Development of Multi-Inflow Prediction Ensemble Model Based on Auto-Sklearn Using Combined Approach: Case Study of Soyang River Dam

Round 1

Reviewer 1 Report (Previous Reviewer 1)

Thanks, All responses are satisfying and the paper can be accepted for publication

best  

Reviewer 2 Report (Previous Reviewer 3)

I would like to thank the authors for addressing all the comments and improving the quality of the manuscript. For future works I would suggest extending the framework to larger datasets and maybe including convolutional or recurrent networks to provide a benchmark for prediction accuracy. 

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

I read this paper completely and found that it is well-written and well-organized, however, there are some drawbacks that can be considered in the revised version including:

1- It is useful to add some computations about the performance of AS and MPE approaches in the estimation of Peak flows and Low flows. 

2- Please summarized the optimized parameters of the models in the paper and compare the results between to above approaches. 

3- Where is the uncertainty analysis in your simulations? It is recommended to add some outputs about this important issue.

4- Improve the quality of figures in the revised version.

Author Response

Itemized Response to Comments (hydrology-2171241)

Development of Multi-inflow Prediction Ensemble model based on Auto-sklearn using Combined approach: Case study of Soyang River Dam

We appreciate the Editor and the reviewers’ evaluations and valuable comments of this manuscript for publication. We have adopted the comments and suggestions in our revised manuscript to improve the quality of our manuscript.

Note: Line numbers correspond to the manuscript with track changes

---------------------------------------------------------------------------------------------------------------------

Reviewer #1

Comment 1: It is useful to add some computations about the performance of AS and MPE approaches in the estimation of Peak flows and Low flows.

Response 1:

Thank you for your insightful review. We agree that there are various metrics and methods used to evaluate model performance, and the selection of specific metrics and methods may depend on the problem being addressed. In our study, we chose to evaluate the performance of two ensemble models, the conventional Auto-Sklearn (AS)-based ensemble model, and the multi-inflow prediction ensemble (MPE) model, for predicting the high and low inflow of dams using commonly used metrics (R², NSE, RMSE, and MAE). Additionally, we validated the predicted inflow data using the Flow Duration Curve (FDC) analysis and compared it with actual data, separated by season flow characteristics. We aimed to demonstrate the superiority of the AS-based ensemble model trained on high and low-inflow datasets over the AS-based ensemble model trained on the entire dataset for predicting dam inflow.

We appreciate your suggestion to perform additional analyses to reveal different aspects of the model results. However, we also recognize that the selection of appropriate additional analysis techniques and their interpretation can be time-consuming and challenging. Furthermore, too many additional analyses and misinterpretations can obscure the manuscript's purpose and confuse readers. In future research, we plan to evaluate the performance of each ensemble model in more detail while taking into consideration the trade-off between the complexity of the analysis and the clarity of the results. Once again, we thank you for your valuable feedback.

Comment 2: Please summarized the optimized parameters of the models in the manuscript and compare the results between to above approaches.

Response 2:

Thank you for your valuable feedback, we appreciate your attention to detail. As you mentioned, in our study, we have presented the hyperparameters of each model that were optimized during the training process (Table 2). We provided the optimized parameters for each model in the results and discussion section of the manuscript, and compared the performance results between the different approaches, reporting the corresponding evaluation metrics in the Results section. We believe that our study adequately addressed your request to summarize the optimized parameters of the models and compare the results between the approaches. If you have any further suggestions or comments, please do not hesitate to let us know.

Comment 3: Where is the uncertainty analysis in your simulations? It is recommended to add some outputs about this important issue.

Response 3:

Thank you for the reviewer's insightful comments. We agree that uncertainty analysis is crucial in evaluating the reliability of model predictions. In this study, we utilized AS, which is an Automated Machine Learning (AutoML) technique that automatically performs model configuration and hyperparameter optimization.

Although AS does not provide uncertainty analysis capabilities, we did evaluate the overall prediction performance of the ensemble models using multiple evaluation metrics. However, we acknowledge that uncertainty analysis is an important aspect that should be considered in future research. As you suggested, interpretability tools can be applied to the models generated by AS to evaluate uncertainty. We will consider your feedback and focus on the uncertainty analysis of AS's ensemble models in future research.

Comment 4: Improve the quality of figures in the revised version.

