Comparative Study of Artificial and Graph Neural Networks for Load Flow Analysis in Modern Power Grids: A Case Study of IEEE 14-Bus and Nigerian 28-Bus Systems

Accurate load flow estimation is critical for the secure and efficient operation of modern power grids, especially with the increasing integration of Distributed Generation (DG). This study compares the performance of Artificial Neural Networks (ANN) and Graph Neural Networks (GNN) for load flow analysis on the IEEE 14-bus and Nigerian 28-bus systems. Reference datasets were generated using the Newton-Raphson (NR) method by varying active and reactive power under three scenarios: steady-state, fault disturbances, and Photovoltaic (PV)-based DG integration. For each scenario, bus voltages and phase angles were extracted and split into training, validation, and test sets. Both models were implemented in MATLAB R2025a and evaluated using MSE, RMSE, MAE, MAPE, Line Voltage Stability Index (LVSI), and Computation Time (CT). Across all scenarios, ANN consistently achieved lower errors and faster convergence, with MSE 0.539–1.079, RMSE 0.734–1.039, MAE 0.037–0.105, MAPE 2.364–5.562%, LVSI 0.778–0.870, and CT 0.08–0.27 min. In contrast, GNN delivered slightly higher errors but stronger voltage stability in medium-sized networks, with MSE 1.088–1.828, RMSE 1.088–1.352, MAE 0.104–0.180, MAPE 4.872–8.903%, LVSI 0.621–0.716, and CT 0.12–0.37 min. Across all scenarios, ANN consistently delivered faster, more accurate load flow predictions. At the same time, GNN exhibited slightly higher errors but greater robustness to network disturbances and DG integration, highlighting a trade-off between computational efficiency and topological resilience. These findings highlight a trade-off: ANN offers fast, accurate predictions for smaller grids, while GNN provides enhanced resilience to network topology changes and DG integration, providing practical guidance for AI model selection in modern power system operation and planning.