Improving Forecast Accuracy with an Auto Machine Learning Post-Correction Technique in Northern Xinjiang

Reliable meteorological forecasts of temperature and relative humidity are critically important to take necessary measures to avoid potential damage and losses. An operational meteorological forecast model based on the Weather Research and Forecast (WRF) model has been built in Xinjiang. Numerical forecasts usually have significant uncertainties and errors due to imperfections in models themselves. In this study, a straightforward automated machine learning (AutoML) approach has been developed to post-process the raw forecasts of the WRF model. The method was implemented and evaluated to post-process forecasts from 13 stations in northern Xinjiang. The post-processed temperature forecasts were significantly improved from the raw forecasts, with average RMSE values in the 13 stations decreasing from 3.24 degrees C to 2.34 degrees C by a large margin of 28%. As for relative humidity, the mean RMSE at 13 stations decreased from 19.54% to 11.54%, or it showed a percentage decrease of 41%. Meanwhile, biases were also significantly decreased, with average ME values being reduced from around 2 degrees C to similar to 0.33 degrees C for temperature and improved from -15.6% to similar to 0% for relative humidity. Moreover, forecast performance values after post-correction became much closer to each other than raw forecast performance values, improving forecast applicability at regional scales.