Dynamic Tracking: A Machine Learning Approach for PV Yield Optimization

This work presents a novel framework for photovoltaic (PV) tracking optimization called dynamic tracking. Its core is a method of irradiance prediction using a machine learning surrogate model trained on ray tracer simulations. This approach outperforms ray tracing models in terms of computational speed while upholding accuracy for diffuse and direct irradiance. It is therefore able to predict irradiance distributions with a resolution of 11 × 11 cm2 for all possible tilting configurations over a given weather time series in real time. Spatial and temporal validation under overcast and sunny conditions show that the network accurately reproduces direct and diffuse shading patterns under arbitrary weather conditions. The proposed framework simulates all possible tilt sequences for any single-axis tracking system over daily periods with a 10 min resolution in only 2 min. Using discrete dynamic programming, it then calculates the truly optimal tilting sequence with respect to average irradiation. This method predicts and avoids inter-row shading, thereby increasing yield and outperforming established strategies like backtracking or diffuse tracking methods for any weather condition.