A Temporal Fusion Transformer Model to Forecast Overflow from Sewer Manholes during Pluvial Flash Flood Events

Round 1

Reviewer 1 Report

A deep learning transformer-based model was developed to predict the upcoming overflow behavior at the manhole level based on measurements and forecasts of precipitation, as well as measurements in the sewer network. In addition, sensitivity analyses were carried out on the influence of the resolution of the measurement network used and the measurement signal taken into account. In general, the topic is interesting, and the suggested model is new, but not original. The manuscript needs a moderate revision to deserve acceptance.

- In the abstract: add full expressions for LSTM or a DA-RNN.

-Line 47: replace "historical event simulations." with "historical events simulation."

- Line 77-78: remove repetitive sentences.

Line 81: Temporal Fusion Transformer, especially in the field of time series analysis. Your model is not a Time series model. revise this sentence.

- Line 93: The abbreviation for machine learning model (ML) should be given at first appearance.

- Manhole Spilling, Manhole Overflow, Manhole Hydrograph! confusing! use a single term in the entire text.

- Add the coordinate system to Fig. 3. In addition, check the red line legend.

- Line 190: what do you mean by "after design rainfall events and 153 after natural rainfall events" clarify.

- Figure 4a. revise x axes values.

-Remove Eq. 1. (known issue).

- Compare different model's forecasted manhole hydrographs in a single graph.

- The introduction and discussion sections could be strengthened using the following studies:

1.       Oborie, E. ., & Rowland, E. D. . (2023). Flood influence using GIS and remote sensing based morphometric parameters: A case study in Niger delta region. Journal of Asian Scientific Research, 13(1), 1–15. https://doi.org/10.55493/5003.v13i1.4719

2.       Samadi, M., Sarkardeh, H. and Jabbari, E., 2021. Prediction of the dynamic pressure distribution in hydraulic structures using soft computing methods. Soft Computing, 25, pp.3873-3888. https://doi.org/10.1007/s00500-020-05413-6

3.       Samadi, M., Sarkardeh, H. and Jabbari, E., 2020. Explicit data-driven models for the prediction of pressure fluctuations occur during turbulent flows on sloping channels. Stochastic Environmental Research and Risk Assessment, 34, pp.691-707. https://doi.org/10.1007/s00477-020-01794-0

Minor editing of English language required.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

I believe this paper is a good in terms of comparing the different ML models. Here are my review comments:

• Sensors were used but authors have to clarify what type of data were ontained and how they were obtained.
  
• The abbreviation have to be clarified in the manuscript. These include acronyms such as MLP(line 139), VSN block (line 397). Please check the manuscript throughly.
  
• Author mentioned "chapter 2.2(line 397).  I am assuming this is section 2.2 in the manuscript. Please correct this.
• Initially, authors mentioned that there was a sensor network which consist of sensors from 20 locations. Can you please explain why the sensors were not used for the HD model calibration?
• Figure 8 made comparion between the proposed TFT model and the HD simulation. However, the HD simulation was used as a surrogate to train the model in hopes of forecasting oevrflow. Comparison should be made to the field data instead to serve its practical purpose. Please revise this.

only one minor revision is needed. Line 361 "no sensors" --> no sensor

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

- The limited investigated lead time (i.e., just one hour) may be an issue for the present study. This theme should be largely discussed with respect to potential applications of the proposed approach in real-time aimed to the issue of warnings. Lead times greater than one hour characterize several real-time flood forecasting and decision support systems. A minimum warning lead time can be established that reflects the capacity for receiving timely data and forecasts (in the operational practice, at least one hour needs to gather the observed data, to run the forecasting models and to disseminate the information about the forecast) and the time to implement necessary response actions. This lower limit is particularly relevant in urban areas.

- Lines: 46-48: this statement should be supported by references and briefly recalling the motivations for the unsuitability of models for real-time applications

- Lines 198-199: it is not clear why a period of 2 hours is cited here, given that the investigated lead time is limited to 1 hour in the following analyses . Please solve this apparent inconsistency.

 - Line 372: the underlined results appears just for 3 cases. Is there a typing problem?

- Lines 430-432: the terms “longer” and “short” are relative to the phenomenon  that occur within just the investigated  1-h time window? Please clarify in the text.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors revised their manuscript, addressing only my minor comments as listed in their revision note. However, they overlooked addressing my major concern regarding the originality of the suggested model. Additionally, some of their responses to minor points are unconvincing.

Unfortunately, the manuscript at the present form is just a simple utilization of  TFTs for Forecasting Overflow. There is no discernible innovation or advancement in this approach. Also, the comparison results are not convincing and sensible. It is obviously clear that transformers generally outperform CNNs or RNNs in such tasks. As I suggested in comment 12, to discuss TFT, it is required to compare your results with other versions of transformers. A more meaningful comparison would involve benchmarking against alternative transformer architectures rather than conventional neural networks.

On the other hand, while the paper has mentioned the advantages of TFTs its drawbacks are neglected in this study. TFT is known for its high computational requirements compared to simpler models. Training and inference with TFT may require significant computational resources, especially for the paper’s large datasets. The internal mechanisms of the model and diagnosing errors should be clarified. The process of tuning the hyperparameters of TFT, such as learning rate, number of layers, and attention mechanisms require extensive explanations. While TFT incorporates mechanisms to capture temporal dependencies in time series data, it may still have challenge with capturing complex or non-linear temporal patterns effectively. It should be mentioned that how the proposed structure addresses this challenge.

Returning to the mentioned minor comments:

in comment 7: red box is not countur.

in comment 8: the revised text is still vauge.

in comment 11: for some points, the prediction hydrographs for the best model must be shown vs benchmarks.

Minor correction is required.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 3

Reviewer 1 Report

- In response 1, the authors agreed that their aim was only to utilize TFT. However, in the abstract, their claim "This study presents a Temporal Fusion Transformer (TFT) for predicting overflow from 11 sewer manholes during heavy rainfall events. This study presents a Temporal Fusion Transformer (TFT) for predicting overflow from 11 sewer manholes during heavy rainfall events. The developed TFT ..." did not clearly show this aim. To avoid misleading the reader, these sentences should be revised. For instance:

"This study employs a Temporal Fusion Transformer (TFT) for predicting overflow from sewer manholes during heavy rainfall events. T. The utilized TFT ..."

- Similarly, revise line 70 and line 83.

The claim "until now no study used a transformer architecture" is also incorrect. Many studies used transformers in hydrology and even streamflow prediction.

- In line 524, the authors claimed "TFT is suitable for real-time operation"; however, they also reported that "the training on a GPU (NVIDIA RTX 2080 Ti) for 100 epochs lasted 18 hours for the TFT." To me, this time passes real-time prediction time.

- High similarity with reference 30 must be reduced. Special PERMISSION may be needed to use figure 3 and figure 4.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf