Fast time domain simulation of power systems using multilevel preconditioners with adaptive reconstruction strategies

Abstract

Time domain simulation (TDS) is an important tool for the analysis of the dynamic behavior of power systems. TDS is a hard computational problem due to the complexity in solving a sequence of large linear systems based on Jacobian matrices. Iterative solvers with various preconditioning techniques have been applied to solve these systems, and among which GMRES is reported to be the most robust. This paper explores the use of multilevel preconditioning technique based on INDependent SETs (INDSETs) and with fill-in guidance. To reduce the system size more effectively, we propose the use of large supernodes instead of single vertices for INDSET choice. Furthermore, a dynamic preconditioner reconstruction strategy is proposed to incorporate the runtime convergence information of the linear system solving process during the simulation. Experiments show that the proposed multilevel preconditioners have a much lower memory usage and computational overhead than their ILU counterparts. By using supernodes and fill-in guidance, we further reduce the size of the multilevel system and total number of nonzeros by 40% and 13%, respectively. The proposed preconditioner reconstruction strategy shows good adaptivity and performance compared with the strategies based on fixed time intervals.