You are on page 1of 2

Discussions and Closures

(e.g., Asefa 2009; Sharkey 1999; Shu and Burn 2004) that
Closure to “One-Day-Ahead Streamflow
the robustness and reliability of an ANN can be often signifi-
Forecasting Using Artificial Neural cantly improved by appropriately combining several ANN
Networks and a Meteorological Mesoscale models into an ANN ensemble model” (Linares-Rodriguez
Model” by Alvaro Linares-Rodriguez, et al. 2015).
Vicente Lara-Fanego, David Pozo-Vazquez, Only with that optimization procedure was the proposed model
and Joaquin Tovar-Pescador able to generate reliable and accurate one-day-ahead streamflow
forecasting. The independent dataset (last 100 days) was used to
DOI: 10.1061/(ASCE)HE.1943-5584.0001163 evaluate the ensemble model.
The discussers mentioned that, “In general, owing to the fact
Downloaded from ascelibrary.org by URI LIBRARIES on 10/20/15. Copyright ASCE. For personal use only; all rights reserved.

Alvaro Linares-Rodriguez 1; Vicente Lara-Fanego 2; that the testing subset should not have any influence on training
David Pozo-Vazquez 3; and Joaquin Tovar-Pescador 4 the ANNs, the calibration phase should only benefit from the
1
Ph.D. Candidate, Polytechnic Univ. College, Univ. of Jaén, Building training and validation subsets : : : It is obvious that using
A3, Campus Lagunillas, 23071 Jaén, Spain. E-mail: alinares@ the higher percentages of datasets in the training subset within
ujaen.es the calibration phase will increase the accuracy of forecasting.
2 Hence, it would be better to allocate the portion of the testing sub-
Researcher, Synermet Weather Solutions S.L., C/ Fuente de Don Diego,
30. 23001 Jaén, Spain. E-mail: vlara@synermet.com set to the training subset to provide a robustly trained network”
3
Associate Professor, Polytechnic Univ. College, Univ. of Jaén, Building (Linares-Rodriguez et al. 2015). The authors agree with the dis-
A3, Campus Lagunillas, 23071 Jaén, Spain. E-mail: dpozo@ujaen.es cussers that, in general, the common practice is to split the cal-
4
Professor, Polytechnic Univ. College, Univ. of Jaén, Building A3, Campus ibration set into training and validation subsets, but in this case the
Lagunillas, 23071 Jaén, Spain (corresponding author). E-mail:
optimization procedure for building an optimal ensemble model is
jtovar@ujaen.es
based explicitly on the evaluation of the ANN models over the test
The authors would like to thank the discussers for providing some subset. The 10 best ANN models selected were optimal for an
comments regarding the paper (Bozorg-Haddad et al. 2015). The ensemble model due to their diversity; all ANN models were built
following are responses/clarifications to the points raised by the with random initial weights and random (75%) training sets. As
discussers. the authors stated in the original paper: “There are two major as-
The aim of the original paper was to develop a one-day-ahead sumptions behind this procedure: (1) by varying the initial param-
streamflow forecasting model using artificial neural networks eters, the training algorithm is enabled to search other parts of the
(ANNs). The case study was conducted on the northeast Guadal- parameter space, increasing overall chances of finding a global
quivir catchment (southeast Spain) over a two-year period. Seven error minimum; (2) since measured data are affected by noise
hundred and twenty-six days were used to train and validate the and an ANN is too sensitive to calibration data quality, a cleaner
model. As the authors stated in the original paper, “a larger input choice of training dataset should better reflect the real hydrologic
datasets are preferable for training” an ANN model (Linares- state and thereby attain more reliable results.”
Rodriguez et al. 2015). In our case study, daily streamflow values The discussers mentioned that “the applied multilayer percep-
of previous years were not used because there was no flood event tron (MLP) is not capable of providing a proper extrapolation
since 2001. Training the ANN with a large dataset of low flows to predicate the out-of-range samples : : : ” For these authors, the
could produce unreliable estimates for flood events. Therefore, results of the proposed ensemble model described in the original
only streamflow values from 2009 to 2010 were used, including paper lead to a more favorable assessment of the model. The dis-
two flood events in the calibration period and another in the inde- cussers support that statement in another ad hoc case study; how-
pendent period. Although this period is sufficient to train the ANN ever, (1) they do not build an ensemble model with the proposed
(Kişi 2007), a more detailed optimization should be conducted optimization model; (2) they do build a ANN model with the
to avoid problems from the limited number of training patterns common training, validation, and testing subsets; and (3) the num-
(Linares-Rodriguez et al. 2013).” For that reason, a new optimiza- ber of input patterns is quite limited (408). For the reasons ex-
tion method was proposed. With that optimization procedure, the plained above, a simple ANN model is not able to ensure the
model was able to guarantee its generalization capability and avoid generalization capability of the model. In fact, the model proposed
overfitting problems as the authors stated. The original manuscript by the discussers achieves improvable results. For example: (1) The
described in details such a new procedure: r2 score over the independent set is 0.872 ¼ 0.76, what may be
1. The calibration dataset (626 inputs) was split into three subsets considered unsatisfactory; and (2) the discussers do not calculate
(training, validation, and testing), and randomly sampled ac- the Nash-Sutcliffe coefficient of efficiency (R2 ) nor the persistence
cording to the proportions 75%, 15%, and 10%, respectively. index (PI), especially defined for streamflow forecasting (Dawson
The ANN was trained with the training subset, the validation and Wilby 2001; Legates and McCabe 1999; Tayfur 2012). In par-
subset was used within the learning procedure to avoid over- ticular, the PI score plays an important role to evaluate the well-
training via early stopping technique, and the test subset was known problem of the timing errors often occurring when previous
used to evaluate and rank the ANN model. flow values are used.
2. Step (1) was repeated 5,000 times, and the 5,000 ANNs were The discussers mentioned that “the forecasting accuracy would
evaluated and ranked according to the value of a fitness func- deteriorate owing to two reasons: (1) using the testing subset
tion (RMSE) over the test subsets. through the calibration phase, and (2) ignoring the efficient use
3. The 10 best ANN models achieving best scores over of the available datasets.” As explained above, the testing subset
the test subsets were selected and used to build an ANN plays a key role in the optimization procedure to build an optimal
ensemble model. The authors stated: “It has been shown ensemble model. Furthermore, by repeating the procedure 5,000

