Parameter Estimation of Spatial Ensemble Model Output Statistics Fajar Dwi Cahyoko(a*), Sutikno (a) and Purhadi (a)
a) Department of Statistics
Faculty of Science and Data Analytics
Institut Teknologi Sepuluh Nopember
Surabaya, Indonesia
*fajardwicahyoko[at]gmail.com
Abstract
Numerical Weather Prediction is a weather forecasting method that is translated into a system of mathematical equations that are solved by numerical methods. The transformation of the basic theory of NWP into computer code still produces errors. Several statistical methods have improved accuracy and reduced bias in NWP forecasts. One of them is statistical postprocessing using Ensemble Model Output Statistics (EMOS). EMOS is a variant of multiple linear regression traditionally used for deterministic forecasting. In performance, EMOS can provide a predictive probabilistic density function and cumulative distribution function from a continuous weather variable ensemble forecast at a single observation site, without considering spatial correlation. Unlike EMOS, Geostatistical Output Perturbation (GOP) considers spatial correlations between multiple locations simultaneously. However, the GOP only applies to a single deterministic forecast. Spatial Ensemble Model Output Statistics (SEMOS) is a method that combines EMOS and GOP. SEMOS is expected to be able to make up for the shortcomings of the EMOS and GOP methods. The parameter estimator method for SEMOS is divided into several stages, namely the estimation of EMOS parameters using Maximum Likelihood Estimation with newton Raphson^s numerical iteration, followed by the estimation of spatial correlation parameters using the Weighted Least Square approach with Limited Memory BFGS (L-BFGS) iterations and finally the estimation of SEMOS model regression parameters with the same stages as EMOS parameter estimation.
