CALIBRATION OF ECMWF SYSTEM - 4 ENSEMBLE SEASONAL RAINFALL PREDICTION USING BAYESIAN MODEL AVERAGING

• Main Author: Muharsyah, Robi
• Format: Theses
• Language: Indonesia
• Subjects: Geologi, hidrologi & meteorologi
• Online Access: https://digilib.itb.ac.id/gdl/view/46615
• Institution: Institut Teknologi Bandung

The European Center for Medium-Range Weather Forecasts (ECMWF) system - 4 (ECS4) is a state-of-the-art operational climate model, one output of which is ensemble seasonal rainfall prediction. However, to produce a reliable probabilistic seasonal rainfall prediction, the raw output ECS4 must be calibrated for systematic errors and biases. Bayesian Model Averaging (BMA) is a post-processing method for calibrating the raw distribution of ensemble prediction system (EPS) in the form of the predictive probability density function (PDF). In previous studies, BMA has been applied to produce probabilistic weather predictions that are limited to short and medium range. In this research, BMA implementation technique has been developed to enable calibration of seasonal rainfall prediction by using data provided by the Indonesian Agency for Meteorology, Climatology, and Geophysics (BMKG) from 26 stations over Java Island observed during 1981-2018. In this method, predictive PDFs are generated with a Gamma distribution by applying two training schemes differentiated by the use of Sequential (BMA-JTS) and Conditional (BMA-JTC) training windows. In general, the results show that BMA calibration improves the skill of ensemble seasonal rainfall prediction compared to raw model output. However, the predictive PDFs of BMA-JTC training scheme shows better verification scores than the BMA-JTS alternative. Therefore, further assessment was focused on the results of BMA-JTC calibration, which shows that improvements to probabilistic prediction quality can be found in terms of bias reduction, increased resolution as measured by Brier Score (BS), more predictions with “perfect” category by Reliability Diagram, and better overall predictions skill by Brier Skill Score (BSS). With respect to extreme events, it can be demonstrated that the calibration scheme also improves the skill and reliability of predictions for Below Normal (BN) and Above Normal (AN) seasonal rainfall, including below-normal conditions associated with El Nino. Nevertheless, there are two major issues that need to be addressed in the future development of the method, i.e., relatively poor sharpness in the resulted predictive PDFs and large bias and low skill for rainy season period.