Machine learning-based prediction of optimal tilt angles for monofacial and bifacial PV systems

This study introduces a novel machine learning-based framework to predict optimal tilt angles for monofacial and bifacial photovoltaic systems, using six years of high-resolution (5 min interval) satellite irradiance data from 184 U.S. locations. Unlike previous work, it combines fine temporal data, wide geographic coverage, and a comparative evaluation of thirteen machine learning models to optimize tilt angles under three adjustment strategies: yearly, seasonal, and monthly. The irradiance on tilted surfaces was estimated using the isotropic sky model, to efficiently simulate the front- and rear-side exposure across a wide range of albedo values varying from 0 to 1 with 0.1 increments. The resulting tilt angles served as simulated optimal angles for training thirteen machine learning (ML) models using location coordinates and albedo as input features. Model accuracy was validated at 15 independent U.S. cities. For yearly prediction, Support Vector Machine with Gaussian Kernel (SVMG) and k-Nearest Neighbors Regression (KNN) emerged as the top-performing models for monofacial and bifacial systems, respectively, yielding angular prediction absolute errors below 1.0°for monofacial and 1.7°for bifacial, and negligible irradiation discrepancies (less than 0.01% for monofacial and 0.02% for bifacial). Further analysis showed that the amount of energy gained by bifacial panels is strongly dependent on the reflectivity of the ground (albedo), reaching more than 80%, and that the best panel orientation varies in subtle ways depending on the location’s latitude. The proposed ML framework offers an accurate, scalable, and computationally efficient solution for PV tilt optimization in both fixed and periodically adjustable systems.