Optimization of Reconfigurable Islanded Microgrids using Random Forest Classifier

In this paper, a classifier is developed as an approach to find the optimal configuration of islanded microgrids. In islanded microgrids with high penetration of renewable sources, the power generation may be intermittent and unpredictable. Moreover, even when forecast information is available, the non-dispatchable nature of these generation units further limits the control variables needed to formulate and address an optimization problem. In this regard, reconfigurable microgrids allow controlled changes in the grid topology to redirect and redistribute the power flow, in order to optimize and/or improve the system resiliency. In these scenarios, the optimization variables are the binary status (closed/open) of the controllable switches, which makes the problem particularly suitable to be addressed by decision classification trees. In this study, the optimization objective is power loss minimization, subject to the system constraints of power flow and supply/demand balance. Initially, a decision tree classifier is introduced and tested on a simple 9bus islanded system, to identify and categorize different generation and loading level profiles of the system and learn from them the optimal configurations. After that, a random forest classifier is designed as an ensemble of decision trees with enhanced capabilities. A time-series learning component is also implemented to boost the time-related learning characteristics of the classifier, such as trend and seasonality, which are inherent to the power generation levels of renewable energy sources. The proposed random forest classifier is tested on the modified IEEE 33bus islanded microgrid test system. Simulation results show the random forest classifier, when sufficiently trained, is able to find the optimal configuration of the microgrid to any new generation and loading profile.