Response 4:

Thank you for the valuable feedback regarding the quality of the figures in our manuscript. We apologize for any inconvenience caused and have made improvements in the revised version. Firstly, to enhance visual clarity, we have increased the font size (Figure 4) and maintained consistency across all figures. Additionally, we have added the longitude and location of the study area to Figure 1 and improved the quality of the figure. Furthermore, we have provided a more detailed revision of Figure 5 to facilitate a better understanding of the development process of the MPE model, which is a crucial aspect of our research. Lastly, we have used actual dates instead of Julian days in Figure 6 to enhance comprehension of the training and testing periods. All figures have been modified to meet the journal's guidelines, with a resolution of at least 350 dpi. We hope these improvements meet your expectations and thank you again for your feedback.

Author Response File: Author Response.docx

Reviewer 2 Report

1.       How was the modeling verified?

2.       All maps need improvement .

3.       What are the objectives, importance, hypotheses and questions of this research?

4.      I suggest the author demonstrate what does the paper add to the current literature? and what new knowledge is added by this study?

5.       It is suggested to present the structure of the article at the end of the introduction. At the end of the introduction add a para including 1-Gaps in the backgrounds you try to compensate them, 2-your novelty and unique aspects 3-Hypothesis 4-Objectives.

6.       Discuss the merits and limitations of the technique applied.

7.      The presentation fails to discuss the summary and tries to some of the vague reasons, which is not the explanation. Need to compare the results with new references.

8.       The explanation for the critical analysis is not sufficient, although some of the good points have been identified.

9.       Please rewrite the conclusion with the proper explanation in the R & D and innovation process.

a.      .

10. The material and method section is too weak in the manuscript and you need to focus on it more.

11.  Please revise your conclusion part into more detail. You should enhance your contributions, hypothesis retain/reject, limitations, implications/applications, advantages/disadvantages, policies, underscore the scientific value added to your paper, and/or the applicability of your findings/results and future study in this session

Comments for author File: Comments.docx

Author Response

Itemized Response to Comments (hydrology-2171241)

Development of Multi-inflow Prediction Ensemble model based on Auto-sklearn using Combined approach: Case study of Soyang River Dam

We appreciate the Editor and the reviewers’ evaluations and valuable comments of this manuscript for publication. We have adopted the comments and suggestions in our revised manuscript to improve the quality of our manuscript.

Note: Line numbers correspond to the manuscript with track changes

---------------------------------------------------------------------------------------------------------------------

Reviewer #2

Comment 1: How was the modeling verified?

Response 1:

In this study, we verified the accuracy of our modeling by comparing the predicted dam inflow time series with observed values using both the conventional Auto-sklearn (AS)-based ensemble model and the multi-inflow prediction ensemble (MPE) model during the training and testing periods. To evaluate the prediction performance during the training period, we conducted 10-fold cross-validation. The performance of both models during the testing period was verified using observed dam inflow data, with evaluation metrics including R², NSE, RMSE, and MAE. Additionally, we evaluated the seasonal and annual dam inflow prediction performance of both AS-based ensemble models using the Flow Duration Curve (FDC). Thank you for your review and feedback.

Comment 2: All maps need improvement.

Response 2:

Thank you for the valuable feedback regarding the quality of the figures in our manuscript. We apologize for any inconvenience caused and have made improvements in the revised version. Firstly, to enhance visual clarity, we have increased the font size (Figure 4) and maintained consistency across all figures. Additionally, we have added the longitude and location of the study area to Figure 1 and improved the quality of the figure. Furthermore, we have provided a more detailed revision of Figure 5 to facilitate a better understanding of the development process of the MPE model, which is a crucial aspect of our research. Lastly, we have used actual dates instead of Julian days in Figure 6 to enhance comprehension of the training and testing periods. All figures have been modified to meet the journal's guidelines, with a resolution of at least 350 dpi. We hope these improvements meet your expectations and thank you again for your feedback.

Comment 3: What are the objectives, importance, hypotheses and questions of this research?

Response 3:

The objective of this study is to develop a multi-inflow prediction ensemble (MPE) model for dam inflow prediction using Automated Machine Learning (AutoML), which can automatically generate various machine learning models and effectively create an ensemble model by combining them for accurate predictions. The significance of this study lies in the fact that accurate prediction of dam inflow is critical for effective water resource management and dam operations, particularly in arid and semi-arid regions facing significant challenges due to the frequent occurrence of floods and droughts caused by climate change.

The hypotheses and questions of this study include whether AutoML can effectively generate an ensemble model for dam inflow prediction, whether the ensemble composition based on AutoML for dam inflow prediction can be improved by datasets assigned to flow regimes, and whether the MPE model outperforms an AS-based ensemble model developed using a conventional approach.

In summary, this study aims to develop an AutoML-based ensemble model for predicting dam inflow, considering the characteristics of the flow regime to improve prediction accuracy. It also aims to compare the performance of the MPE model with an Auto-sklearn (AS)-based ensemble model developed using a conventional approach. The results of this study are expected to provide valuable insights into developing an AutoML-based ensemble model for predicting dam inflow, contributing to sustainable water resource management and dam operation in regions with high climate variability.

Comment 4: I suggest the author demonstrate what does the manuscript add to the current literature? and what new knowledge is added by this study?

Response 4:

The manuscript presents a novel contribution to the current literature on dam inflow prediction by introducing the MPE model. The MPE model is developed using AutoML techniques such as AS, which trains independent ensemble models for high and low inflow prediction and combines their predictions. The study aims to evaluate the performance of the MPE model compared to an AS-based ensemble model developed using a conventional approach.

The manuscript provides valuable insight into developing a robust ensemble model for predicting dam inflow, which is critical for sustainable water resource management and dam operation in regions with high climate variability. By evaluating the performance of AutoML in developing ensemble models for dam inflow prediction, the manuscript adds new knowledge to the current literature. The effectiveness of the MPE model for dam inflow prediction is also demonstrated, which further contributes to the literature.

Overall, the manuscript's contribution is significant in advancing the understanding of dam inflow prediction and improving water resource management in regions with high climate variability.

Comment 5: It is suggested to present the structure of the article at the end of the introduction. At the end of the introduction add a para including 1-Gaps in the backgrounds you try to compensate them, 2-your novelty and unique aspects 3-Hypothesis 4-Objectives

Response 5:

Thank you for your suggestion. We agree that it's important to present the structure and key elements of our article in the introduction. We will revise the introduction accordingly to include a paragraph that highlights the following:

This study aims to evaluate the performance of an AutoML approach for developing an MPE model for dam inflow prediction, which has not been previously evaluated in previous studies. The novelty of this study is that the MPE model trains independent ensemble models for high and low inflow prediction based on AS and combines their predictions, taking into account the characteristics of the hydrological flow regime.

The hypothesis is that the MPE model combining the AS-based approach outperforms the conventional AS-based ensemble model by effectively capturing complex and nonlinear characteristics of dam inflow time-series data in both flood and non-flood periods.

The main objectives of this study are to develop the MPE model, evaluate its performance by comparing it to a conventional AS-based ensemble model, quantify the impact of datasets assigned to flow regimes on the ensemble composition based on AutoML and provide insight into developing an AS-based robust ensemble model for predicting dam inflow. The methodology of this study is expected to contribute to sustainable water resource management and dam operation in regions with high climate variability.

Thank you again for your helpful feedback, which will improve the clarity and effectiveness of our manuscript.

Comment 6: Discuss the merits and limitations of the technique applied.

Response 6:

The MPE model in this study combines ensemble models for high and low inflow prediction, resulting in improved accuracy for dam inflow prediction. We explored the impact of flow-regime-assigned datasets on ensemble composition and compared the performance of the MPE model to an AS-based ensemble model developed through conventional approaches.

The advantages of this methodology are twofold: first, even users without specialized knowledge can efficiently develop ensemble models for dam inflow prediction by using AS. Second, the learning of datasets assigned to flow regimes addresses issues related to data imbalance and prediction model generalization, contributing to capturing various aspects of the data. Consequently, the integration of ensemble models for high and low inflow prediction can capture diverse aspects of dam inflow data and produce more accurate dam inflow prediction data.

In this study, we compared the performance of an AS-based ensemble model developed using the conventional approach and the MPE model, which uses a combined approach of high and low inflow prediction. Both models were trained using the same time budget parameter to investigate the impact of datasets assigned to flow regimes on ensemble composition and performance. The time budget parameter in AS refers to the maximum time allowed for fitting and evaluating ML models for a specific dataset and can be adjusted based on dataset characteristics and available computational resources. However, more precise time-budget parameter adjustments will be required to generate accurate and efficient ensemble results for high and low inflow prediction.

The quality and diversity of individual models combined are crucial to the performance of the ensemble model. If the individual models are too similar or too weak, the ensemble may not be effective. However, understanding the contribution of each model to the final prediction can be difficult, which may limit the interpretability of the AS-based ensemble model. Therefore, future research should evaluate the applicability of various analysis techniques such as Shap (Shapley Additive exPlanations) and LIME (Local Interpretable Model-Agnostic Explanations to increase the interpretability of the ensemble model. (Lines 478 to 482)

Comment 7: The presentation fails to discuss the summary and tries to some of the vague reasons, which is not the explanation. Need to compare the results with new references.

Response 7:

Thank you for your valuable feedback on our manuscript. We appreciate your input and agree that the summary could benefit from a more detailed discussion and additional references. However, we would like to clarify that our discussion section provides a comprehensive analysis of the results obtained from both the conventional and MPE models. While it may be difficult to directly compare our results with previous studies due to differences in models, hyperparameters, and dataset size, we have cited several relevant studies in our introduction to support our findings.

