Exploration of Multi-Scale Reconstruction Framework in Dam Deformation Prediction
Round 1
Reviewer 1 Report
The present paper “Exploration of multi-scale reconstruction framework in dam deformation prediction” deals with the proposal of a multi-scale deformation prediction model based on Variational Modal Decomposition (VMD) signal decomposition technology.
In this reviewer opinion, the contribution is interesting, but some considerations must be made before publishing. In general, in this reviewer opinion the expression “We” should be avoided in scientific texts and changed by impersonal or passive ways of expressions. This appears along the document several occasions.
Abstract. Abstract is well presented, but in this reviewer opinion, the novelty should be clearly indicated.
Introduction. This section is well proposed, with adequate references and indicating the main objectives and procedures to get these objectives.
- Methodology. In this reviewer opinion, this section could be renamed Materials and Methods.
On line 188 there is a small typo: It is written “formular” when it should be “formulae”
In this opinion, 2.4 should be Random Forest instead of RF,
- Research and design. In this reviewer opinion, this section should be part of the Materials and Methods section. These two sections are well presented.
- Results and discussion. This section is well presented, figures, discussion and results are clear and concise.
- Conclusions. Conclusions are clear and well presented. In this reviewer opinion, conclusions could add the degree of accomplishment of the main objectives and the main novelty of the work here presented.
Author Response
Please see the attachment.
Reviewer 2 Report
In this work, the Authors advance a multi-scale deformation prediction model based on a variational modal decomposition signal decomposition technique as well as Long Short-Term Memory neural network and Random Forest algorithms, in order to analyze and re-construct, for prediction purposes, the dam deformation process based on several data monitored at a dam in China. The subject of this work is of great interest and suitable for this journal. Please see some major and minor issues that I have encountered while reading the manuscript:
1) In the text is mentioned that "dam deformation monitoring data usually reflects non-stationary and non-linear characteristics.". Please consider rephrasing this to "dam deformation monitoring models usually reflects non-stationary and non-linear characteristics." and throughout the whole text, since "stationarity" or "non-stationarity" are model attributes/characteristics, and therefore, the physical process or monitored data of "a dam deformation" cannot be (non-)stationary but rather a model (that tries to imitate the physical process of dam deformation) can be (non-)stationary. In other words, we may choose to reproduce a physical process by a stationary and a non-stationary model and then, by fitting this to real-time observations, to determine which one of the 2 models best describes our data. Please see more details if required in Koutsoyiannis and Montanary (2015).
2) In the text is written that "The collapse of the dam is usually caused by long-term deformation.".
From the 31 current related research, the common dam deformation analysis and prediction models are mainly 32 divided into three categories: Statistical model [4-6], Deterministic model [7-8], and Mixed model [9-33 11].
3) However, due to the influence of 36 water level, temperature, aging and other aspects, the deformation of the dam often shows strong 37 nonlinearity.
4) Although the regression model can establish a nonlinear relationship between variables, 38 its generalization ability is weak.
5) But the above models also have certain flaws, for example, the ARIMA model is usually applied to relatively stable time series [18] and the forecasting time is limited and it is difficult to capture the volatile time series such as dam deformation.
6) There seems to be a periodic behaviour of the dam displacement shown in Figure 6. Is there any particular reason for this? Also, please provide the units for the displacement (e.g., mm).
7) Please provide more information on how the LSTM, RF and Extreme Learning Machine (ELM) prediction algorithms work (e.g., mathematical expressions, number of parameters, physical justification of parameters, etc.), since it is not clear from the text.
8) Please provide more information on the dam that is used in the application (for example, location, height, historical facts, etc.).
9) Because LSTM is suitable for dealing with complex nonlinear problems and has good long-term memory capabilities while RF has reliable prediction results in the prediction of stationary series and the model is stable, not prone to overfitting, and the model performance is robust.
