Using GPUs to simulate photovoltaic power plants: Special cases of performance loss

Abstract

The installed capacity of renewable energy is soaring, and Photovoltaic Power is a sizeable part of it. Photovoltaic panels are being installed on both buildings and in large Photovoltaic Power Plants (PVPPs), reaching up to 2 GW of installed capacity. We need to perform simulations of the state of PVPPs to ensure the stability of the power system and their profitability. Traditional simulations that take into consideration each panel can accurately predict the generated power but are too computationally taxing. To lower the computing cost of these simulations, a GPU-based solver that processes the dynamic model of a PVPP connected to a power system has been developed. It simulates the output of the whole photovoltaic power plant down to the photovoltaic panel level. The characteristics of the power system to be simulated were analyzed in order to parallelize the system. The two-diode model is used for the PV panels, and this model was linearized, resulting in a large sparse matrix with millions of columns that must be solved. Our proposal can simulate a model of the PVPP with both accuracy and low computational cost with no PV panel aggregation required. The system was solved using an adaptive step and helper conductances to aid convergence of the matrix. Parallel column factorization and column grouping is key to achieving a high performant simulation, which makes it feasible to simulate events such as cloud occlusion efficiently. We also studied how the wiring of PV panels affects the global power plant output when panels are shaded and/or broken. It was found that the aggregated model returns a higher generated output, which is key to predicting its output and its profitability. PVPP simulations on the GPU solver are up to 14.3 times faster than on the CPU.