We also highlight in our study the advantages of the MPE model over the conventional model in predicting low and high inflow, which demonstrates the importance of optimizing models for different flow regimes to improve predictive ability. Furthermore, we discuss the potential of the AS-based ensemble model compared to other standalone and combined ML models.

We believe that our manuscript cites a sufficient number of relevant studies to support our findings, but we will certainly consider your suggestion for future work. Thank you again for your feedback, and we hope that our response addresses your concerns.

Comment 8: The explanation for the critical analysis is not sufficient, although some of the good points have been identified.

Response 8:

Thank you for your comment. We appreciate your feedback and agree that interpreting the results of the AS-based ensemble model presented some challenges due to the lack of previous research on predicting non-linear time-series data, such as dam inflow, using AS-based ensembles. However, the main goal of this study was to develop the MPE model, compare its performance with a conventional AS-based ensemble model, quantify the impact of datasets assigned to flow regimes on the ensemble composition based on AutoML and provide insights into developing a robust AS-based ensemble model for predicting dam inflow. In future research, we will verify the reliability of the ensemble model by thoroughly validating the composition of the model and the predictions of its sub-models. Additionally, we will explore ways to perform critical analysis of the ensemble results using AS and various sensitivity analysis techniques.

Thank you again for your feedback, and we hope that our response addresses your concerns.

Comment 9: Please rewrite the conclusion with the proper explanation in the R & D and innovation process.

Response 9: The study aims to address the challenge of accurately predicting dam inflow for flood and non-flood periods by developing the MPE model using the AS with a combined approach to predict multi-inflow. The results of the study demonstrate that the MPE model outperforms conventional models in predicting both high and low inflow conditions, highlighting the importance of using an ensemble approach to address the imbalance between high and low inflow observations in the dataset.

Furthermore, the study provides a foundation for further research on the development and application of advanced AS-based ensemble models for inflow prediction in various fields, contributing to the advancement of data-driven decision-making and planning. The results of the study are expected to be useful not only for experts but also for non-experts without domain knowledge related to ML in ensuring the predictive performance of the AS-based ensemble model. Overall, the study's findings have important implications for water resource management and can contribute to better decision-making and planning for flood and non-flood periods.

Comment 10: The material and method section is too weak in the manuscript and you need to focus on it more.

Response 10: Thank you for your comment regarding the material and method section of our manuscript. We appreciate your feedback and will focus on improving this section to provide more clarity and detail on the development of our MPE model. As you mentioned, the key aspect of our study was the development of the MPE model using a combined approach to predict both high and low inflow conditions through the use of separate ensemble models. To achieve this, we utilized three techniques of AS: meta-learning, Bayesian optimization, and ensemble selection. We have revised the structure and wording of the relevant sentences in this section to better convey our methodology to readers.

In future studies, we will explore various sensitivity analysis techniques to improve the interpretability of the results from our ensemble model and will consider including these in the materials and methods section. Thank you again for your valuable feedback, and we will work to ensure that our materials and methods section is more robust and informative for readers.

Comment 11: Please revise your conclusion part into more detail. You should enhance your contributions, hypothesis retain/reject, limitations, implications/applications, advantages/disadvantages, policies, underscore the scientific value added to your manuscript, and/or the applicability of your findings/results and future study in this session.

Response 11:

 Thank you for your insightful comments. We have carefully revised and restructured the conclusion based on your suggestions to enhance the overall quality of the manuscript. Your attention to detail has been greatly appreciated and we thank you for contributing to the improvement of our research.

Author Response File: Author Response.docx

Reviewer 3 Report

This study provides a framework for developing an ensemble machine learning based model for predicting reservoir inflow using an automated process that would be easy to implement for users with less expertise in machine learning. The key contributions of this study are categorizing the data into low and high inflow and developing an ensemble model for each in order to improve accuracy; and using a fast and automated process for selecting the machine learning models and their hyperparameters without expertise in machine learning.

Major Revisions:

1)     Introduction must be improved. Recurrent Neural Networks and Convolutional Neural Networks especially must be reviewed as they have been proven to be the most powerful machine learning tools for dealing with sequential data such as inflow timeseries.

2)     The authors suggest that their approach is advantageous because it needs less machine learning knowledge, and it would be difficult for users to tune the hyperparameters and structure of RNNs. To support this claim, the authors must compare their results at least with a simple LSTM or CNN model. Finding the optimal structure for the mentioned networks may be difficult but performing a test using a simple structure that doesn’t require a lot of machine learning knowledge is essential for justifying the framework suggested in this study.

