Power Flow Analysis using Dynamically Weighted Deep Neural Networks

The computation of nonlinear power flow equations is crucial for optimization and power system control applications. However, obtaining system information such as lines and topology parameters can be challenging for distribution system operators. To address this issue, various data-driven techniques have been proposed, but they face challenges related to generalizability, data quality requirements, and inconsistent solutions. As a result, there is a growing interest in integrating physics-informed neural networks (PINNs) into deep learning techniques. In this study, we introduce a dynamic weight-adjusted physics-informed deep neural network (DW-PIDNN) for solving power flow equations. This approach mitigates gradient imbalance issues during training, leading to improved generalizability, computational efficiency, and convergence rate. We validate the effectiveness of DW-PIDNN using IEEE 5, IEEE 33, IEEE 57, and IEEE 118 test cases for power flow analysis.