A Compound Approach for Monthly Runoff Forecasting Based on Multiscale Analysis and Deep Network with Sequential Structure

Round 1

Reviewer 1 Report

Summary The authors combine a deep learning technique with more traditional methods in the development of a compound approach for forecasting monthly runoff. They test the performance of the compound approach against other methodologies in predicting the monthly runoff in the Baishan reservoir using various metrics and demonstrate improved performance in all the metrics. Conclusion I think this is a great paper. The authors have provided a thorough description of their methodology along with careful evaluation and comparison with existing techniques. That said, I think the paper needs better organization. Grammar also needs to be corrected at many places for reading comprehension. As the paper stands now, the organization, typos and grammatical errors prevent the reader from straightforward understanding. I would suggest having the paper proof-read by a native speaker of English before publication. I list a few minor changes below Revisions Line 46: difficult to complete --> difficult to model Line 46: the word chaotic has a very specific meaning in hydrology and I don’t think it’s suitable to use here. I suggest removing it Lines 68 – 85: I feel this paragraph is somewhat tangential to the paper. I suggest removing it. Line 109: proposed on --> proposed one Line 113: remanent --> remaining Line 116: experimental --> experiments Line 168: Please correct “Error! Reference source not found..” Line 182: not sure what the authors mean by “above-combined approaches” Line 190: non-adaptively --> non-adaptivity Line 203: remove “Besides” Line 237: proposed in this study --> proposed method in this study Line 254: testify --> test Line 258: Adam --> ADAM Line 301: several hypothesis --> several hypotheses Line 304: what do authors mean by “minimum indictor” Line 346: better distinction of --> better distinction from Line 380: “in the light of balancing accuracy and efficiency” --> “with improved balance between accuracy and efficiency”

Author Response

Thank you very much for your aborative comments and positive affirmation on the paper. In this response, we have followed all your suggestions to make the whole paper proof-read by a native speaker of English. All the changes mentioned in the comments have been carefully revised, which are exhibited below:

Line 45: “However, the collection of the factors mentioned above is difficult to model, and the complicated correlations among the hydrological information may bring challenges to the construction of the physically-based models [6].”

Line 110: “Furthermore, eight relevant contrastive models and the proposed one are performed on the monthly runoff data collected from the Baishan reservoir, ...”

Line 113: “The remaining parts of our study are summarized: ...”

Line 116: “Section 4 exhibits the efficiency and effectiveness of all the experiments based on comprehensive evaluation methods.”

Line 168: “The structure of a single GRU cell is depicted in Figure 1.”

Line 187: “It can be seen that such recombination strategies based on the observation of researchers possess intense subjectivity and non-adaptivity to various datasets.”

Line 202: “The flatten layer is set following the max-pooling layer to reduce the dimension of tensors, ...”

Line 234: “For this purpose, the monthly runoff series collected from April 1933 to March 2001 is employed to evaluate the performance of the proposed method in this study, ...”

Line 254: “Additionally, TVFEMD-CNNGRU is constructed based on TVFEMD and CNNGRU to test the effectiveness of the SE-based subseries recombination employed in the proposed model.”

Line 259: “…, where the optimizer ADAM is employed to optimize the basic parameters of the neural networks.”

Line 301: “…, several hypotheses and conclusions can be drawn as follows:”

Line 348: “…, where the actual values are represented in the blue histogram for better distinction from the predicted and the actual values.”

Line 383: “To implement runoff forecasting with improved balance between accuracy and efficiency, …”

Additionally, for the issues mentioned in the comments, we have reviewed our paper carefully, while the corresponding responses for these issues are appended.

• Lines 68 – 85: I feel this paragraph is somewhat tangential to the paper. I suggest removing it.