Minor Revisions:

1)     In line 20, it is unclear what the authors mean by “integrally predict high and low inflows”.

2)     In lines 48- 50, the authors state “The traditional approaches towards modeling studies for dam inflow prediction are classified into two major methods: physical models and data-driven machine learning (ML) models.” How about conceptual models? This paragraph must be improved. RNNs are discussed in the results section but should be discussed here. More literature review is required. For example, I would suggest looking at the paper below:

https://doi.org/10.3390/w11112201.

3)     In lines 147-150, the description of meta-learning must is very unclear and must be improved.

4)     The texts in Figure 4 should be larger.

5)     I suggest the authors improve the resolution of Figure 5.

6)     In line 211 it is stated that RMSE and MAE are a measure of error variability. How can error variability be inferred from those criteria that calculated based on average error? This should  clarified.

7)     In lines 217-220 it is clear what the authors mean by “The results showed that the MLP and GB models were used to construct ensembles with different weights. This result indicates that the same algorithm can be used for the ensembles because of various factors such as the hyperparameters and data preprocessing method”. This sentences should be rewritten and clarified. I think the authors mean in creating the ensembles, the same type of model can be used multiple times with different sets of hyperparameters and weights.

8)     In lines 235-239 it is unclear what the authors are concluding. I suggest the whole paragraph be rewritten and rephrased in a clear manner. 

9)     In table 2, the term cost is not defined anywhere. Every parameter should be defined clearly.

10) In Figure 6, it’s better to use actual dates instead of Julian day.

11) In line 255, it’s been said that the model is “overfitting” to high inflows, while no evidence is provided to support that claim, the problem seems to be underfitting, i.e., the model is not able to learn the pattens for low flows. Also, In Line 283 is has been claimed that the MPE model alleviates the overfitting problem. The problem with the “conventional” seems to be underfitting as has been shown in Figure 6-a and suggested by the training performance.

12) The sentence starting in line 285 should be part of the introduction not the results and discussion section. Same with lines 308-311 where the authors are justifying the method by discussing other papers.

13) In line 309 the author mentions “fine-tunning parameters of RNNs such as layers, etc.”. That is an unusual usage of the term “fine-tunning” which usually refers to readjusting the weights of a pre-trained network.

14) In line 289 it is not clear what is meant by “nonlinear dam inflow”.

15) In lines 296-299, the authors state that AS-based models may not always be the best option since they may be affected by calibration of hyperparameters of the algorithm among other things, while in line 308 it is said that while RNNs may have a better performance the calibration of hyperparameters is not an easy task for non-experts. These are inconsistent claims and must be clarified.

16) A valuable analysis would be to compare both model performances during training for high and low flows separately. Table 4 should include training scores as for low and high flows as well.

17) In line 337 “seasonal lower portion” is not clear. Does it mean the lower flows for each season?

18) In line 371, the authors suggest that the MPE model has decreased the “prediction uncertainty”. How do the authors measure the prediction uncertainty?

19) In line 387 the authors suggest their approach may be useful for developing “optimal ensemble model”. If RNN networks yield better results, how this approach develops the “optimal” model? I suggest that the authors avoid using the term “optimal” and focus on the advantages of their approach in the conclusion section.

Comments for author File: Comments.pdf

Author Response

Itemized Response to Comments (hydrology-2171241)

Development of Multi-inflow Prediction Ensemble model based on Auto-sklearn using Combined approach: Case study of Soyang River Dam

We appreciate the Editor and the reviewers’ evaluations and valuable comments of this manuscript for publication. We have adopted the comments and suggestions in our revised manuscript to improve the quality of our manuscript.

Note: Line numbers correspond to the manuscript with track changes

---------------------------------------------------------------------------------------------------------------------

Reviewer #3

Comment 1: Introduction must be improved. Recurrent Neural Networks and Convolutional Neural Networks especially must be reviewed as they have been proven to be the most powerful machine learning tools for dealing with sequential data such as inflow timeseries.

Response 1:

Thank you for your valuable feedback on our study. We agree that RNN and CNN models can be powerful tools for dealing with sequential data such as inflow timeseries. However, our study aimed to develop an Automated Machine Learning (AutoML)-based ensemble model for predicting dam inflow, taking into account the characteristics of the flow regime, to improve prediction accuracy. We also aimed to compare the performance of the multi-inflow prediction ensemble (MPE) model with an Auto-sklearn (AS)-based ensemble model developed using a conventional approach.

