A precondition generalized minimization residual method based on GPU for static security analysis
Sen Qi,Xin Li,Gengfeng Li,Liang Yi,Zhaohong Bie,Huili Tian
DOI: https://doi.org/10.1109/ciced50259.2021.9556797
2021-01-01
Abstract:In power system, in order to avoid equipment damage caused by overload and over voltage, fast online real-time static Security analysis (SSA) is very important. With power system expanding constantly, the scale of power grid is getting bigger and bigger, which even reach tens of thousands nodes. Therefore, the number of states that need to be calculated is huge, and the traditional serial computing method can’t meet the real-time computing requirements of large power grid any more. The expected fault set increases greatly, resulting in a larger computational burden. To solve this problem, A generalized minimization residual method (GMRES) based on GPU for SSA is proposed. First, the SSA is conducted in coarse-grained parallel, and the power flow calculation in each fault case are allocated by thread. Then, the solution of the modified equation and the formation of the Jacobian matrix in the process of each power flow are designed in a fine-grained parallel way to improve the computing speed and achieve a better acceleration effect. For the solution of the modified equation, due to its large scale and sparse mode, the iterative method is adopted to solve it. In view of the situation that the Jacobian matrix is asymmetric and positive definite, the internal iteration method adopts the generalized minimization residual (GMRES) method to further accelerate the internal iteration convergence. Incomplete LU precondition method further improves the efficiency of fine grain parallelism. At the same time, the Jacobian matrix with high computing time is parallelized to achieve the best overall acceleration effect. Compared with CPU serial calculation, the acceleration effect of parallel large-scale power system SSA based on GPU can reach a large acceleration ratio, and the acceleration ratio of case2383 power system can reach more than 6.17 times.