© ASCE 07015022-1 J. Hydrol. Eng.

J. Hydrol. Eng., 07015022


times leaving random the initial weights and the training subset, Legates, D. R., and McCabe, G. J., Jr. (1999). “Evaluating the use of
the proposed method makes an efficient use of the available ‘goodness-of-fit’ measures in hydrologic and hydroclimatic model
datasets. validation.” Water Resour. Res., 35(1), 233–241.
Linares-Rodriguez, A., Lara-Fanego, V., Pozo-Vazquez, D., and Tovar-
Pescador, J. (2015). “One-day-ahead streamflow forecasting using
artificial neural networks and a meteorological mesoscale model.”
References J. Hydrol. Eng., 10.1061/(ASCE)HE.1943-5584.0001163, 05015001.
Linares-Rodriguez, A., Ruiz-Arias, J. A., Pozo-Vazquez, D., and Tovar-
Asefa, T. (2009). “Ensemble streamflow forecast: A GLUE-based neural Pescador, J. (2013). “An artificial neural network ensemble model
network approach.” J. Am. Water Resour. Assoc., 45(5), 1155–1163. for estimating global solar radiation from Meteosat satellite images.”
Bozorg-Haddad, O., Hosseini-Moghari, S. M., and Azarnivand, A. (2015). Energy, 61, 636–645.
“Discussion of “one-day-ahead streamflow forecasting using artificial Sharkey, A. J. C. (1999). Combining artificial neural nets: Ensemble and
neural networks and a meteorological mesoscale model”.” J. Hydrol. modular multi-net systems, Springer, New York.
Eng., in press. Shu, C., and Burn, D. H. (2004). “Artificial neural network ensembles and
Dawson, C. W., and Wilby, R. L. (2001). “Hydrological modelling using their application in pooled flood frequency analysis.” Water Resour.
artificial neural networks.” Progr. Phys. Geogr., 25(1), 80–108. Res., 40(9), W09301.
Downloaded from ascelibrary.org by URI LIBRARIES on 10/20/15. Copyright ASCE. For personal use only; all rights reserved.

Kişi, Ö. (2007). “Streamflow forecasting using different artificial neural Tayfur, G. (2012). Soft computing in water resources engineering: Artifi-
network algorithms.” J. Hydrol. Eng, 10.1061/(ASCE)1084-0699 cial neural networks, fuzzy logic and genetic algorithms, WIT Press/
(2007)12:5(532), 532–539. Computational Mechanics.

© ASCE 07015022-2 J. Hydrol. Eng.

J. Hydrol. Eng., 07015022

You might also like