Deep Learning for High-Resolution Solar Irradiance over the Arabian Peninsula: From Radiative-Transfer Emulation to Observation-Constrained Probabilistic Mapping

Abstract

Surface global horizontal irradiance (GHI) is a primary input to solar resource assessment, plant design, and generation forecasting, yet in high-potential but sparsely gauged regions such as the Arabian Peninsula, it is difficult to obtain accurately and with quantified uncertainty. The gridded fields of numerical weather prediction offer complete coverage but are costly to produce at high resolution and are systematically biased, mainly through the misplacement of clouds. Satellite retrievals resolve cloud structure but are indirect and footprint-averaged, and ground stations measure GHI directly but only at scattered points. This thesis addresses both the cost of the modelled field and its reconciliation with observations, in two parts. The first part develops a deep-learning surrogate of the radiative-transfer computation of GHI in WRF-Solar. It introduces a two-stage spatio-temporal neural network design that separates a clear-sky baseline from a learned cloud-radiative correction, trained on a 5 km hourly reanalysis of the Arabian Peninsula spanning 1980–2023. The surrogate reproduces the modelled irradiance to within a few percent under clear skies and about ten percent under cloud, with a monthly coefficient of determination above 0.95 throughout the year, and reduces a 36-hour simulation from roughly fifteen minutes on a 192-core cluster to about ten seconds on a single GPU, without retraining for a new domain or period. A faithful surrogate, however, inherits the biases of its host model, so the second part turns from computational cost to accuracy. It formulates a two-stage probabilistic framework that fuses WRF-Solar outputs, the CERES and SARAH-3 satellite retrievals, and in situ K.A.CARE observations into a calibrated GHI map. A deterministic backbone first combines WRF-Solar outputs and satellite observations via cross-attention to predict a station-calibrated mean alongside a per-pixel aleatoric variance. Then, a residual diffusion model samples the correction, treating ground stations as assimilated truth while using the satellite and WRF fields as complementary priors. This ensures regime-dependent fusion rather than collapsing onto any single source. At inference, stations are assimilated through soft guidance to produce a spatially coherent ensemble whose mean represents the best estimate and whose spread defines the uncertainty field. Evaluated against withheld stations, the framework improves upon both the reanalysis and the satellite retrievals, yielding an uncertainty that contracts under clear skies and near observations, while widening under cloud cover and across data-sparse regions. Together, the two parts form a unified pipeline: an inexpensive, high-resolution modelled GHI, corrected against observations and equipped with calibrated confidence.