Deep Reinforcement Learning for the Optimal Angle Control of Tracking Bifacial Photovoltaic Systems

Abstract

An optimal tilt-angle control based on artificial intelligence (AI control) for tracking bifacial photovoltaic (BPV) systems is developed in this study, and its effectiveness and characteristics are examined by simulating a virtual system over five years. Using deep reinforcement learning (deep RL), the algorithm autonomously learns the control strategy in real time from when the system starts to operate. Even with limited deep RL input variables, such as global horizontal irradiance, time, tilt angle, and power, the proposed AI control successfully learns and achieves a 4.0–9.2% higher electrical-energy yield in high-albedo cases (0.5 and 0.8) as compared to traditional sun-tracking control; however, the energy yield of AI control is slightly lower in low-albedo cases (0.2). AI control also demonstrates a superior performance when there are seasonal changes in albedo. Moreover, AI control is robust against long-term system degradation by manipulating the database used for reward setting.