Abstract: The Modern power systems face rising complexity from demand expansion and renewable integration, requiring faster and more accurate Available Transfer Capability (ATC) estimation. Real-time congestion management is limited by traditional approaches' lack of scalability and inability to capture dynamic spatiotemporal changes. This paper suggests an integrated architecture that combines Graph Neural Networks (GNN) to describe spatial grid dependencies, Conditional Variational Autoencoders (CVAE) to reduce dimensionality, and hybrid VARMAx-LSTM to anticipate multivariate temporal trends. While Physics-Informed Neural Networks (PINNs) effectively solve optimal power flow under physical restrictions, Reinforcement learning via Proximal Policy Optimization (PPO) allows real-time control. A scalable, high-fidelity solution for real-time grid operations is demonstrated via validation on IEEE 30, 39, 57, and 118-bus systems, which demonstrates a 40% reduction in computing time, ~95% congestion prediction accuracy, and 10–15% ATC improvement.

Keywords: Transmission Network Congestion, Available Transfer Capability, Power System Reliability, Transmission Losses, Operational Costs, Congestion Pricing Generation Rescheduling, Artificial Intelligence

Pages: 119-132