Probabilistic Flood Forecasting
A problem brought to the 85th European Study Group with Industry by The Environment Agency.
Problem Description
The Environment Agency provides a forecasting and warning service to people at risk from flooding. However, flood forecasts are inherently uncertain. There are differences (errors) between forecast time series of river level and subsequent observations. These differences can be relatively large, so understanding uncertainties is useful when interpreting forecasts for decision support.
We recently investigated whether information on historic flood forecast performance could be analysed to give an estimate of uncertainty around current flood forecasts in real time. This 'rear mirror' view assumes that previous error relationships continue to hold. Since it is relatively easy to compute, we saw it as a quick way to quantify errors using existing historic performance information. In the longer term, ensembles of forecast rainfall may be introduced to complement this approach.
Our question to the study group: How can the Quantile Regression approach of Weerts et al. (2011) be improved to provide representative and meaningful estimates of the uncertainty of flood forecasts?
Study Group Report
An investigation into factors that may be correlated with the uncertainty lead to the observation that there are structural biases in the model. It is possible to remove these, and thereby reduce the mean square error of the predictions, but the benefit of this is apparent in the prediction of 'normal' conditions, rather than in flood predictions. Additionally, a tweak to the linear fit in the quantile regression is suggested which is better suited to the data.