Warm-starting conventional solver for ACOPF prediction using Modified Seagull Optimization Algorithm and physics-informed spatiotemporal graph neural network

Abstract

The penetration of renewable energy has intensified the computational challenges associated with solving the AC Optimal Power Flow (ACOPF) problem. The Newton–Raphson (NR) solver is highly sensitive to initialization and may exhibit slow convergence under dynamic conditions. This research proposes a physics-informed spatiotemporal graph neural network (PIStGNN) framework to generate initial conditions for warm-starting the NR solver to improve computational efficiency. The model integrates graph convolutional networks to capture system topology and recurrent architectures (LSTM/GRU) to model temporal variability. A physics-informed loss function enforces power flow constraints, while a Modified Seagull Optimization Algorithm (MoSOA) enhances hyperparameter tuning. The proposed method is compared with DC and flat start warm-start strategies for IEEE 33-, 57-, and 118-bus systems. The results show that the PIStGNN method maintains high accuracy, with MAE values of 0.0126, 0.0446, and 0.0063 for IEEE 33-, 57-, and 118-bus systems respectively. The proposed PIStGNN initialization maintains the lowest runtime across all test systems. On average, the proposed method achieved 1.3–18% computational time reduction compared to conventional initialization techniques. Overall convergence is faster even with more iterations. The high-quality starting point makes the path to convergence more stable. The physics-informed learning provides efficient and scalable ACOPF solutions.