ESEL Paper Review_20131209
By Hong Guo
Phone: (+82) (0)10 82276568
1, Title and Author
Title: A comparative study of artificial neural networks and support vector machines for predicting groundwater levels in a coastal aquifer
Journal: Journal of Hydrology
Heesung Yoona, Seong-Chun Junb, Yunjung Hyuna , Gwang-Ok Baeb, Kang-Kun Leea,*,
a School of Earth and Environmental Sciences, Seoul National University , Seoul 151-747, Republic of Korea
b GeoGreen21 Co., Ltd., EnC Venture Dream Tower 2nd 901, Seoul 152-719, Republic of Korea
2. Summary of Paper
? Two nonlinear time-series models for predicting ground water level (GWL) fluctuations using artificial neural networks(ANNs) and support vector machines (SVM) were applied to GWL prediction of two wells at a coastal aquifer in Korea.
? The results of the model performance show that root mean squared error (RMSE) values of ANN models are lower than those of SVM in model training and testing stages. However, the overall model performance criteria of the SVM are similar to or even better than those of the ANN in model prediction stage.
? The present study focused on examining the relative importance of input variables. Therefore, the lag time of GWL was set equal to other input variables. For the ANN, the input combinations with related lag times define the nodes in the input layer, whereas they define the components of the input vectors for the SVM. In this study, multiple lead-time predictions of 1, 2, 4, 6 and 8 time steps were also performed for each input structure. These input structures and lead times constituted a total of 25 cases of a model design.
? In the model ?building process, root mean squared error was used for the model calibration through the training and testing stages
? The volume of model calibration data for well BH5 is about 2.6 times as much as for well OB2. However, the correlation coefficient between input variables and GWL for well BH5 is lower than well OB2. This implies that the time-series data of well BH5 can have a higher nonlinearity than those of well OB2 and include more noisy data that hinder the
model training process. The result of model performances for well BH5 indicates that the SVM model is more likely to catch the nonlinear relationship for the given data and to filter out the noise than the ANN model in this case.
? Although the minimum calibration values for the two models are close to each other, the calibration values of the ANN model include higher and more widely distributed values than those of the SVM model. On the other hand, the calibration values of the SVM models are close to unity and narrowly distributed
? For the validation values, the discrepancy of the two models is higher than that for the calibration values. These results indicate that the generalization ability of the ANN model is more sensitive to the selection of the input structure and lead time than that of the SVM model.
This research would be fully studied and applied to the research of the modeling for the food supernatant assessment
5. Contact (Mail address): email@example.com (H. Yoon)