Modelling and hardware implementation of sector and region prediction based on SVPWM in three-level NPC inverters

The rapid increase in the utilization of renewable energy sources has markedly heightened the significance of efforts aimed at enhancing the efficiency of energy conversion systems. In this context, inverters have become critical components, as they play a key role in reducing total harmonic distortion (THD) and minimizing power losses. The objective of this study is to achieve efficient and highly accurate sector and region detection in three-level Neutral Point Clamped (NPC) inverters employing the Space Vector Pulse Width Modulation (SVPWM) technique, while simultaneously reducing computational burden. To this end, regression models, artificial neural networks (ANN), and deep learning (DL) architectures were comprehensively evaluated, and the developed models were examined through comparative analyses. In the regression-based first stage, the linear regression model exhibited the best performance with R², RMSE, and MAE values of 0.9722, 0.2452, and 0.2846, respectively, while THD values of 0.29% in simulations and 1.9% in experimental studies verified the quality of the inverter output. In the second stage, ANN-based sector and region detection was conducted to overcome the limitations of classical approaches, and the results indicated that the Narrow and Wide ANN models achieved a leading accuracy of 99.97% in sector detection, whereas the Wide ANN model reached an accuracy of 98.81% in region detection. In the third stage, DNN, 1D-CNN, LSTM, and GRU architectures were analyzed to develop solutions suitable for real-time operation, which is crucial for industrial applications. The DL-based prediction results revealed that the GRU model, distinguished by its low parameter count and high performance, achieved accuracies of 99.27% in sector detection and 97.62% in region detection, making it the most promising candidate. Overall, the findings clearly demonstrate that sector and region detection in three-level NPC inverters can be performed with high accuracy using regression, ANN, and DL methods, and that the GRU-based approach, in particular, offers substantial potential for real-time hardware implementations.