Hierarchical multimodel ensemble probabilistic forecasts for precipitation over East Asia

Abstract

Bayesian model averaging (BMA) and ensemble model output statistics (EMOS), as two state-of-the-art approaches, were applied to improve the prediction skills of 24-h accumulated precipitation over East Asia with lead days of 1–7 days. The multimodel ensemble precipitation probabilistic forecast experiments were constructed using ensemble forecasts from multiple ensemble prediction systems, revealing that the standard BMA (s-BMA) and the standard EMOS (s-EMOS) outperformed the raw ensemble forecasts. In comparison with the raw ensembles, the improvement by the s-BMA model increases as lead days increase, while the s-EMOS model consistently enhances prediction accuracy by around 30% for all lead days. Overall, the s-EMOS model demonstrates superior performance compared with the s-BMA model, which struggles with forecasting heavy daily precipitation exceeding 25 mm. Accordingly, the hierarchical BMA (h-BMA) model is introduced in this study, designed for different precipitation classifications. Compared with the s-BMA model, the h-BMA model notably improves the probabilistic forecast skill for all precipitation thresholds throughout East Asia, particularly for heavy precipitation events. Moreover, the h-BMA model also improves the forecast reliability across various precipitation thresholds. A hierarchical EMOS (h-EMOS) model is also developed to validate the benefits of the precipitation classifications and further improves the forecast accuracy as expected. The prediction probability density functions of the hierarchical models are much sharper and more concentrated than those of the standard models. In general, the improvement in precipitation probabilistic forecast skill of the h-BMA model relative to the s-BMA model surpasses that of the h-EMOS model compared with the s-EMOS model.