Static and dynamic forecasting of solar radiation: a comparative study

The ever-growing global interest in renewable energies and the need to integrate their power generation systems into local electric grids necessitate powerful forecasting tools for estimating the future values of the resources. Motivated by the lack of such studies, this article investigates the potential of dynamic artificial neural networks (ANNs) in global solar irradiation forecasting compared to the well-established static modelling approach. The adopted networks are focused time delay network (FTD), nonlinear autoregressive network (NAR), and nonlinear autoregressive network with exogenous inputs (NARX), as well as the newly suggested distributed time delay network (DTD) and layer recurrent network (LRN), while the static multilayer perceptron network (MLP) was considered as a reference. Long-term high-resolution data of the ener-MENA Network Project of three MENA locations, namely: Adrar (Algeria), Ma’an (Jordan), and Zagora (Morocco), have been used for developing the models. It was found that except for LRN, all dynamic algorithms have lower prediction accuracy in sunshine-based modelling compared to MLP. Meanwhile, most of the dynamic models have the same or significantly better performances in independent and temperature-based modelling. Although dynamic ANNs are slow to train and cost extensive to optimize, they are generally less sensitive to the sky clearness index and have more stable performances when new data are introduced in the forecasting phase.