The motivation of this paragraph is to explain the reason for the application of the time-frequency decomposition in our proposed approach. The collected runoff time series possesses strong non-stationarity, randomness, and nonlinearity, which makes it difficult to achieve accurate prediction with the single AI models. Hence, a series of signal preprocessing techniques, such as EMD, CEEMDAN and TVFEMD, are developed to decomposed the raw series into several subsequences with various frequency-scales. Hence, the tendency components with weakened non-stationarity can be effectively predicted. Among the above three decomposition methods, the drawback of modal-aliasing existing in EMD and CEEMDAN may result in ambiguous decomposition results in each component, which will restrict the performance of the forecasting models to some extent. By introducing a time-varying filter into the shifting process of EMD, the newly developed TVFEMD can solve the modal-aliasing problem effectively. What's more, the comparisons among CNNGRU, EMDEMD-CNNGRU, CEEMDAN-CNNGRU and TVFEMD-CNNGRU discussed in Section 4.3 can demonstrate the above conclusion adequately. In summary, the paragraph expressed in Lines 68 – 85 is an important part for the whole Introduction, which will be retained in the revision.

• Line 182 (Now 181): not sure what the authors mean by “above-combined approaches”.

As described in the sentence before “above-combined approaches”, we introduced the traditional hybrid models based on time-frequency decomposition technologies. The positive effects of the decomposition method are affirmed, while the shortcoming of such combined model in term of computational complexity is claimed in sequence. To make this sentence unambiguous, “above-combined approaches” is replaced by “traditional decomposition-based approaches” in the revision, which is as follows:

Line 180: “Nevertheless, the time computation of the traditional decomposition-based approaches will increase significantly with the number of decomposed subseries.”

• Line 304 (Now 301): what do authors mean by “minimum indictor”

As illustrated in Table 5, the metric values in terms of RMSE, MAE and MAPE obtained by the newly developed CNNGRU are minimum. Hence, “minimum indictor” described here are equal to minimum indictor values. In the revision, we have improved the corresponding description to make it easy to understand, which are as below:

Line 303: “…, it can be observed that the newly developed CNNGRU achieves the minimum indictor values in terms of RMSE, MAE, and MAPE as 167.3551, 108.9287 and 0.7550, …”

 

Thank you very much for all your constructive comments and suggestions to improve our paper. We hope that it would be a satisfactory response. Anyhow, if there is any question about this paper, please don’t hesitate to let us know, and we will add it in the next revised manuscript.

Author Response File: Author Response.pdf

Reviewer 2 Report

This article aims to propose a compound approach for monthly runoff forecasting based on multi-scale analysis and deep sequential structure incorporating CNN into GRU. The article has an interesting topic and presents some original results. There are few points that need revision to make it ready for publication:

 

1- Title is a bit long and confusing. Moreover, using two abbreviations CNN and GRU is also another issue. It is better to avoid abbreviations in the title.

2- Page 6, the figure number should be 2  not 1. Then the numbering for the next figures need to be updated as well.

Moreover, in the same figure, ReLU function needs to be explained somewhere in the caption.

3- Page 7, Figure 4 (which will be corrected to Fig 5 later), for the time series presented in this figure it is better to show the years in the horizontal axis rather than a count on number of month. 

4- Since the study is on monthly data, and studied data is quite long, it is suggested to present the average monthly values of each month over the whole period of the data length. This additional information beside Table 1 give a better picture of the rainfall patterns in the study site.

5- Line 242: is the partitioning of the data between training and testing made following chronological order (the first 4 chunks for training and the last chunk for testing)? If yes, it is good to elaborate whether caution has been taken to assure fair distribution of extreme values in both datasets. 

6- Page 8, Figure 5 (which will be corrected to Fig 6 later): The font size on the axis of all 4 diagrams are too small to read.  

7- Table 3. performance criteria such as RMSE and MAE have unit which will be dependent on the parameter that they measure. It is better to highlight it in this table.

8- Table 5. In many of pioneer works on using AI-based techniques in time series analysis, it is highlighted that R2 alone is incapable of assessing the model performance in extreme values. Therefore, it is generally recommended to refer to Coefficient of Efficiency where errors in extreme values will be penalized more. It is suggested to add CE.

9- Further elaboration is needed on why models such as SVR and CNN fail to perform as well as the proposed compound model    

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf