Calibrating Weather Forecast using Bayesian Model Averaging and Geostatistical Output Perturbation

Abstract

Numerical Weather Prediction (NWP) has not yet been able to produce the weather forecast accurately. In order to overcome that, one approach could be taken is ensemble postprocessing. Ensemble is a combination of several methods to improve its accuracy and precision yet still possesses underdispersive nature. Bayesian Model Averaging (BMA) is intended to calibrate the ensemble prediction and create more reliable interval, though, does not consider spatial correlation. Unlike BMA, Geostatistical Output Perturbation (GOP) reckons spatial correlation among many locations altogether. Analysis applied to calibrate the temperature forecast at eight meteorological sites within Jakarta, Bogor, Tangerang and Bekasi (Jabotabek) are BMA and GOP. The ensemble members of BMA are the prediction of PLS, PCR, and Ridge. For training period over 30 days and based on some assessment indicators, BMA is better than GOP in terms of accuracy, precision, and calibration.