Logo image
Soft actor critic based deep reinforcement learning for optimization of global maximum power point tracking in PV systems during partial shading conditions
Dissertation   Open access

Soft actor critic based deep reinforcement learning for optimization of global maximum power point tracking in PV systems during partial shading conditions

Vidhya Viswambaran
Doctor of Philosophy (PHD), University of Greater Manchester
06/03/2025

Abstract

One of the major challenges in Photovoltaic (PV) systems is the effective tracking of the Maximum Power Point (MPP) during Partial Shading Conditions (PSC). Partial shading occurs when only a portion of the PV array is shaded due to obstacles like trees, buildings, or clouds. One possible result of long term partial shading is the hotspot phenomenon, where the undesirable temperature increase causes permanent damage to the solar cells. The utilization of bypass diodes is a common solution to prevent hotspot problems and associated power loss. However, this results in multiple local maxima in the power voltage curve, making it difficult for traditional Maximum Power Point Tracking (MPPT) algorithms to accurately locate the Global Maximum Power Point (GMPP). Addressing this issue is crucial for enhancing the overall energy conversion efficiency and reliability of PV systems. Utilizing novel approaches to improve energy conversion efficiency is crucial in photovoltaic systems and subsequently lower the costs of PV energy. An efficient MPPT controller which can track the optimum operating point of the DC-DC converter which in turn ensure maximum power is transferred from the source to the load during PSC. The research explored the feasibility of employing AI driven strategies to address the limitations of traditional MPPT methods controllers during PSC in PV systems. The novelty of this research is in applying an actor-critic based deep reinforcement learning agent, specifically the Soft Actor Critic (SAC), to detect the global maximum during partial shading. The SAC reinforcement learning agent replaces the traditional controller and acquires knowledge through its interactions within the environment. The proposed Deep Reinforcement Learning (DRL) based approach offers advantages, including the elimination of the need for real world training data. The SAC agent was designed and trained using MATLAB and Reinforcement Learning (RL) Toolbox. The proposed MPPT controller has been designed and tested through simulations in MATLAB Simulink. The proposed DRL MPPT method enhances tracking efficiency to 99.9%, achieving rapid GMP tracking with an average tracking time of 0.304 seconds and a settling time of 0.14 seconds. The MPPT controller tracked the GMP with minimal oscillations around the Maximum Power Point (MPP), resulting in a steady state efficiency of 97%. The results demonstrate superior GMP tracking under both static and dynamic PSC scenarios. The SAC MPPT shows a rapid tracking response and minimal oscillations, enhancing system stability. The SAC MPPT controller improves GMP tracking efficiency compared to traditional and other AI-based MPPT methods. Furthermore, the results are validated using Hardware in the Loop (HIL) testing, showing good agreement with the simulation outcomes.
pdf
SOFT ACTOR CRITIC BASED DEEP REINFORCEMENT LEARNING5.71 MBDownloadView

Metrics

11 File views/ downloads
21 Record Views

Details

Logo image

Usage Policy