A bio-inspired neuro-adaptive deep reinforcement learning approach for real-time solar tracking system to enhance photovoltaic efficiency

Traditional solar tracking systems have the disadvantages of either hard-coded control logic or wasteful actuator operation, both of which cause poor photovoltaic performance. This paper proposes a neuroadaptive solar tracking system, which is a hybrid of deep learning and reinforcement learning. The suggested NAT-DRL model will incorporate CNN-LSTM to forecast the solar irradiance in real-time and a Proximal Policy Optimization (PPO)-based reinforcement learning agent to adjust the tilt of the two-axis PV panel. The system is fully based on an edge hardware (Raspberry Pi) and has the capability to perform inference with zero latency and autonomous behavior. One-year deployment was carried out at Sitapura Solar Lab Jaipur with three same PV modules under various tracking plans: NAT-DRL, time-based and statistical. The NAT-DRL model interred 4385 kWh of energy per year, which was 36.6 and 12.7 times superior to the static and time system, respectively. A mean tracking error of only 1.2 was obtained, actuator movement was minimized by 24.1 and in real-time policy execution had a latency of only 47.2 ms. ANOVA (p < 0.01) was used to statistically find the significance and Monte Carlo simulation confirmed the system robustness with the uncertainty of 3 % or less. Unlike prior studies, this study presents a real-time, deployable, predictive control framework that unifies forecasting, learning, and embedded actuation executed entirely at the edge. NAT-DRL is a scientifically sound, scalable, and bio-inspired artificial intelligence (AI) method for active solar tracking that improves efficiency in dynamic environmental settings in real-time.