A Machine Learning Approach for Load Margin Assessment in Balanced Islanded Microgrids

Load margin assessment in islanded AC microgrids is crucial to ensure stable operation. Based on this paradigm, this paper first proposes an artificial neural network (ANN)-based approach for this task. Three optimal power flow (OPF) formulations, considering steady-state frequency as a state variable, are employed to generate the training/testing dataset. In addition, distributed generators (DGs) operating in droop control mode are considered together with wind generators operating in PQ control mode. The ANN's inputs comprise a set of phasor voltages and complex power generation of DGs operating in droop control mode, sampled by phasor measurement units. The outputs are the associated load margin for the input vector. The results of two microgrids from the literature show that the prediction stage is carried out in milliseconds, regardless of the system scale, which makes the proposed approach suitable for real-time applications. Secondly, this work proposes a genetic algorithm (GA)-based optimization approach to minimize the required number of measurement units, considering a pre-specified accuracy level for the testing stage. The results show that for the 33-bus microgrid, our optimization technique for PMU placement achieved an accuracy rate of 98.4% while requiring PMUs in only 34% of the nodes of the system. For the 69-bus microgrid, an accuracy rate of 98.5% was achieved while requiring PMUs in only 16% of the system nodes. Finally, it was found that the presence of Gaussian errors has not impacted the accuracy of the predicted load margin.