Improving radar rainfall forecasting for hydrological applications

Abstract

Rainfall forecasting plays an essential role to forecast extreme precipitation events for real-time flood forecasting. Uncertainties from both radar rainfall estimations and forecasts propagate into runoff modelling and impact the ability of an event to be correctly forecasted. Weather radars provide measurements with high temporal and spatial resolutions necessary for hydrological applications; however, radar rainfall is subject to different sources of uncertainties. Short-term radar-based rainfall forecasts (known as nowcasting) are subjected to uncertainties deriving from radar rainfall estimations, uncertainties related to the nowcasting model and uncertainties related to the temporal evolution of the precipitation field. This thesis proposes new methods to quantify and account for rainfall forecast uncertainties and assesses how some of these uncertainties propagate into hydrological modelling in small urban areas and large river catchments. Uncertainties related to both radar rainfall estimations and the temporal evolution of velocity fields were studied.
The study area focused on the north of England, where data from three weather radars and more than 200 rain gauge stations were available. A radar rainfall estimation ensemble generator was implemented to model the uncertainties in radar rainfall. The radar rainfall estimation ensembles were computed based on comparing historical weather radar rainfall estimations and rain gauge measurements. The radar rainfall estimation ensembles were used to drive a radar-based forecasting model to produce ensemble rainfall forecasts. These radar estimation ensemble forecasts were compared against the forecasts produced with a stochastic ensemble generator. The results showed an improvement in the forecasting ability of the radar rainfall estimation ensembles during the first hour of the forecasting time. For flow forecasting applications, the radar rainfall estimation ensembles overperformed the stochastic ones in the first forecasted hour and could reproduce flow peaks more accurately. To assess uncertainties related to the temporal evolution of velocity fields, a new methodology to generate ensembles using rainfall advection fields from a time window that goes from 10 min up to 2 hr before the forecast’s start was proposed. The results show that the extra information provided by the rainfall velocity fields from the previous hours can improve the rainfall forecast skill up to 3h ahead. The forecasts were used to predict sewer flows in an urban area, and the results showed that these forecasts provide improvement compared to deterministic forecasts. Merging radar rainfall with rain gauge measurements was studied to improve rainfall estimations and tests to assess if sub-hourly temporal resolutions could be used to merge radar and rain gauge data. A radar – rain gauge merging method combines the spatial distribution of precipitation from weather radars with the accuracy of rain gauge measurements to produce a product with the best from both information sources. However, using radar-rain gauge merging techniques to produce rainfall forecasts is a challenge because the temporal correlation of the radar rainfall advection field is lost. A new rainfall forecasting method that merges radar and rain gauge rainfall using kriging with external drift (KED) and using advection velocity fields from original radar data was developed. The results showed that this method produces a better rainfall forecast than using KED rainfall or radar rainfall.
The methods used to account for uncertainties in radar estimations had a more substantial influence in improving the forecasting skill up to 1 hour lead time; during this period, radar estimations are the main errors sources in nowcasting. Ensembles produced by varying the rainfall velocity fields showed improved estimations compared to a stochastic ensemble generator at longer lead times.

Date of Award	11 May 2021
Original language	English
Awarding Institution	The University of Bristol
Supervisor	Miguel A Rico-Ramirez (Supervisor), Flavia De Luca (Supervisor) & Dawei Han (Supervisor)