<![CDATA[Direct Normal Irradiance (DNI) plays an important role in global climate observation and the optimization of solar photovoltaic systems. Pyrheliometer calibration commonly refers to ISO 9059 by comparing an operational instrument with a reference standard. This study develops a Deep Learning-based virtual verification instrument to reconstruct the output of a CHP1 pyrheliometer while capturing nonlinear relationships between thermopile sensor response and meteorological dynamics. Two bidirectional architectures, Bidirectional Long Short-Term Memory (Bi-LSTM) and Bidirectional Gated Recurrent Unit (Bi-GRU), were trained using multivariate irradiance and meteorological data obtained from a calibration campaign in Kupang City during September 2025. Model performance was evaluated using the coefficient of determination, RMSE, and MAPE. The results show that bidirectional architectures achieved stronger performance than their unidirectional counterparts. Bi-GRU delivered the best performance, with a coefficient of determination of 0.986, an RMSE of 1.153 W/m², and a MAPE of 0.096%, while Bi-LSTM achieved a coefficient of determination of 0.885 and a MAPE of 0.309%. These findings demonstrate the potential of Bi-GRU as an adaptive virtual verification model for pyrheliometer response reconstruction. Further development may focus on validation across different periods, locations, weather conditions, and instruments.]]>
Copyrights © 2026
