the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 27 Mar 2024
Effective Storm Surge Evacuation Planning Coupling Risk Assessment and DRL: A Case Study of Daya Bay Petrochemical Industrial Zone
Abstract. Storm surge is one of the most destructive marine disasters, characterized by abnormal and temporary rises in water levels during intense storms, leading to extreme inland flooding in the coastal area. Emergency evacuation planning, based on storm surge risk assessments, plays a crucial role in saving lives and mitigating disasters. Conventional emergency evacuation plans primarily adopt the perspective of administrators, providing evacuees with complete environmental information. However, in practical situations, evacuees often lack access to complete environmental information and need to select appropriate paths based on their limited awareness of their surroundings. This study coupled a risk assessment of storm surges with a road network to optimize evacuation routes in the Daya Bay Petrochemical Industrial Zone, a low-lying coastal region of Huizhou City, China, which is frequently affected by storm surge-driven flooding. A combination of the Deep Q-Network (DQN) model and raster environment was employed to develop real-time evacuation plans based on limited surrounding environments during storm surge events. To address the DQN model's convergence challenges, masked state space, masked action space, and tri-aspect reward mechanism were proposed, profoundly enhancing the model's convergence capabilities. The coupled ADCIRC-SWAN model and the Jelesnianski hurricane model were utilized to simulate storm surges for risk assessments under various typhoon scenarios. Additionally, potential safe shelters were identified to offer alternative evacuation options. Two distinct storm surge scenarios were employed as test environments, evaluating path plans for 1000 randomly selected starting points in each case. The results indicate that the proposed method is highly effective in devising optimal evacuation routes with minimal deviation, offering valuable guidance for evacuees during real-world storm surges.
-
Notice on discussion status
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
-
Preprint
(64321 KB)
-
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(64321 KB) - Metadata XML
- BibTeX
- EndNote
- Final revised paper
Journal article(s) based on this preprint
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2023-2280', Anonymous Referee #1, 01 May 2024
-
AC1: 'Reply on RC1', Chuanfeng Liu, 22 Jun 2024
We would like to express our sincere gratitude for your professional advice. These comments help to improve rigor of our study. We have carefully considered your comments and have thoroughly revised the manuscript according to your suggestions. This study is a deep dive into the realm of interdisciplinary research, incorporating theories and methods from various disciplines. We acknowledge that while such interdisciplinary integration offers a broad perspective and enriches insights, facing challenges in articulating it systematically is inevitable. We regret any parts where we have not clearly explained. We are committed to further refine our article to ensure its presentation is cogent. We would like to show the details as follows:
- Traditional evacuation path planning methods are generally based on static scenarios and rely on manually drawn lines from a macro perspective, which makes it difficult to achieve road-level precision, lacking timeliness and availability. The benefits of using deep reinforcement learning to address path planning issues lies in its ability to automatically select safe roads and plan appropriate evacuation routes for affected individuals based on their limited surrounding environment. This approach can adapt to dynamically changing disaster environments, providing guidance for individuals during storm surge disasters. We will enhance the introduction of the benefits of using DRL, and highlight the innovation of this work.
- Considering your feedback on the excessive number of figures and the need for more detailed information, we decided to merge Figures 1 and 4 and added more descriptive captions to all figures. We believe these improvements will more effectively convey the information and enhance the overall quality of the article.
- We had actually used tidal observation data from hydrological stations to validate the ADCIRC+SWAN model. The reason why this part of the verification was not included in the main text is primarily because the focus of this study leans more towards evacuation path planning rather than storm surge simulation. The verification of tidal levels is not as crucial. We will include the validation results for the ADCIRC+SWAN model in the appendix.
- We will move the content you identified as belonging to the methodology section from the results section to ensure the logical flow and coherence of the sections.
- The Markov Decision Process is the mathematical foundation of the DQN algorithm. To leverage the advantages of deep reinforcement learning, we transform the path planning problem into a Markov Decision Process problem, and then employed the DQN model to address it. We apologize for not making this part clear in the article and we will enhance the explanation of the relationship between the Markov Decision Process and the DQN algorithm in the text.
- The optimal path is obtained through the method of exhaustive search. We define the optimal path as the route where the agent, starting from the initial cell, achieves the maximum cumulative reward, hence the exhaustive search method can be used to find the optimal path for a given start. This aspect was not detailed in the main text, making our work appear perplexing and sloppy. We will revise this section to include the relevant information.
- We will integrate the full name "Deep Reinforcement Learning (DRL)" into the manuscript's title to more accurately reflect the content of the study.
Thank you again for your suggestions, these will help in improving the manuscript, and we will ensure there are no grammatical errors in the revised manuscript.
