Power flow calculation is the basis of power system calculation, and its core is LU decomposition calculation. Therefore, the key to power system power flow calculation acceleration is LU decomposition acceleration. Currently, parallel algorithms based on central processing units (CPU) have matured and limited space for performance improvement. As a coprocessor, the graphics processor (GPU) has powerful advantages in scientific computing and is widely used in power system power flow calculation. This paper first analyzes the GPU structure and parallel operation architecture, then introduces the LU decomposition principle, and selects the appropriate matrix sorting algorithm and sparse matrix storage model. The GPU-based single LU decomposition is realized by the unified computing device architecture (CUDA) programming model. Parallel acceleration with batch LU decomposition. Finally, five different cases were tested on the simulation device, and the acceleration effect of the parallel algorithm was compared and analyzed. The simulation test results show that the GPU-based batch sparse LU decomposition parallel algorithm can obtain an acceleration effect of 25~50 times on average.