Solar energy prediction: An international contest to initiate interdisciplinary research on compelling meteorological problems

The AMS Committee on Artificial Intelligence and Its Applications to Environmental Science has sponsored a contest to determine which approach produces the best total daily solar energy forecast. The forecast data used in this study came from the second-generation NCEP Global Ensemble Forecast System (GEFS) reforecast dataset. For this contest, a small spatial subset of the 11-member ensemble data were extracted over Oklahoma and surrounding regions, consisting of forecasts at the +12-, +15-, +18-, +21-, and +24-h lead times. To be coincident with the observational data, the reforecast data were extracted only back to 1994. The forecast variables saved were mean sea level pressure, skin and 2-m temperature, 2-m specific humidity, daily maximum and minimum 2-m temperature, total precipitation in the last 3 h, total column precipitable water, total column integrated con?densate, total cloud cover, downward and upward short- and long-wave radiation flux at the surface, and upward long-wave radiation flux at the top of the atmosphere. The data were split into training, public testing, and private testing sets. Mean absolute error (MAE) over all stations and days was chosen as the evaluation metric because it does not penalize extreme forecasts as greatly as root mean squared error. In addition to the contest data participants. The top contestant methods exhibited similar monthly error characteristics. The con?test results showcased GBRT, which has not been used extensively in the atmospheric science community to this point. Optimized Gradient Boosted Regression Trees (GBRTs) have been shown to provide superior performance on this dataset compared to random forests, linear regressions, and neural networks, which were all used by other contestants.