Examination of empirical and Machine Learning methods for regression of missing or invalid solar radiation data using routine meteorological data as predictors

Sensors are prone to malfunction, leading to blank or erroneous measurements that cannot be ignored in most practical applications. Therefore, data users are always looking for efficient methods to substitute missing values with accurate estimations. Traditionally, empirical methods have been used for this purpose, but with the increasing accessibility and effectiveness of Machine Learning (ML) methods, it is plausible that the former will be replaced by the latter. In this study, we aimed to provide some insights on the state of this question using the network of meteorological stations installed and operated by the GIS Research Unit of the Agricultural University of Athens in Nemea, Greece as a test site for the estimation of daily average solar radiation. Routine weather parameters from ten stations in a period spanning 1,548 days were collected, curated, and used for the training, calibration, and validation of different iterations of two empirical equations and three iterations each of Random Forest (RF) and Recurrent Neural Networks (RNN). The results indicated that while ML methods, and especially RNNs, are in general more accurate than their empirical counterparts, the investment in technical knowledge, time, and processing capacity they require for their implementation cannot constitute them as a panacea, as such selection for the best method is case-sensitive. Future research directions could include the examination of more location-specific models or the integration of readily available spatiotemporal indicators to increase model generalization.