Power flow analysis via typed graph neural networks

Abstract

Power flow analyses are essential for the correct operation of power grids, however, electricity systems are becoming increasingly complex to analyze with the conventional numerical methods. The present work proposes a typed graph neural network based approach to solve the power flow problem. The neural networks are trained on benchmark power grid cases which are modified by varying the injections (load and generation), branch characteristics and topology. The solution to the power flow analysis is found when all voltage values are known. The proposed system infers the voltage magnitude and phase and is trained so that the obtained values minimize the violation of the physical laws that govern the system, this way the training is achieved in an unsupervised manner. The proposed solver has linear time complexity and is able to generalize to grids with considerably different conditions, including size, from the grids available during training. Though the training is unsupervised and does not suppose any ground truth data, the solutions obtained are found to have a close correlation with the conventional Newton–Raphson method. The results are additionally validated by finding the root mean square deviation from the Newton–Raphson method, and the faster, though less accurate, DC approximation method.