A researcher at Turkey’s Selçuk University has conducted a comparative analysis of multiple machine learning algorithms for predicting plane-of-array (PoA) irradiance on bifacial PV panels. The researcher used identical input conditions to predict PoA irradiance on both the front and rear sides of the panels.

“This study presented a comprehensive machine learning–based framework for predicting front-side and rear-side PoA irradiance in a bifacial photovoltaic system using routinely measured meteorological, surface-related, and temporal input variables,” researcher Ayşegül Toprak said in the paper. “This study demonstrates that accurate and interpretable prediction of both front and rear PoA irradiance can be achieved using a compact set of easily measurable inputs.”

The study used synchronized field measurements collected between Nov. 17, 2023, and May 29, 2024, from a vertical bifacial PV testbed operated by the US Department of Energy’s National Renewable Energy Laboratory (NREL) in Golden, Colorado. The testbed consisted of a vertically mounted bifacial PV array positioned close to the ground, along with meteorological sensors, ground-reflected irradiance measurements and six IMT reference cells.

The input dataset included global horizontal irradiance (GHI), diffuse horizontal irradiance (DHI), ambient temperature, wind speed, testbed albedo and a binary reflector variable, as well as hour_sin and hour_cos, which represent the daily solar cycle. Before model training, the researcher removed physically implausible zero values, sensor faults and records containing missing or inconsistent measurements.

Toprak then evaluated six regression algorithms: linear regression, k-nearest neighbors (KNN), support vector regression (SVR) with a radial basis function kernel, random forest (RF), extreme gradient boosting (XGBoost) and a feedforward multilayer perceptron (MLP). Each model was run separately for front and rear PoA irradiance using the same input variables and identical five-fold cross-validation partitions in MATLAB. Predictive accuracy was assessed using root mean square error (RMSE), mean absolute error (MAE) and the Pearson correlation coefficient (r).

The results showed that nonlinear models clearly outperformed linear regression in predicting both front- and rear-side irradiance. Random forest achieved the best performance for front PoA irradiance, with an RMSE of 0.188, an MAE of 0.061 and a correlation coefficient of 0.982. It was followed closely by MLP, XGBoost and SVR, all of which recorded correlation coefficients above 0.97.

Rear PoA irradiance proved more difficult to predict. Random forest again performed best, with an RMSE of 0.236, an MAE of 0.080 and an r value of 0.973, while MLP recorded the same RMSE but a slightly higher MAE of 0.086. Linear regression ranked last for both targets, with RMSE values of 0.682 for front PoA irradiance and 0.593 for rear PoA irradiance.

“The results indicate that front-side irradiance is primarily governed by global irradiance and diurnal solar geometry, whereas rear-side irradiance is strongly influenced by surface-related factors such as ground albedo and the presence of reflective ground cover, confirming the conditional and interaction-driven nature of rear-side irradiance formation in bifacial systems,” Toprak concluded.

The study, “Front and rear plane-of-array irradiance in bifacial photovoltaic systems: A machine learning-based prediction approach,” was published in Energy Reports.