Abstract
Accurate estimation of power plays an important role in designing and monitoring a bifacial PV (BFPV) system. The way to obtain BFPV power by PVlib, as the common simulation approach, is identical to monofacial PV (MFPV), where the bifaciality concept is utilized as a simplifying assumption, and an effective irradiance is employed to consider both front and back radiation together. In addition, during the recent years, a wide range of data-driven approaches, including semi-empirical models, digital twins, and machine learning techniques such as artificial neural networks and ensemble methods, have been explored for BFPV power estimation. While these methods have shown promising accuracy, most of them still rely on rear irradiance measurements, albedo assumptions, or extensive input parameters, which limits their generalizability in real-world BFPV scenarios. Considering the mentioned points, we presented a novel physical data-driven model for estimation of BFPV power output (called the DBPG2 model). The big advantage of DBPG2 model compared to models in the literature is that it does not need rear tilted irradiance (RTI) or albedo and module bifaciality coefficient. The DBPG2 model is trained and tested for a variety of BFPV systems in different parts of the world, with various configurations and albedo conditions. It includes: (1) A fixed-tilt (FT) system at Technical University of Denmark (DTU), (2) A single axis tracking (SAT) system at DTU, (3) A vertically mounted system in Turku, Finland, (4) One row from Bifacial Experimental Single-axis Tracking (BEST) system of National Renewable Energy Lab (NREL) in the Colorado, the USA, and (5 and 6) SAT systems with black and white ground in EURAC Research, Bolzano, Italy. They are called DTU FT, DTU SAT, Turku VER, NREL BEST SAT, EURAC SAT Black Ground, and EURAC SAT White Ground, respectively. According to the results, due to high changes in the albedo over time and small size of the string, the DBPG2 model has a higher error than PVlib for Turku VER (2.47% vs. 1.60%). However, it outperforms PVlib in all other case studies. The normalized root mean square error (NRMSE) for the DBPG2 model and PVlib are: 0.61% and 0.95% (DTU FT), 0.61% and 0.97% (DTU SAT), 1.88% and 3.01% (NREL BEST SAT), 0.77% and 2.07% (EURAC SAT Black Ground), and 1.44% and 3.85% (EURAC SAT White Ground). The findings show a more significant improvement of the DBPG2 model in comparison to PVlib for high albedo and single axis tracker, which is usually the preferred working condition of BFPV systems.