability to dynamic power system conditions.