Power flow analysis via typed graph neural networks

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