Comparison of conventional and simple artificial neural network models for high-performance separation of global solar irradiance components at minutely resolution

Solar radiation components are required by most solar applications, but global horizontal irradiance (GHI) is the only measurement or model output that is usually available. Empirical component separation models separate GHI into its components and are the only practical solution that can ensure the availability of the solar components on a global scale. The growing availability of public observed datasets and the rise of machine learning (ML) have paved the way for a new modeling framework, where conventional and ML-based models now coexist. Many ML techniques have been proposed so far, but they have not been clearly found to improve the best conventional models on a global scale. Moreover, the complexity of some ML approaches is out of the reach of normal users, which is detrimental to their practical adoption in regular applications. This study investigates whether a basic artificial neural network (ANN) with a reduced number of easily accessible input variables can outperform the best conventional separation models on a global scale. Three ANN versions with different input combinations are tested using a global database of 117 radiometric stations, and are evaluated against 13 of the best conventional models. Although two of the ANN models are not conclusively better than the best conventional separation model, proving that ML-based models are not necessarily better than conventional models, the third one is consistently better at nearly all ground sites, reducing the average root mean square error of the predicted direct normal irradiance from $\approx$16% with the best conventional model to $\approx$14%.