Machine learning based performance analysis for bifacial photovoltaic systems
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Bifacial photovoltaic technology offers enhanced performance compared to traditional monofacial technology. These modules generate more energy than monofacial modules with a small increase in cost relative to the performance gain because they absorb light energy from both the front and rear sides of the cell. Bifacial arrays have been shown to have a lower levelized cost of energy compared to other technologies, making power prediction for bifacial arrays a high-priority research area. A major challenge in developing these power prediction models is quantifying the impact of ground cover reflectance and installation height on rear-side irradiance. The main research objective for this thesis was to use machine learning to evaluate how ground cover and array height affect bifacial PV array performance. The experimental setup includes a fixed-tilt bifacial PV array with colored tarps to simulate different surface albedos and elevated platforms which simulate different installation heights. Data were collected using irradiance sensors, temperature sensors, and module-level optimizers. Data were logged at 15-minute intervals for a test period spanning over two years. ☐ The final power prediction model employs a two-stage approach: a gradient boosting machine predicts the bifacial gain based on ground cover reflectance, installation height, shading losses, and solar angle and a neural network predicts monofacial power output based on ambient temperature and solar irradiance. The stages are integrated to calculate total bifacial PV production in real-time. This model can simulate bifacial energy yield over any period under any installation conditions. To evaluate the overall impact, a year-long prediction was performed for each combination of ground cover and installation height. The result was a 5.67\% increase in annual energy yield when a white ground cover was installed compared to gravel. The effect of install height is more nuanced; in cases with high ground cover reflectance, a lower installation height had higher production whereas in cases with low ground cover reflectance, a greater installation height had higher production. This is likely due to the relative impact of albedo irradiance and diffuse horizontal irradiance. ☐ The highest-producing simulation (white tarps with low installation height) had a bifacial gain of 19.74\%, which is a 6.02\% increase in total annual energy yield compared to the control (gravel with high installation height). The lowest-producing simulation (black tarps with low installation height) had a bifacial gain of 12.90\% and a -0.003\% decrease in total annual energy yield compared to the control. These insights can inform economic viability studies for determining levelized costs of energy for new bifacial PV installations. These studies can compare the relative increase in energy production with the cost associated with installing and maintaining the specific site conditions. The power prediction methodology developed in this study can also be deployed in utility power forecasting models for improving day-ahead or minute-to-minute energy forecasts needed for maintaining grid stability. Finally, combining the machine learning techniques used here with physical vector-based simulation techniques already in place in software such as SAM or PVSyst further improves accuracy and model interpretability.
Description
Keywords
Bifacial gain, Bifacial photovoltaics, Ground cover reflectance, Machine learning, Photovoltaics, Power prediction