As the reviewer pointed out, RNN models can also be useful in predicting dam inflow, and we have added an example of this to the study. However, we did not provide a detailed explanation of specific models such as RNN or CNN as they are currently not available in AS. To avoid confusion and help readers understand the purpose of our study, we have restructured the introduction to provide a clearer outline of our study's objectives. We appreciate your constructive feedback and hope our revised introduction addresses your concerns.

Comment 2: The authors suggest that their approach is advantageous because it needs less machine learning knowledge, and it would be difficult for users to tune the hyperparameters and structure of RNNs. To support this claim, the authors must compare their results at least with a simple LSTM or CNN model. Finding the optimal structure for the mentioned networks may be difficult but performing a test using a simple structure that doesn’t require a lot of machine learning knowledge is essential for justifying the framework suggested in this study.

Response 2:

Thank you for the reviewer's valuable comments. Our study aimed to propose a new development approach for ensemble models that can produce better prediction results when dealing with highly complex and nonlinear time-series data. We did not compare and analyze the prediction performance of individual algorithms and Auto-sklearn (AS)-based ensemble models in this study. However, we agree with the reviewer's suggestion that a model such as RNN with a simple structure can also be used to develop a dam inflow prediction model, and it can be compared to an AS-based ensemble model.

Nevertheless, it is important to note that the ensemble model may outperform individual algorithms or it may not, as shown in the results section. Additionally, it is challenging to conclude which model is superior in prediction based solely on evaluation performance. The primary purpose of AS is to automate complex processes such as data preprocessing, hyperparameter optimization, and ensemble to produce a robust ensemble prediction model. Therefore, future research may need to compare the prediction results of the RNN model and the AS-based ensemble model for other research purposes.

Thank you again for your valuable feedback, and we hope our response addressed your concerns.

Comment 3: In line 20, it is unclear what the authors mean by “integrally predict high and low inflows”.

Response 3:

Thank you for your helpful feedback on our abstract. We understand that the wording used in line 20 may have been unclear. To provide clarification, the "integrally predict high and low inflows" phrase refers to the ability of the multi-inflow prediction ensemble (MPE) model to predict both high and low inflow conditions using an ensemble of Auto-sklearn (AS)-based models. The MPE model is trained on datasets that are separated according to high and low inflow conditions and integrate the predictive performance of each model to improve the accuracy of the dam inflow prediction. We appreciate your input and hope this explanation provides a clearer understanding of the methodology used in our study. Additionally, in the revised manuscript, we would like to inform you that the sentence "integrally predict high and low inflows" has been removed from the abstract. (Lines 16 to 29)

Comment 4: In lines 48- 50, the authors state “The traditional approaches towards modeling studies for dam inflow prediction are classified into two major methods: physical models and data-driven machine learning (ML) models.” How about conceptual models? This paragraph must be improved. RNNs are discussed in the results section but should be discussed here. More literature review is required. For example, I would suggest looking at the manuscript below:

https://doi.org/10.3390/w11112201.

Response 4:

Thank you for your valuable feedback on our manuscript. We appreciate your suggestions for improvement. In response to your comments, we have revised the paragraph on dam inflow prediction methods to include physical models, conceptual models, and data-driven machine learning (ML) models. Additionally, we have included examples of dam inflow prediction using each of these methods. Thank you again for your feedback, which has helped us to improve the clarity and quality of our manuscript. (Lines 82 to 85)

Comment 5: In lines 147-150, the description of meta-learning must is very unclear and must be improved.

Response 5:

Thank you for your feedback on our manuscript. We have reviewed the section on meta-learning and agree that it could be more clearly explained. To address this, we have rephrased the sentences and provided more details on how meta-learning was used in our study. Specifically, we utilized meta-learning to train an algorithm that can automatically select the best ML models for a given task based on the characteristics of the data. This approach helps to improve the performance of our ensemble model by ensuring that the individual models selected for the ensemble are well-suited for the task at hand. We hope this explanation provides a clearer understanding of the role of meta-learning in our study, and we appreciate your feedback on how we can improve the clarity of our manuscript. (Lines 300 to 306)

Comment 6: The texts in Figure 4 should be larger.

Response 6:

Thank you for your feedback on the font size of the text in Figure 4. We have made the necessary adjustments and increased the font size to improve the legibility of the text in figure 4.

Comment 7: I suggest the authors improve the resolution of Figure 5.

Response 7:

Thank you for your feedback on the resolution of Figure 5. I have improved the resolution of all the figures, including Figure 5, to meet the journal's guidelines, ensuring a resolution of at least 350 dpi.

Comment 8: In line 211 it is stated that RMSE and MAE are a measure of error variability. How can error variability be inferred from those criteria that calculated based on average error? This should  clarified.

