Open Access

ARTICLE

Safe and Explainable Reinforcement Learning-Based Intelligent Switching Control for Standalone and Grid-Tied Z-Source Inverter under Uncertain Solar Conditions

Biswanath Hajoary^1,*, Ranjay Das¹, Ganesh Roy², Daijiry Narzary³
1 Department of Electrical Engineering, Central Institute of Technology, Kokrajhar, Assam, India
2 Department of Instrumentation Engineering, Central Institute of Technology, Kokrajhar, Assam, India
3 Department of Electronics and Instrumentation Engineering, National Institute of Technology, Chümoukedima, Dimapur, Nagaland, India
* Corresponding Author: Biswanath Hajoary. Email:

Energy Engineering https://doi.org/10.32604/ee.2026.075305

Received 29 October 2025; Accepted 04 February 2026; Published online 21 May 2026

Abstract

The increasing integration of photovoltaic systems into smart grids requires accurate evaluation of power conversion efficiency and output performance. In this context, Z Source Multilevel Inverters function as voltage boosting converters and offer a certain degree of fault tolerance. However, conventional control strategies such as proportional integral controllers and hybrid optimization-based methods including POA-RFA (Pelican Optimization Algorithm-Random Forest Algorithm) are limited in their ability to maintain dynamic stability, efficiency, and operational safety under varying solar irradiance and load conditions. This study proposes a safe and explainable Deep Q Network based intelligent switching control framework for the Modified Capacitor Assisted Extended Boost Z Source Multilevel Inverter operating in both standalone and grid tied modes. A unified reinforcement learning controller is designed to ensure effective voltage regulation, strict enforcement of safety constraints, and transparent decision making through SHAP (SHapley Additive exPlanations) based interpretability. Simulation results demonstrate that the proposed Safe DQN (Safety-Aware Deep Q-Network) controller achieves a total harmonic distortion of 1.63 percent and an efficiency of 97.8 percent without any safety violations across diverse operating scenarios. In addition, it provides 40 percent faster settling time and 50 percent lower switching losses compared to proportional integral and POA RFA controllers. Explainability analysis confirms that the control decisions are consistent with the underlying physical dynamics of the system. Overall, this work advances safe, adaptive, and interpretable control strategies for renewable energy converters suitable for real time intelligent power electronic applications.

---

Keywords

Deep reinforcement learning; explainable AI; grid-tied inverter; safe control; solar PV systems; Z-source multilevel inverter

---