Jingyi Chen, Dimitri Solomatine, Yi Xu
Thursday 2 july 2015
17:45 - 17:48h at Central America (level 0)
Themes: (T) Water resources and hydro informatics (WRHI), (ST) Catchment hydrology, Poster pitches
Parallel session: Poster pitches: 13H. WRHI - Catchment
Computer science and better understanding of physics make rainfall-runoff model more sophisticated and closer to reality. However, no matter what kind of models they are all subject to uncertainty from multiple sources. Fortunately, all these uncertainties can be propagated to the prediction and can be represented by model error. In this case study, a novel version of UNEEC method using instance-based learning (k-nearest neighbor algorithm) is developed. This method is only analogous locally and all calculation is deferred until classification. During the developing process, two thoughts are applied to select the neighbor for uncertainty prediction: one is to choose the optimized k values for neighbors and the other one is to choose the optimized distance weight function using locally weighted learning method. Say specifically, to get more flexible k value for neighbors, the creative ideal combining optimized PICP and limited distance is applied audaciously. What’s more, one smoothing weight function named Gaussian kernel with additional parameter is determined to computer the weight of each potential neighbor. After building the uncertainty predicted model, the Talagrand diagram will be developed to compare the performances of several different residual uncertainty prediction methods such as UNEEC and DUMBRAE. Combining the meteorological model and hydrological model then the residual uncertainty prediction results can be applied to get the long-time forecasting runoff in the downstream of the catchment. After applying the early warning system into the whole forecasting system the more accurate and timelier prevention action can be taken by decision makers and residents to avoid the loss of life and property.