Dynamic machine learning calibration of satellite-derived solar irradiance in arid climates with accuracy–cost tradeoffs

Accurate solar radiation data are critical for optimizing solar energy systems, especially in arid regions where irradiance exhibits strong spatial and temporal variability and ground observations are sparse. This study intro duces dynamic ML calibration for biased satellite solar irradiance using ground networks, explicitly addressing the trade-off between predictive accuracy and the operational costs of ground monitoring infrastructure. Utilizing five years of minute-resolution ground measurements from 13 sites across Qatar, we evaluate five bias-correction mod els: Feed-Forward Neural Networks (FFNN), Gradient Boosting Machines (GBM), Random Forest (RF), Empirical Quantile Mapping (EQM), and Linear Regression (LR). Model performance for global horizontal irradiance (GHI) and direct normal irradiance (DNI) is assessed using relative root mean square error (rRMSE) and relative mean bias error (rMBE). Our results demonstrate that while rRMSE systematically decreases as the number of adaptation sites increases, the network exhibits diminishing returns beyond a moderate size. Among the candidates, the FFNN achieves the most significant error reduction over the baseline (similar to 30-40%) for both GHI and DNI, followed closely by GBM and RF (similar to 23-37%), while EQM and LR exhibit only marginal improvement. To support application-specific network design, a generalized linear objective function is formulated to jointly optimize prediction error, the number of active monitoring stations, and their spatial distribution using user-defined weighting factors. Based on this framework, a techno-economic analysis identifies an optimal configuration of seven solar monitor ing stations for the considered deployment scenario. These findings provide a practical basis for periodic satellite irradiance adaptation and cost-effective monitoring-network planning in predominantly arid environments.