Response 8:

Thank you for your comment. We appreciate your feedback on our methodology. As noted by the reviewer, the above paragraph has been corrected for clarity. To clarify, while RMSE and MAE are indeed calculated based on the average error, they can still provide some indication of error variability when used together. The RMSE is always larger than or equal to the MAE, and the greater the difference between them, the greater the variance in individual errors in the sample. While these measures alone may not provide a complete picture of error variability, they can be used in combination with other techniques to better understand and quantify error variability in a given dataset. We hope this explanation provides more clarity on our approach. (Lines 420 to 424)

 Comment 9:  In lines 217-220 it is clear what the authors mean by “The results showed that the MLP and GB models were used to construct ensembles with different weights. This result indicates that the same algorithm can be used for the ensembles because of various factors such as the hyperparameters and data preprocessing method”. This sentences should be rewritten and clarified. I think the authors mean in creating the ensembles, the same type of model can be used multiple times with different sets of hyperparameters and weights.

Response 9: Thank you for your comment. As you mentioned, I have modified the sentence as follows: "The results showed that even when the same type of models are included in the ensemble configuration, their weights can be assigned differently based on data and feature preprocessing methods, as well as hyperparameters. (Lines 437 to 439)

Comment 10: In lines 235-239 it is unclear what the authors are concluding. I suggest the whole paragraph be rewritten and rephrased in a clear manner.

Comment 10:

As per your comment, the entire paragraph has been revised. Through the revised paragraph, we emphasized the importance of separating the training dataset before constructing the ensemble and the significance of appropriately optimizing the time budget parameters that determine the learning time in meta-learning for each ensemble model. (Lines 475 to 477)

Comment 11: In table 2, the term cost is not defined anywhere. Every parameter should be defined clearly.

Response 11:

The cost in AS represents the time required for model training and validation. A lower cost indicates that AS found the optimal model in a shorter time. Additionally, Rank represents the evaluation ranking of each hyperparameter configuration in the meta-learning algorithm, and Weight indicates the importance of each model in the ensemble. As a detailed explanation of the weights is necessary for the flow of the content, we have included a brief description of the weights on lines 256 to 260 in the manuscript.

Comment 12: In Figure 6, it’s better to use actual dates instead of Julian day.

Response 12:

Thank you for your comment, as you mentioned, I have updated the Julian day in Figure 6 to the actual date.

Comment 13: In line 255, it’s been said that the model is “overfitting” to high inflows, while no evidence is provided to support that claim, the problem seems to be underfitting, i.e., the model is not able to learn the patterns for low inflows. Also, In Line 283 it has been claimed that the MPE model alleviates the overfitting problem. The problem with the “conventional” seems to be underfitting as has been shown in Figure 6-a and suggested by the training performance.

Response 13:

I have revised the relevant sentences in the manuscript to better address the reviewer's comments. Instead of claiming that the model is "overfitting" to high inflows, I have clarified that the problem may be underfitting, i.e., the conventional AS-based ensemble model is not able to learn patterns for low inflows. In addition, I have removed the claim that the MPE model alleviates the overfitting problem and acknowledged that distinguishing between overfitting and underfitting can be difficult. Overall, I have made sure to convey the necessary information while addressing the reviewer's concerns clearly and respectfully. (Lines 507 to 511)

Comment 14: The sentence starting in line 285 should be part of the introduction, not the results and discussion section. Same with lines 308-311 where the authors are justifying the method by discussing other manuscripts.

Response 14:    

Thank you for your comment, as suggested by the reviewer, the relevant citations that should have been included in the introduction instead of the results and discussion section have been moved to the introduction to better support the purpose of the study. Thank you for the suggestion.

Comment 15: In line 309 the author mentions “fine-tunning parameters of RNNs such as layers, etc.”. That is an unusual usage of the term “fine-tuning” which usually refers to readjusting the weights of a pre-trained network.

Response 15:

Thank you for your feedback. As suggested, we have removed the term "fine-tunning" from the manuscript as it is an unusual usage and may cause confusion.

Comment 16: In line 289 it is not clear what is meant by “nonlinear dam inflow”.

Response 16:

In line 289, I used the term "nonlinear dam inflow" to refer to the fact that the inflow to a dam can have complex, nonlinear characteristics that are influenced by various factors such as rainfall and watershed properties.

However, I understand that this terminology may not be immediately clear to all readers. To address this, I have revised the paragraph to better explain the concept and its relevance to the MPE model. The updated paragraph reads as follows:

This improved performance can be attributed to the MPE model's combined approach, where models are trained separately for high-inflow and low-inflow datasets, allowing for the optimization of the models using the bayesian optimization algorithm in AS. In contrast, conventional model approaches trained on the entire dataset are less likely to be optimized for each flow regime, leading to poor predictive ability and low-flow regimes. This highlights the importance of optimizing models for different flow regimes, especially low-flow regimes, to improve their predictive ability.

