Parallel Power Flow on Graphics Processing Units for Concurrent Evaluation of Many Networks

The power flow (PF) analysis provides the steady state of the power system and is key to the simulation of transmission networks. It is a tool commonly used by system operators to visualize the effect of generator settings on the network prior to making a change. In situations involving large networks, hundreds or even thousands of PF analysis may have to be run on the network before finding the optimal power dispatch. This process requires significant computation time and does not allow for rapid control of the network. To address this problem, this paper presents two parallel PF solvers that exploit the massively parallel architecture of graphics processing units (GPU) in a hybrid GPU-central processing unit (CPU) computing environment using compute unified device architecture and OpenMP in order to significantly speedup the concurrent analysis of many instances of a network. Both implementations use sparse matrices, double precision operations, and enforce the reactive power limit of generators. The parallel Gauss-Seidel (G-S) and Newton-Raphson (N-R) PF algorithms are tested on networks ranging from 4 to 2383 buses. The accuracy is validated using MATPOWER and the maximum speedup achieved, compared with a sequential execution on CPU, is $45.2 \boldsymbol {\times }$ for G-S and $17.8 \boldsymbol {\times }$ for N-R.