Deep reinforcement learning for global maximum power point tracking: design and experiments in real PV systems
- Tesis/Trabajos de Grado 
To ensure the efficient operation of photovoltaic systems, the maximum power point (MPP) must be found under different environmental conditions using a maximum power point tracking (MPPT) algorithm. Conventional methods such as the perturb and observe (P&O) algorithm tend get stuck at local MPP (LMPP), not being able to find the global MPP (GMPP). In this research, we explore the integration of deep reinforcement learning (DRL) using a deep-q-network (DQN) agent to tackle the GMPP problem in real-time experiments. The main contribution of this work is a comparison between the DQN agent against the P&O algorithm for GMPPT under real uniform and partial shading (PS) conditions and a pipeline for testing DRL models. An open repository that includes PCBs schematics and layout, and also the code used to train and deploy the models will be provided in a future publication. We show that the DQN agent was able to outperform the P&O algorithm in simulations in every scenario, while in real test benches, it did not happen. However, when the P&O algorithm got stuck in a LMPP in PS scenarios, the DQN algorithm was able to extract up to 63.5% more power than the P&O algorithm.