Transferring day-ahead solar post-processing to sites without local history: a pre-registered test of aerosol encoding, climate-analog training pools and shift-aware conformal intervals

A site without a measurement history can obtain free day-ahead forecasts from ECMWF, NOAA GFS and the Copernicus Atmosphere Monitoring Service (CAMS), but any statistical post-processor for it must be trained on other sites. We pre-registered three hypotheses about how to do this and tested them once on ten reserved cities with no earlier weather-forecast or post-processing result, after an exploratory development stage on 23 sites that were already used in a companion study. The target is daily NASA POWER global horizontal irradiance, and the post-processors are ridge regressions on day-ahead weather forecasts and lagged POWER history, trained on one year and tested on the next. One of the three primary hypotheses was confirmed, and it concerns the form of the aerosol input. Raw issue-day CAMS inputs (aerosol optical depth, surface dust and PM2.5) make transferred post-processors fail badly at some targets: RMSE skill against climatology fell to −620% for Dhaka to Riyadh in development, and to −231% at Karachi averaged over 23 source sites. A bounded, daylight-weighted aerosol transmittance removed these failures. Its transferred skill exceeded that of the raw inputs at 10 of 10 reserved cities (Holm-adjusted p = 0.003). Column aerosol optical depth alone did the same in a secondary check (9 of 10), but neither encoding beat a post-processor without aerosol inputs (4 of 10). Training on the five climatically nearest sites did not beat the five geographically nearest (4 of 10). Climate-weighted, scale-normalised conformal intervals were not better calibrated than naive ones (3 of 10). Intervals calibrated without target-site irradiance labels under-covered at the new cities (median coverage 0.85–0.88 for a nominal 0.90), most of all at Lima and Colombo, whereas intervals calibrated on one year of target data stayed close to nominal (median 0.885). Three development-stage gains did not replicate on the reserved cities.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-19
DOI
https://doi.org/10.5281/zenodo.24374337
Primary Topic
Solar Radiation and Photovoltaics
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Transferring day-ahead solar post-processing to sites without local history: a pre-registered test of aerosol encoding, climate-analog training pools and shift-aware conformal intervals

S. M. Abdulla Al Mamun
Zenodo (CERN European Organization for Nuclear Research)
Solar Radiation and Photovoltaics
preprint

Transferring day-ahead solar post-processing to sites without local history: a pre-registered test of aerosol encoding, climate-analog training pools and shift-aware conformal intervals

S. M. Abdulla Al Mamun
preprint en

Abstract

A site without a measurement history can obtain free day-ahead forecasts from ECMWF, NOAA GFS and the Copernicus Atmosphere Monitoring Service (CAMS), but any statistical post-processor for it must be trained on other sites. We pre-registered three hypotheses about how to do this and tested them once on ten reserved cities with no earlier weather-forecast or post-processing result, after an exploratory development stage on 23 sites that were already used in a companion study. The target is daily NASA POWER global horizontal irradiance, and the post-processors are ridge regressions on day-ahead weather forecasts and lagged POWER history, trained on one year and tested on the next. One of the three primary hypotheses was confirmed, and it concerns the form of the aerosol input. Raw issue-day CAMS inputs (aerosol optical depth, surface dust and PM2.5) make transferred post-processors fail badly at some targets: RMSE skill against climatology fell to −620% for Dhaka to Riyadh in development, and to −231% at Karachi averaged over 23 source sites. A bounded, daylight-weighted aerosol transmittance removed these failures. Its transferred skill exceeded that of the raw inputs at 10 of 10 reserved cities (Holm-adjusted p = 0.003). Column aerosol optical depth alone did the same in a secondary check (9 of 10), but neither encoding beat a post-processor without aerosol inputs (4 of 10). Training on the five climatically nearest sites did not beat the five geographically nearest (4 of 10). Climate-weighted, scale-normalised conformal intervals were not better calibrated than naive ones (3 of 10). Intervals calibrated without target-site irradiance labels under-covered at the new cities (median coverage 0.85–0.88 for a nominal 0.90), most of all at Lima and Colombo, whereas intervals calibrated on one year of target data stayed close to nominal (median 0.885). Three development-stage gains did not replicate on the reserved cities.

Zenodo (CERN European Organization for Nuclear Research)
University of Manchester (GB)
Solar Radiation and Photovoltaics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon