Spatial Bayesian model averaging to calibrate short-range weather forecast in Jakarta, Indonesia

Abstract

Bayesian Model Averaging (BMA) is a statistical post-processing method to calibrate the ensemble forecasts and create more reliable predictive interval. However, BMA does not consider spatial correlation. Geostatistical Output Perturbation (GOP) considers spatial correlation among several locations altogether. It has spatial parameters that modifies the forecast output to capture spatial information. Spatial Bayesian Model Averaging (Spatial BMA) is a method which combines BMA and GOP. This method is applied to calibrate the temperature forecast at 8 stations in Indonesia that is previously predicted by Numerical Weather Prediction (NWP). Temperature forecasts of BMA are used to obtain simulated spatially correlated error that modify temperature forecasts. Spatial BMA is able to calibrate the temperature forecast better than raw ensemble whose coverage comes closer to the standard 50%. Based on Root Mean Square Error (RMSE) criteria, Spatial BMA is able to correct forecast bias NWP with RMSE value of 1.399° lower than NWP of 2.180°.