DOI: 10.29130/dubited.1907458 ISSN: 2148-2446

A Hybrid Model Combining Neural Learning and Uncertainty Modeling for Estimating the Average Phase Voltage in Three-Phase IGBT Inverters

Faruk Kürker
Show PDF Cite
Accurate modeling of power electronics-based inverter systems is very important in the analysis of electric drive systems, as well as in the development of sophisticated control systems and condition monitoring systems. Nevertheless, the non-linear nature of inverter systems, owing to their switching characteristics, and the variability in system parameters make it difficult to accurately model average phase voltages using analytical models. In this study, a hybrid data-driven model is presented, which combines residual learning with heteroscedastic uncertainty modeling in the estimation of average phase voltages using the pulse width modulation (PWM) period in two-level three-phase insulated gate bipolar transistor (IGBT) inverters. In this study, the reference voltage is estimated using the physical operating principle of the inverter system, and this is combined with the use of an artificial neural network in the development of the residual learning model. In addition, probabilistic voltage estimation is performed in this study to assess the reliability of the model outputs. The framework is validated using an open-access three-phase inverter dataset. The results show that the physics-guided residual model substantially improves the standalone physical reference and direct MLP estimation, achieving RMSE values in the range of 1.05–1.11 V, MAE values in the range of 0.59–0.62 V, and (R2≈ 0.99990) for three-phase voltage estimation. The uncertainty-aware extension provides reliability-related uncertainty indicators that may be useful for monitoring, diagnostic, and supervisory-level applications; however, these uncertainty estimates should be interpreted as practical reliability indicators rather than perfectly calibrated probabilities.

More from our Archive