References
Koutsoyiannis, D., and A. Montanari, Negligent killing of scientific concepts: the stationarity case, Hydrological Sciences Journal, 60 (7-8), 1174–1183, doi:10.1080/02626667.2014.959959, 2015.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Round 2
Reviewer 2 Report
Thank you for the reply to the comments that have been all addressed. Please consider taking into consideration some additional suggestions that seemed to be missing from the 1st review round possibly due to some confusion during the submission process:
1) Please consider enhancing the literature review with some recent works on the dam deformation mechanism and please discuss the differences in the applied methods. For example, Yan et al. (2021) provides an interpretation and prediction of the dam deformation based also on a Long Short-Term Memory model but coupled with an attention mechanism, and therefore, this difference (and others if exist) with the Authors' work could be discussed in the Introduction.
2) In the text, it is written that: “However, due to the influence of water level, temperature, aging, and other aspects, the deformation of the dam often shows strong nonlinearity”.
As the Authors suggest, this strong volatility and linearity of the dam deformation may be caused by other factors such as temperature, water-level, streamflow, precipitation, etc. It is interesting to see that all these aforementioned factors/processes are also strongly non-linear, as quantified by the so-called Long-Term Persistence (or else LTP or Hurst coefficient after the work of Hurst, 1951). Actually, a recent global-scale analysis (Dimitriadis et al., 2021) has shown the existence of LTP in all the aforementioned factors/processes, and thus, the appearance of the strong volatility/non-linearity of the dam deformation may be justified in this manner (i.e., if all factors affecting the dam deformation are strongly non-linear and with LTP behaviour, then the deformation should also exhibit such behaviour).
Therefore, the presence of LTP in all these factors may give a physical justification to the Authors' observation that the dam deformation also exhibits strong volatility. Please consider discussing this, and linking the observed strong volatility to the LTP behaviour.
3) In the text, it is written that “the ARIMA model is usually applied to relatively stable time series”. Please also note that the ARIMA model may be able to capture non-stable processes with strong non-linearity (i.e., with the presence of LTP) but only if many parameters (i.e., many ARIMA terms) are included, which will make this model non-parsimonious and difficult to handle (for such discussion and a proposed solution for stochastic modelling of LTP processes through linear models and the explicit preservation of the autocorrelation structure and the marginal moments, please see discussion and proposed model for processes with strong volatility in Dimitriadis and Koutsoyiannis, 2018).
Please consider discussing this type of models since they are capable of capturing processes with strong volatility (through the LTP behaviour), such as temperature, streamflow, precipitation, water-level etc. (there are many such examples in the literature about these models and processes but not for the dam deformation, and thus, I believe this could be a benefit for the Authors' work to first discuss).
4) Please provide a strong grammar and syntax check throughout the whole document. For example:
- a) In the sentence “Because LSTM is suitable for dealing with complex nonlinear problems and has good long-term memory capabilities while RF has reliable prediction results in the prediction of stationary series and the model is stable, not prone to overfitting, and the model performance is robust.”, the word “Because” should be replaced to “The”.
- b) In the text, it is written that “… the forecasting time is limited and it is difficult to capture the volatile time series such as dam deformation.”. Please consider rephrasing to “… the forecasting time-window is limited and it is difficult to capture the volatile time series such as dam deformation”.
References
Dimitriadis, P., and D. Koutsoyiannis, Stochastic synthesis approximating any process dependence and distribution, Stochastic Environmental Research & Risk Assessment, 32 (6), 1493–1515, doi:10.1007/s00477-018-1540-2, 2018.
Dimitriadis, P., D. Koutsoyiannis, T. Iliopoulou, and P. Papanicolaou, A global-scale investigation of stochastic similarities in marginal distribution and dependence structure of key hydrological-cycle processes, Hydrology, 8 (2), 59, doi:10.3390/hydrology8020059, 2021.
Hurst, H.E., Long-Term Storage Capacity of Reservoirs, Trans. Am. Soc. Civ. Eng., 116, 770–799, 1951.
Yan Su, Weng Kailiang, Lin Chuan, and Zeqin Chen, Dam Deformation Interpretation and Prediction Based on a Long Short-Term Memory Model Coupled with an Attention Mechanism, J. Applied Sciences, 2076-3417, V 11, 14, 6625, doi:10.3390/app11146625, 2021.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
