ML‐Based Grid‐Interactive EV Chargers Under Nongrid Conditions
Gaurav Yadav, Philemon Yegon
Abstract
The increasing adoption of electric vehicles (EVs) has heightened the relevance of vehicle‐to‐grid (V2G) technology, which facilitates bidirectional power flow between grid‐interactive EV chargers and the electrical grid. Ensuring consistent power quality under nonideal grid conditions remains a significant challenge, as traditional proportional‐integral (PI) controllers exhibit limitations in mitigating total harmonic distortion (THD) and adapting to dynamic grid variations. This study investigates the application of machine learning methods, including decision trees (DTs), artificial neural networks (ANNs), and linear regression (LR), as alternatives to conventional PI controllers. Among these, DTs prove particularly advantageous due to their simplicity, interpretability, and efficiency in managing complex, nonlinear grid behaviors with minimal preprocessing requirements (i.e., training and testing data for mean squared error (MSE) indicator are 0.0039 and 0.003, respectively). While ANN effectively captures intricate patterns, it demands higher computational resources and lacks transparency (i.e., training and testing data for MSE indicator are 0.076 and 0.171, respectively), whereas LR, though computationally efficient, is less suited for addressing complex grid dynamics (i.e., training and testing data for MSE indicator are 0.84 and 0.844, respectively). To further improve system performance, a CNISOGI filter is incorporated for enhanced harmonic attenuation and DC‐offset rejection. The proposed system, evaluated through MATLAB/Simulink simulations and validated using a 1.1‐kW hardware prototype, demonstrates that DT‐based controllers, in conjunction with advanced filtering techniques, significantly enhance THD reduction (i.e., 2.44% in G2V and 3.83% in V2G modes), grid stability, and operational efficiency, offering a promising solution for future smart grid applications.