Deep reinforcement learning using deep-Q-network for Global Maximum Power Point tracking: Design and experiments in...

Area of Science:

• Electrical Engineering
• Renewable Energy Systems
• Artificial Intelligence

Background:

• Conventional maximum power point tracking (MPPT) algorithms like Perturb and Observe (P&O) struggle with partial shading conditions (PSC), often getting stuck at local maxima.
• Photovoltaic (PV) systems require efficient MPPT to maximize energy harvest under dynamic environmental factors.

Purpose of the Study:

• To experimentally validate the integration of a Deep-Q-Network (DQN) agent for real-time MPPT in PV systems.
• To compare the performance of the DQN agent against the P&O algorithm under uniform and partial shading conditions.
• To establish a robust testing pipeline for DRL-based MPPT models.

Main Methods:

• Implementation of a DQN agent within a synchronous DC-DC Buck converter for real-time MPPT.
• Experimental testing under both uniform and partial shading conditions.
• Comparative performance analysis against the established P&O algorithm.

Main Results:

• The DQN agent demonstrated superior performance in simulations compared to the P&O algorithm.
• In real-world tests under PSC, the DQN agent achieved up to 63.5% more power extraction than the P&O algorithm, which became trapped at a local maximum.
• A functional testing pipeline for DRL models was successfully established.

Conclusions:

• DRL, specifically DQN, offers a promising approach for advanced MPPT in PV systems, outperforming conventional methods under challenging conditions.
• Experimental validation confirms the potential of DQN for real-time GMPP tracking, even with limitations in direct replication of simulation trends.
• Open-source resources, including hardware designs and code, are provided to facilitate further research and development.