A Low-Complexity Artificial Neural Network-Based Optimal Droop Gain Design Strategy for DC Microgrids Onboard the More Electric Aircraft

This article proposes a new droop control design method based on a “reversed data training” of artificial neural network (ANN). Conventionally, after data collection, the ANN is used for forward mapping the control variables (inputs) and system response (outputs). After training, the ANN model can be used for optimal control design for each specific system performance requirement either through curve fittings or other optimization methods. In our proposed method, however, a reversed data training process is used. The ANN uses system responses as its inputs and control variables as outputs. By doing so, the ANN can identify the requested control variables directly for a given system performance request. In the example aircraft dc microgrid, multiple generation systems feed a common dc bus with droop control implemented. During the data-generating process, different droop coefficient combinations are used, and the resulting power sharing ratios are stored as outputs. However, the ANN is data reversely trained with power sharing ratios as inputs and droop coefficients being the outputs. Through this example, we have shown that the proposed approach is straightforward and effective to derive the optimal droop gains based on the desired power sharing requests. The proposed approach is tested in both simulation and experiment.