PINN4PF is an end-to-end deep learning architecture for power flow (PF) analysis that effectively captures the nonlinear dynamics of large-scale modern power systems. It consists of a neural network (NN) architecture that incorporates two key advancements: (A) a dual-head feedforward NN consistent with PF analysis, with activation functions tailored to the net active and reactive power injection patterns; and (B) a physics-based loss function that partially incorporates power system topology information via a novel hidden function. We demonstrate the effectiveness of the proposed architecture on 4-bus, 15-bus, 290-bus, and 2224-bus test systems, and compare it with two benchmarks: a linear regression model (LR) and a black-box neural network (MLP). The comparison is based on (i) generalization ability, (ii) robustness, (iii) the impact of training dataset size on generalization ability, (iv) the accuracy of approximating derived PF quantities (specifically, line current, line active power, and line reactive power), and (v) scalability. The results show that PINN4PF outperforms both direct metrics, such as generalization ability, as well as derived physical quantity approximation, by up to two orders of magnitude.