I hope this revised explanation is clearer and more concise, while still conveying the intended message. (Lines 537 to 539)

Comment 17: In lines 296-299, the authors state that AS-based models may not always be the best option since they may be affected by calibration of hyperparameters of the algorithm among other things, while in line 308 it is said that while RNNs may have a better performance the calibration of hyperparameters is not an easy task for non-experts. These are inconsistent claims and must be clarified.

Response 17:

In our manuscript, we initially suggested that AS-based ensemble models could have superior performance compared to other algorithms for predicting dam inflow. However, upon further investigation and analysis, we found that the performance of AS-based models does not show a significant difference compared to other standalone and combined ML models. Therefore, we now caution against making overly confident claims about the superiority of AS-based models in all cases. Instead, we suggest that the performance of an ML model for a specific dataset can be improved by the efforts of experts, rather than relying solely on meta-learning-based AS. Our revised statement reflects this caution, and we provide references to support our claims. Specifically, we reference a study that demonstrated the importance of expert input in improving the performance of ML models. We hope that this clarification addresses the reviewer's concerns and contributes to a more accurate and nuanced understanding of the performance of AS-based models.

The updated sentence reads as follows:

It is noteworthy that the prediction performance of the AS-based ensemble models for dam inflow does not show a significant difference compared to other standalone and combined ML models. The performance of an ML model for a specific dataset can also be improved by the efforts of experts, rather than relying on meta-learning-based AS. This means that AS cannot guarantee the development of the optimal prediction model for a given dataset. (Lines 552 to 557)

Comment 18: A valuable analysis would be to compare both model performances during training for high and low inflows separately. Table 4 should include training scores for low and high inflows as well.

Response 18:

Thank you for your feedback. We appreciate your suggestion to analyze and compare the performance of both models separately for high and low inflows during training, as well as to include training scores for low and high inflows in Table 4. In our study, we have already analyzed the prediction performance of the two models based on four evaluation metrics (R², NSE, RMSE, and MAE) during training and testing periods for both high and low inflow rates. Moreover, we have emphasized the importance of considering the flow duration curve (FDC) to compare and evaluate the performance of the AS-based ensemble model with other models for accurate dam inflow prediction during a flood and dry seasons. We have also followed your recommendation to add the training scores for high and low inflow rates for both models in Table 4. However, we acknowledge that valuable analysis regarding the prediction uncertainty of ensemble models for high and low inflow rates can be a potential area of future research.

Comment 19: In line 337 “seasonal lower portion” is not clear. Does it mean the lower flows for each season?

Response 19:

Thank you for your feedback on my previous response. I agree that the term "seasonal lower portion" may not be clear to all readers, and I appreciate your suggestion to use the phrase "lower flows for each season" instead. To address this, I have revised the sentence as follows:

More specifically, Figure 10 compares the observed and predicted low inflow values using the models for each season (>60% exceedance probability), showing that the MPE model had higher R² and NSE than the conventional model, reducing RMSE and MAE by 58.8–88.5% and 54.1–89.9%, respectively. (Lines 659 to 662)

Comment 20: In line 371, the authors suggest that the MPE model has decreased the “prediction uncertainty”. How do the authors measure the prediction uncertainty?

Response 20:

I appreciate your concern about the potential confusion caused by the term "prediction uncertainty". To address this, I have removed this term from the sentence and revised it to better convey the intended message. The updated sentence reads as follows:

"Additionally, the MPE model was found to capture the characteristics of each flow regime and make more accurate predictions for each condition."

I also appreciate your suggestion to consider analyzing the uncertainty of the prediction models in future research, especially given its importance in ensemble model development. We will consider this in our future work.

Comment 21: In line 387 the authors suggest their approach may be useful for developing an “optimal ensemble model”. If RNN networks yield better results, how this approach develops the “optimal” model? I suggest that the authors avoid using the term “optimal” and focus on the advantages of their approach in the conclusion section.

Response 21:

 I appreciate your suggestion to avoid using the term "optimal" and instead focus on the advantages of our approach in the conclusion section. I fully agree that our proposed ensemble approach cannot be considered the "optimal" model, and we have removed this term from the manuscript to avoid any confusion.

To clarify the contributions of our approach, we have revised the conclusion section to highlight the advantages of our proposed method in developing an ensemble model for dam inflow prediction.

Author Response File: Author Response.docx