Machine Learning for Rapid QSTS Simulations using Neural Networks

Distribution system analysis with high penetrations of distributed PV requires quasi-static time-series (QSTS) analysis to capture the time-varying and time-dependent aspects of the system, but current QSTS algorithms are prohibitively burdensome and computationally intensive. This paper proposes to identify the key time periods throughout the year that need to be run with QSTS simulation and then predict the number of tap changes for the rest of the time periods using an ensemble of neural networks (NN). This ensemble NN approach is a new method of performing time-step simulations based solely on the input data and demonstrates high accuracy for reproducing the full baseline QSTS analysis while performing the simulation up to 4 times faster.