The invention discloses a GPU accelerated power flow jacobian matrix LU decomposition method. The method includes the steps that LU symbol decomposition is carried out on a jacobian matrix J in a CPU to obtain sparse structures of a lower triangle transformation matrix L and an upper triangle matrix U, and the sparse structure of J is equal to L+U after symbol decomposition; according to the sparse structure of the matrix U, all columns of the matrix J are layered in a parallelization mode, and data needed by calculation is transmitted to a GPU; a layered LU decomposition kernel function Sparse LU is calculated in the GPU in the sequence of level progressive increase. The mode of combining program process controlling and basic data processing through the CPU with dense floating-point calculation processing through the GPU is used for improving the efficiency of power flow jacobian matrix LU decomposition, and the problem that flow calculation consumes time greatly in electric system static safety analysis is solved.