Parallel-Levenberg-Marquardt-Method-Based Power Flow Algorithm for Large-scale Ill-conditioned Systems With GPU-CPU Framework

With the expansion of power systems, a parallel Levenberg-Marquardt (L-M) method based power flow algorithm with powerful robustness and high convergence rate for large-scale ill-conditioned systems is proposed in this paper. Further, an implementation of the proposed algorithm based on the GPU-CPU framework is realized. GPU and CPU multi-threading parallel techniques are used to accelerate the algorithm from two perspectives: Jacobian matrix parallel formation and the parallelization of the L-M method. Numerical experiments show that the proposed algorithm has better robustness and efficiency than the traditional Newton-Raphson method and some other power flow methods for ill-conditioned systems. The implementation based on the GPU-CPU framework achieves a significant acceleration effect. Especially for large-scale systems, GPU parallel computing accelerates Jacobian matrix formation with a speedup ratio of over 20 times, and CPU multi-threading computing reduces numerous failed trial steps, over 1/3 of the total trial steps, in the L-M method.