Development and evaluation of ensemble down-selection methods for precipitation forecast

Ensemble forecasting is a robust technique for quantifying forecast uncertainty, with larger ensemble sizes providing more accurate error estimation by reducing sampling error.However, large ensembles are computationally expensive for operational forecasting due to the substantial computer resources required.At the Canadian Meteorological Center (CMC), an ensemble of 256 members is only generated for error estimation, while a fixed subset of 20 members is utilized for medium-range forecasts twice daily.This simple selection of a subset is named as down-selection, which reduces the computational expenses.Other complex down-selection methods have also been developed to improve forecasts, but not for hourly precipitation forecasts, which present additional challenges due to their high spatiotemporal variability.This necessitates the development of downselection methods specifically tailored to hourly precipitation forecasts.The objective of this dissertation is to develop and evaluate three ensemble down-selection methods-Localized Ensemble Mosaic Assimilation (LEMA), Principal Component Analysis (PCA), and Continuous Ranked Probability Score (CRPS)-designed for the forecast of hourly precipitation.These methods share a common principle: members that better capture surface precipitation patterns relative to observations are on average closer to the "true" state, thereby improving precipitation forecasts.LEMA emphasizes the preservation of local precipitation structures, while PCA selects members based on domain-wide precipitation patterns.However, PCA's global approach used in this study reduces forecast error but compromises ensemble spread, thereby limiting error estimation capabilities.In contrast, LEMA maintains ensemble spread but sacrifices hourly precipitation accuracy.To address these trade-offs, we propose a third method, ABSTRACT viii based on the CRPS, that strikes a balance between precipitation error minimization and ensemble spread retention.Results demonstrate that the CRPS-based down-selection method outperforms LEMA and PCA, achieving lower hourly precipitation errors while preserving ensemble spread.Compared to a random subset, our approaches improve hourly precipitation forecasts by up to 6 forecast hours, notably from a reduction in incorrect precipitation forecasts (false alarms).These findings underscore the potential of three down-selection methods in different contexts to improve hourly precipitation forecasts under computational constraints.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.