Our forecast system is designed to blend any variety of inputs at any forecast lead time. Most of our inputs are various models ranging from HRRR at short time ranges to GFS and ECMWF at long time ranges, among others. The forecast irradiances are bias corrected at sites with observations using multivariate and corrections, adjusted against clear sky irradiance, partitioned into direct and diffuse components, projected onto tilted or sun-tracking panel geometries, and run through empirical site-specific power curves. On clear days, the forecasts are nearly perfect for sites with a good history of good data. However, all too often, the forecasts are more sunny than the verification because the underlying models busted – most of them, much more often than they should, in dry and moist climates in southern and northern US states, in short and long forecast lead times. Some cases were driven by mesoscale events, others by large scale cut-off lows that didn't move out as predicted, others by excessive boundary layer mixing in the models. Cases will be shown and some steps we have taken to improve this situation will be presented.
Since the MDA forecast system uses irradiance observations for bias correcting model forecasts and power observations for generating empirical power curves, data quality is crucial for generating the correct relationships. However, we have found flaws with many data sources widely used by solar forecasters and have found ways to correct some and flag others that appear to be unaddressed by those who own, have, or use these data. We will show some examples to illustrate how we improve this situation.