End to end prediction of optimal power flow in sustainable power system using physics-informed spatiotemporal graph neural network

Abstract

This paper proposes a physics-informed spatiotemporal graph neural network (PIStGNN) for end-to-end ACOPF prediction in sustainable power systems. The proposed framework embeds physical power flow constraints, grid topology, and temporal dynamics of renewable generation. The model is validated on IEEE 33-, 57-, and 118-bus systems using time-series 60,048 datasets generated in Pandapower. All experiments are implemented in Python and performance are evaluated using prediction accuracy, physical feasibility, and computational efficiency. The results show that system voltage profiles remain within acceptable operational limits of 0.96–1.04 p.u. across all scenarios.  The system efficiency analysis reveals that moderate renewable penetration reduces transmission losses, while excessive penetration increases losses due to reverse power flow and congestion in larger systems. The proposed PIStGNN achieves prediction errors on the order of 10⁻² p.u. for voltage magnitude and angle while delivering speedups of up to 10,135× compared with Newton–Raphson-based solvers. Furthermore, physics-informed recurrent graph models achieve minimum uncertainty values of 1.06 × 10⁻³, 2.88 × 10⁻³, and 1.19 × 10⁻³ for the IEEE 33-, 57-, and 118-bus systems, respectively. The results demonstrate that integrating physics-informed learning with spatiotemporal graph modeling provides an efficient surrogate for ACOPF prediction in sustainable power systems.