Deep Reinforcement Learning-Based MPPT for PV Systems Under Partial Shading: A HybridMetaheuristic-Optimized Control Framework.

Abstract

Partial shade and quickly changing environmental conditions make Maximum Power Point Tracking (MPPT) in photovoltaic (PV) arrays particularly difficult. In these situations, traditional algorithms frequently fail to find the global optimum or experience sluggish convergence and steady-state oscillations. To obtain reliable, real-time MPPT under non-uniform irradiance, this research suggests a hybrid control architecture that combines a deep recurrent reinforcement learning (DRL) agent with metaheuristic-based parameter optimization and a digital-twin training environment. While a metaheuristic optimizer (such as evolutionary/Dandelion-inspired search) adjusts learning and control hyper parameters to speed convergence and prevent local maxima, the DRRL agent uses sequence modelling (LSTM) and an actor-critic architecture to learn temporally consistent control policies from real-world and synthetic irradiance profiles. This research suggests a hybrid intelligent MPPT architecture that combines a high- fidelity digital twin environment, metaheuristic-based hyperparameter optimization, and a deep recurrent reinforcement learning (DRL) controller in order to address these issues. To provide adaptive and reliable control decisions, the DRL agent models temporal dependencies in irradiance, temperature, and voltage–current dynamics using an actor–critic architecture improved by Long Short-Term Memory (LSTM). To improve tracking accuracy and convergence speed while avoiding local optima, a population- based metaheuristic optimizer that incorporates evolutionary and Dandelion-inspired techniques is used to automatically adjust learning rates, exploration strategies, and reward parameters. For safe offline training and transfer learning to hardware-in-the-loop (HIL) configurations, a digital twin of the PV string and power-electronics interface is utilized. In comparison to traditional and modern ML-based MPPT techniques, extensive simulation studies and HIL experiments under standardized partial-shading situations show that the suggested framework enhances tracking efficiency, decreases transient settling time, and mitigates steady-state oscillations. Lastly, we highlight future work directions and address sample efficiency, safe exploration, and on-board implementation limits.
Keywords: Deep reinforcement learning, Partial shading, MPPT, photovoltaic systems, and
metaheuristic optimization.