Managing power flows across transmission interfaces is a fundamental challenge in grid operations. When congestion threatens critical transmission corridors, operators must redispatch generation to maintain reliability while minimizing costs and disruption. Traditional approaches rely on optimal power flow calculations, which become computationally intensive as grids grow larger and more complex.
Researchers have proposed a machine learning solution using deep reinforcement learning to automate transmission interface control. The key innovation is a Factorizable State Attribution mechanism that breaks down the entire power system state into separate electrical factor graphs for active power, reactive power, voltage magnitude, and phase angle. This decomposition allows the system to learn which physical factors most influence each specific transmission interface adjustment task.
The framework employs a Dueling DQN policy—a deep Q-learning variant—to select optimal generation redispatch actions. By incorporating interface-specific information, the model develops task-adaptive representations tailored to different adjustment scenarios. This targeted approach improves decision quality compared to generic deep learning baselines.
Validation across three test systems demonstrates substantial benefits. On the IEEE 118-bus benchmark, Chinese 300-bus network, and European 9241-bus system, the FSA-enhanced framework achieved higher adjustment success rates than representative DRL baselines. Critically, inference time was substantially faster than conventional optimal power flow solvers, enabling real-time operational response.
Beyond performance metrics, the factorizable attribution framework offers operational transparency—a significant advantage over black-box AI systems. The learned factor attribution maps reveal which electrical variables most influence each interface adjustment, supporting operator understanding and confidence in automated recommendations.
This development bridges the gap between traditional optimization and modern machine learning for grid control. As transmission systems face increasing renewable integration and dynamic loading patterns, faster, more interpretable automation tools could significantly enhance operational efficiency and reliability across complex interconnected networks.
