Analysis · Finance
A Better Solar Forecast Must Work Without Knowing Tomorrow’s Sky
By AWEI · AI-compiled · Published · Analysis prepared · 1 source · doi.org
A solar forecast can lose most of its average bias and still make substantial errors. In their 2011 study, Patrick Mathiesen and Jan Kleissl reported that correcting ECMWF forecasts reduced average bias from 31.4 to 0.5 watts per square metre. Yet root mean square error, which gives greater weight to large misses, remained 106.2 watts per square metre.
The researchers evaluated forecasts of sunlight reaching a horizontal surface against measurements from seven SURFRAD stations across the continental United States. They compared three weather models, NAM, GFS and ECMWF, alongside simple forecasts assuming clear skies or persistence of recent conditions. The model datasets covered different periods during 2009–2011, each lasting no more than a year.
The correction method had a crucial constraint: it could use only information available before the outcome. Observed cloudiness helped diagnose errors afterward, but could not guide an operational forecast. Instead, corrections used the sun’s position and forecast sunlight relative to a clear-sky reference.
At each station, the researchers fitted corrections using the preceding 56 days of errors, keeping training data separate from evaluation data. This statistical adjustment reduced average bias and overall root mean square error for all three models. It identified repeatable error patterns; it did not establish which physical processes caused them.
The gains were uneven. Under the clearest measured conditions, corrections increased root mean square error, altering some forecasts that were already accurate. Seasonal shifts could also outpace the training window. Early morning and evening were excluded from corrected forecasts, and the available model output could not resolve fluctuations shorter than an hour.
For evaluating forecast products, the useful lesson is to compare against a corrected baseline and inspect errors by condition. A small average bias alone cannot establish reliability. This study supplies a forecasting benchmark, without testing commercial returns or proving that a more expensive forecast delivers business value.
Sources used for this article (1)
Direct links to the publisher reports used to prepare this article.
- Source 1
- Evaluation of numerical weather prediction for intra-day solar forecasting in the continental United States Patrick Mathiesen, Jan Kleissl
