Short-term ensemble prediction of convective cells using data-driven methods
File(s)
Author(s)
Shu, Zhou
Type
Thesis
Abstract
Very short-term forecasting (nowcasting) of convective rainfall remains a challenge due to the highly localised and intense nature of convective storms. Capturing swift changes in convective cells continues to stretch operational systems. This thesis introduces an ensemble nowcasting framework for predicting key cell properties, namely size, major axis length and reflectivity, using radar data processed by an enhanced TITAN algorithm.
An initial assessment exposes the limitations of existing field-based methods for isolated, fast growing storms and motivates the development of an analogue-based approach. A basic analogue model is first constructed and then refined by applying adaptive spatial thresholds to the process of selecting analogues. Further analysis demonstrates that accurate prediction of a cell’s track type is essential for modelling its temporal evolution.
To forecast track type, a range of machine learning classifiers are evaluated, with random forests being the most effective. Track type predictions from these classifiers are then incorpo- rated into the analogue framework so that only historical cells sharing the predicted category are retained as analogues. This combined method improves the reliability of categorical forecasts, enhances the accuracy of deterministic forecasts and yields ensembles with better calibration than those based on simple persistence or unfiltered analogue methods.
The resulting system strikes a practical balance between forecast skill, calibration and computational efficiency, making it well suited for urban surface-water flooding applications. Future work will develop automated parameter-tuning routines, integrate higher-resolution environmental predictors and reconstruct full gridded rainfall fields for hydrological modelling.
An initial assessment exposes the limitations of existing field-based methods for isolated, fast growing storms and motivates the development of an analogue-based approach. A basic analogue model is first constructed and then refined by applying adaptive spatial thresholds to the process of selecting analogues. Further analysis demonstrates that accurate prediction of a cell’s track type is essential for modelling its temporal evolution.
To forecast track type, a range of machine learning classifiers are evaluated, with random forests being the most effective. Track type predictions from these classifiers are then incorpo- rated into the analogue framework so that only historical cells sharing the predicted category are retained as analogues. This combined method improves the reliability of categorical forecasts, enhances the accuracy of deterministic forecasts and yields ensembles with better calibration than those based on simple persistence or unfiltered analogue methods.
The resulting system strikes a practical balance between forecast skill, calibration and computational efficiency, making it well suited for urban surface-water flooding applications. Future work will develop automated parameter-tuning routines, integrate higher-resolution environmental predictors and reconstruct full gridded rainfall fields for hydrological modelling.
Version
Open Access
Date Issued
2025-07-08
Date Awarded
01/02/2026
Advisor
Onof, Christian
Li-Pen, Wang
Publisher Department
Department of Civil and Environmental Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)