Performance Enhancement of DVR Using Adaptive Neural Fuzzy and Extreme Learning Machine-Based Control Strategy

This paper proposes adaptive voltage control strategies for a three-wire dynamic voltage restorer based on a self-adaptive fuzzy neural network for estimation of fundamental weight components and optimal extreme learning machine control for smooth voltage regulation. The SAFNN overcomes the sequential tuning process of classical ANFIS controllers by simultaneously optimizing the structure and parameters of unknown system dynamics. The construction of SAFNN does not require control base rules in the initial phase and this enhances the learning capability to track the approximated voltage error. The proposed SAFNN employs the linear transformation of the input signal to optimize the required fuzzy rules and speed up the online learning process to achieve the higher accuracy of the control results. The meta-algorithms whale optimization algorithm and Harris hawks optimization is employed for online dynamic behavior identification. The SAFNN-WOA optimizes the membership function to estimate direct and quadrature axis voltage for prediction performance in terms of fast-tracking predefined signals, fewer oscillations, and voltage compensation ability. The dc-link voltage is accurately estimated by employing OELM-HHO and provides improved power quality performance during voltage disturbances. The OELM has the inherent feature of randomly generating the initial weights which effectively evaluates the output weight during the online training procedure and enhances the learning rate. The statistical tools MSE, RMSE, ME, SD, and R during the training stage were evaluated as 168.9052, 12.9964, − 5.5524e−07, 12.9965 and during the testing stage the value are 168.9611, 12.9985, − 0.11566, 12.9986 and 0.78654. These statistical assessment values confirm the goodness-of-fit of the predictive model. The result shows that the proposed hybrid controls SAFNN-WOA and OELM-HHO-based DVR significantly achieves the prediction accuracy for both training and test dataset. Nevertheless, the proposed intelligent DVR model outperforms the other state-of-the-art methods.