Hybrid physics–machine learning framework for cell-level irradiance modeling and performance prediction in bifacial photovoltaic modules

Abstract

Bifacial photovoltaic (PV) technology enhances solar energy conversion by capturing irradiance from both front and rear surfaces, but accurate performance prediction remains challenging due to complex spatial irradiance distribution and electrical mismatch losses. This study proposes a hybrid modeling framework integrating physics-based irradiance calculations with machine learning techniques for cell-level analysis of bifacial PV modules. The physics-based component combines solar geometry with numerical view-factor analysis to evaluate front- and rear-side irradiance under realistic three-dimensional configurations. An artificial neural network accelerates view-factor computation, reducing simulation time by 65% compared to full numerical methods while maintaining mean absolute error below 0.003. A convolutional neural network detects irradiance anomalies from partial shading, soiling, and cell faults with 96.8% classification accuracy. The framework is validated using experimental measurements from a 395 Wp bifacial module (Voc = 46.8 V, Isc = 10.97 A) under diverse environmental conditions throughout 2025. Results show rear-side irradiance prediction errors below 3% and power output errors within 3.1%. Analysis reveals that nonuniform albedo reflection causes mismatch losses up to 5.8% in south-facing 30°-tilted configurations under medium albedo (ρg = 0.5), while vertical east/west installations reduce losses to 1.2%–2.1%. Ground clearance significantly impacts irradiance distribution, with mismatch losses decreasing from 6.8% at d = 0.3 m to 2.1% at d = 2.0 m. The proposed framework enables accurate, computationally efficient cell-level analysis for bifacial PV system optimization and performance prediction.