Work

What Do Solar Forecasters Learn? Realistic Skill, Cross-Monsoon Transfer and Climate Similarity from Dhaka

Year: 2026
Type: preprint
Abstract: We test how much daily solar-forecasting skill survives realistic input availability and transfer from Dhaka to other monsoon climates. An information-availability ladder spans estimation with target-day meteorology, cloud and clear-sky irradiance (P0c) to measured NASA POWER publication latency (P3: meteorology to t − 3, solar variables to t − 6). After equal-budget validation tuning, best-model R² falls from 0.799 to 0.274, and RMSE skill against climatology from +49.3% to +1.4%; issue-day inputs without publication delay (P2) retain +17.8%. Raw transfer is worse than local references at most of the 22 targets. At six usable ground stations, local climatology plus transferred anomalies brings skill near parity; Ilorin and Gandhinagar are excluded from ground-truth conclusions. Climate distance predicts transfer skill across three sources, but its advantage over geography is source- and model-dependent: Dhaka and Lagos favour climate, while Darwin favours climate only for gradient boosting. Dhaka's advantage shrinks but survives exclusion of dry contrasts. On ten reserved cities, Dhaka P3 climate distance correlates with actual skill at ρ = −0.794 (p = 0.006), versus ρ = −0.261 for geography; isotonic map predictions have Spearman 0.635 and MAE 0.348. These ten cities test transfer from Dhaka alone. Winter haze retains the strongest recurrent-model seasonal skill at Dhaka. An archived operational ECMWF forecast, strictly available day-ahead and linearly corrected online, reaches +23.6% skill against climatology on the 239 days it covers (May–December 2024), against +1.9% for the P3 gradient-boosting model on those days. A ridge post-processor trained at Dhaka improved on raw ECMWF at 17 of 22 targets without local fitting. In an exploratory comparison, transfer failures tracked climate distance more closely than geographic distance. Without aerosol inputs, the forecasts performed close to climatology during the winter haze season at Dhaka. Adding issue-day CAMS aerosol raises corrected ECMWF's haze-season skill to +16.6% and its full-year skill to +23.8%; ECMWF's bias grows with aerosol load, but aerosol corrections do not transfer to dusty climates.
Language: English
Field: Energy
Open Access status: green
OpenAlex ID: W7236580712
Cites:
Cited by:
Related to: 10