ESEL Paper Review_20131225
By Hong Guo
Phone: (+82) (0)10 82276568
1, Title and Author
Title: Support vector regression for real-time flood stage forecasting
Journal: Journal of Hydrology
Pao-Shan Yua, Shien-Tsung Chen, I-Fan Chang
Department of Hydraulic and Ocean Engineering, National Cheng Kung University, No.1, University Road, Tainan 707101,Taiwan
2. Summary of Paper
? The support vector machine, a novel artificial intelligence-based method developed from statistical learning theory, is adopted herein to establish a real-time stage forecasting model.
? Two structures of models used to perform multiple-hour-ahead stage forecasts are developed. Validation results from flood events in Lan-Yang River, Taiwan, revealed that the proposed models can effectively predict the flood stage forecasts one-to-six-hours ahead. Moreover, a sensitivity analysis was conducted on the lags associated with the input variables.
? After the forecasting models had been established, both SVR model structures were used to forecast stages pertaining to six validation events at Lan-Yang Bridge.
? The forecasting results are good, and are only slightly poorer than the calibration results as identified by the values of RMSE, which demonstrate that both model structures perform equally well in one- to three-hour-ahead forecasting, whereas model structure A is superior in four- to sixhour- ahead forecasting. The results of CE bench are analogous to those of RMSE. Model structure B is only slightly better than model structure A in two- and three-hour-ahead forecasting, whereas model structure A outperforms in four to six-hour-ahead forecasting.
? The slight difference between the forecasts obtained by the two model structures demonstrates that the lags of the input variables may be insignificant or do not much affect the results. This finding inspired an investigation of the influence of the lag combination on the forecasts, which will be describe in the following section.
? The parameters calibrated by the two-step grid search method are only local optima, although our experience of parameter optimization demonstrated that the solution surface is smooth and no local optimum is likely to exist. However, the complex characteristics of nonlinear SVR mechanism do not support the easy interpretation of the performance of an SVR model. More work is required to clarify the aforementioned phenomena.
? The river stage, which is a more relevant variable than discharge in practical flood forecasting, was the variable of interest in this work. SVR is used as a method to establish the flood stage forecasting model.
This research would be fully studied and applied to the research of the modeling for the food supernatant assessment
5. Contact (Mail address): firstname.lastname@example.org (P.-S. Yu)