Flood Risk Estimation
Flood events in the UK
are usually associated with extreme rainfall, and can last from minutes to weeks.
Efficient management and mitigation of flood risk, especially surface water flooding in urban areas,
requires accurate and reliable precipitation forecasts as inputs to flood risk models.
Houses in flat areas are particularly at risk, and meeting the shortage of houses in the south-east of the
UK requires building on these areas. However, the connection between record rainfall and flooding is
highly nonlinear, so that rainfall predictions must also say how likely rainfall is at any time –
calculating the probability of rainfall.
Probability Density Rainfall Forecasts
This research looks at the
generation of probability density forecasts of precipitation, taking a multidisciplinary approach.
All weather model predictions are associated with a particular level of uncertainty. This uncertainty
originates from errors in measurement of the current weather conditions and/or structural errors in
the model. The error in these predictions grows with the prediction horizon (the lead time), as the figure
on the right demonstrates. Each different scenario is a prediction of the mean sea level pressure, wind
speed and geopotential for the same day over the UK, but produced using information from three, four, five
and seven days in advance. The data is from the European Centre for Medium-Range
Weather Forecasts (ECMWF). As can be seen, the predictions vary substantially from a low
in one scenario to a storm in another. Clearly, single forecasts misrepresent the variability in these
predictions, and this hampers flood risk planning and estimation.
Automated Rainfall Pattern Classification
Record rainfalls caused devastating floods
in Boscastle in 2004 and Lynmouth in 1952, but the pattern of
rainfall was different. Therefore, engineers also need to know what pattern of rainfall caused the
flooding [1].
The research also addresses the problem of producing an automated system for discovering the most likely pattern
in the predicted rainfalls, using machine learning and pattern classification techniques.
Combining Numerical Weather Prediction and Historical Archive Data
The research aims at correcting the predictions from the ECMWF ensemble system [2], which produces multiple prediction scenarios. This involves data-based and statistical techniques, complemented by a state-of-the-art ensemble numerical weather prediction model. The research converts ensemble forecasts of precipitation from the ECMWF using an analysis of a newly digitised archive of spatio-temporal data. This archive is constructed from British Rainfall, an annual publication of rainfall observations, and will eventually provide daily rainfall depth at up to 6,000 locations with records ranging from 1860 to the present day. Hydro-GIS Ltd is managing the construction of this new data archive. The digital archive and the novel probability density prediction system will be made freely and publicly available at the end of the project.
Physically-Informed Data-Driven Modelling
Significant nonlinearity and nonstationarity exists in typical hydrological time series, such that time-varying, nonlinear models are strongly indicated [3]. The nonlinear Kalman filter is a predictor-corrector system iteratively producing density predictions corrected by available measurement data. No particular physical assumptions are required (the "top-down" by contrast to "bottom-up" or detailed physical modelling approach). The physically-informed data-driven modelling approach is motivated by the need for simple prediction approaches taking into account basic physical processes, without the detail of physical modelling: this is the approach taken in this research. A nonlinear Kalman with physical constraints is trained to produce predicted extreme events and meteorological outputs consistent with newly available data from British Rainfall and from the ECMWF.
For further details, please contact Max Little. Also please see the slides from the recent conference held at Oxford.
[1] W.H. Hand, N.I. Fox, C.G. Collier (2004), A study of twentieth-century extreme rainfall events in the United Kingdom with implications for forecasting. Meteorological Applications, 11(1): pp. 15-31.
[2] R. Buizza, T. Petroliagis, T. Palmer, J. Barkmeijer, M. Hamrud, A. Hollingsworth, A. Simmons, N. Wedi (1998), Impact of model resolution and ensemble size on the performance of an Ensemble Prediction System. Quarterly Journal of the Royal Meteorological Society, 124(550): pp. 1935-1960.
[3] P.C. Young (2002), Advances in real-time flood forecasting. Phil Trans R Soc Lond A, 360: pp. 1433-1450.