Comparison of Machine Learning Algorithms for Discharge Prediction of Multipurpose Dam

Round 1

Reviewer 1 Report

The authors constructed a machine learning model to predict the amount of discharge from Soyang River Dam using precipitation and dam inflow/discharge data from 1980 to 2020. However, there are many unreasonable presentations. Please see specific comments below:

1. The research gaps needed to be identified in a straightforward way in the end of Literature Review. In my opinion, the literature review should be added and associated with the proposed innovation.

[1] Novel Leakage detection and Water Loss Management of Urban Water Supply Network using Multiscale Neural Networks. Journal of Cleaner Production.

[2] Novel Trajectory Representation Learning Method and its Application to Trajectory-User Linking. IEEE Transactions on Instrumentation & Measurement

2. Please compare the proposed method with the new CNN method to better predict the dam discharge about Soyang River Dam.
3. Please add an overall flowchart of the proposed method.
4. Many analysis should be described in detail.
5. An abbreviation table needs to be added to the paper.

Author Response

Dear reviewer,

Thank you so much for your comments. Please see the attachment.

 

Author Response File: Author Response.pdf

Reviewer 2 Report

In this work, to predict the amount of discharge from Soyang River Dam, a machine learning model was constructed using precipitation and dam inflow/discharge data from 1980 to 2020. Six machine learning algorithms were introduced, applied and compared. As known, for the management of discharge of multipurpose dam, physical hydrological models are mostly used and good predictions results can be obtained provided that all the parameters in the model are accurately determined. However, there are many uncertainties in the calibration parameters of the hydrological models, which can dramatically affect the estimate of the dam discharge. Prediction based on machine learning technique is a promising alternative, although physical meaning of the machine learning results is often vague. This work belongs to the subject of water resource management and coincides with the topics of the journal of Water. It is well written and organized. However, the following questions deserve further examination and explanation before being accepted for publication.

 

Questions:

1. As stated in Line 106-109, for the discharge prediction model construction, precipitation data of the day (forecasted), precipitation data of one day ago, precipitation data of two days ago, the inflow of one day ago, the inflow of two days ago, the discharge of one day ago, and the discharge of two days ago from Jan. 1, 1980, to Dec. 31, 2020, were used (Table 1). The using of inflow and discharge data has been well explained in Line 96-104, but why only the data of one day and two days were used? Are they sufficiently enough? Why that of three days ago were not used? Do we have any criteria to choose this period?

 

2. Line 110, the soil moisture largely affects runoff. To reflect this effect, the historical precipitation data were used because there are no measured soil moisture data. Is this reasonable?

 

3. Data preprocessing needs to be performed for effective machine learning to improve the quality of data and to generate comprehensible information. As known, the methods for data preprocessing largely depend on the characteristics of the data. In this study, the scaling method and standardization methods including shape scaling, normalization, and standardization were used for a preprocessing step. What are the effects of performed data preprocessing on the final learning results? Without preprocessing, can we still obtain reliable solutions? What data preprocessing methods should be used and how can we choose them?

 

4. For model training tests, performance evaluation of the machine learning model is necessary and useful. In the current work, five different error measures were employed to evaluate the learning methods. However, as shown in Table 4, these measures give more or less the same trend. That is to say, any error measure will give the same conclusion for comparing the machine learning methods. As a result, can we just use one best error measure for the purpose of model evaluation? If so, which one should be used?

 

5. In Line 243-250 and Fig. 5, the outflow peaks predicted by the MLP and LSTM were on July 6^th, 2016, and the other predicted peaks overestimated the discharge. What are the possible reasons for the over estimations? How can we recognize the date July 6^th, 2016 from the figure? The authors mentioned that there was severe rainfall on July 5^th, which caused the possible first peak. What are the reasons for other peaks shown in Fig. 5?

 

6. As shown in Table 4 and Fig. 5, the LSTM method indeed gives the best prediction among the machine learning algorithms. What are the possible reasons? For other dams, can we still believe the LSTM method? That is, how can we choose a proper machine learning method for general applications? In addition, the R value in LSTM is still less than 0.9 (10% difference from the observed data). What can be done to improve the accuracy of the prediction, say make R larger than 0.95?

 

7. Although the manuscript is well prepared, there are still some minor errors or typos need careful reading in the revision, some of which are listed below

• Line 20, “Mean Absolute Error(MAE)” = Mean Absolute Error (MAE)
• Line 76, there is a reference citation error.
• There are many inconstancy in the style of the references.

Comments for author File: Comments.pdf

Author Response

Dear reviewer,

Thank you so much for your comments. Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors have modified the paper in detail. And the paper can be accepted.

Reviewer 2 Report

All my questions and comments have been well proceeded. Thanks.