Citation: https://doi.org/10.5194/egusphere-2023-2280-AC1
-
AC1: 'Reply on RC1', Chuanfeng Liu, 22 Jun 2024
-
RC2: 'Comment on egusphere-2023-2280', Anonymous Referee #2, 13 May 2024
The study addresses an important problem in coastal areas prone to storm surges, such as the Daya Bay Petrochemical Industrial Zone. The coupling of risk assessment models with a road network and the utilization of DRL for path planning is well-motivated and logical. The authors have proposed several methods to address convergence challenges in their DRL model, such as masked state space, masked action space, and a tri-aspect reward mechanism which effectively enhance the model's convergence capabilities.
The analysis conducted using the coupled ADCIRC+SWAN model for simulating storm surges and evaluating risk assessments provides insights into inundation depths and extents during various typhoon scenarios. Additionally, the evaluation of optimal evacuation routes using DRL demonstrates promising results in terms of path similarity and distance to true destinations.
However, I would like to suggest some revisions that would strengthen the manuscript.
- In abstract, I suggest to add the study aim after the research gap is identified (However, in practical…), claiming the research objective that is tried to solve.
- In introduction, why ADCIRC+SWAN model employed for the study? The advantages of this model over other models need to be stated.
- In introduction, the state-of-the-art of DRL implications in emergency evacuation should be provided. If there are other studies employing DRL in emergency evacuation, the literature review should be conducted and claim the difference and originalities between the current study and previous studies. If this is the first application, please clearly claim in the introduction.
- Please clarify the specific objective of the study and its significance in addressing research gaps in storm surge evacuation in the introduction.
- Please clearly state the originalities and limitations of the current research in the introduction.
- Literature review is not very updated. Please update the literature review until recent studies in 2023 and 2024.
- Line 122-123, 11 major tidal components were included in the study, and please explain the reason of choosing these components.
- In part 4.1, please add ADCIRC+SWAN model calibration with historical data, enhancing the credibility of the flood simulation and risk assessment. After the model calibration, the typhoon scenarios and corresponding risk assessment could be further studied.
- Please include discussion of uncertainty or parameter sensitivity analysis associated with numerical models for storm surge simulations
- In conclusion, please add a section of concise summary of keys findings, and discuss how these findings can be applied in real-world decision-making process.
Â
Â
Citation: https://doi.org/10.5194/egusphere-2023-2280-RC2 -
AC2: 'Reply on RC2', Chuanfeng Liu, 22 Jun 2024
We deeply appreciate your pointing out the areas where our manuscript requires further substantiation, and your constructive suggestions for improvements. Your comments are invaluable in improving our manuscript and advancing research in the field. We are committed to addressing each of the concerns raised and have worked diligently to revise our manuscript accordingly. Taking the Daya Bay Petrochemical Industrial Zone as an example, this study integrates risk assessment with the road network and employs deep reinforcement learning algorithm for evacuation path planning. It can provide effective guidance for affected individuals based on the limited surrounding environment. This study focuses on the evacuation path planning for storm surges, and thus the description of the storm surge simulation part is rather brief, rendering the article somewhat sketchy and unclear. We will make revisions according to the suggestions provided, as follows:
- As suggested, we will revise the abstract to include the study's aim after the research gap. This addition clarifies our research objectives and enhances the abstract's coherence.
- In the introduction, we will provide a detailed rationale for employing the ADCIRC+SWAN model, highlighting its advantages over other models. The ADCIRC+SWAN model is currently a more mature model used for storm surge simulations, incorporating both wave and current factors, with more accurate simulations of storm surges. Additionally, the two models share a same grid for synchronous coupling, simplifying usage.
- There are already some studies that employ DRL methods for evacuation path planning. However, this study represents the first application of DRL to incorporate storm surge risk assessment within large-scale raster environments for this purpose. We will add details of other related works in the introduction, summarize the differences between our research and others, and clarify the innovations of our study.
- In the introduction, we will delineate the specific objectives and highlight the significance of our study in addressing research gaps in storm surge evacuation. By summarizing our works, we aim to offer readers a concise overview of the unique contributions of our research.
- We will update the Introduction section to explicitly state the original contributions and limitations of our current research, ensuring transparency and setting the stage for future work in this area.
- The initial draft of this article was actually completed in 2022, and therefore does not include related research in 2023 and 2024. We will update the literature review to include the latest advancements and discussions in the field.
- In the South China Sea, the predominant tidal components are four diurnal and four semidiurnal components. We have also included three additional components, which actually have a minor impact on the simulation results. Given that our research concentrates on storm surge risk assessment and evacuation path planning, the precision requirements for storm surge simulation are less stringent. Consequently, we employed these 11 tidal components.
- We had calibrated and validated the ADCIRC+SWAN model using tidal observation data from hydrological stations. The exclusion of this calibration part from the main text is primarily because our study is focused more on evacuation path planning than on storm surge simulation. Therefore, the verification of tidal levels was considered less critical. We will include the calibration results for the ADCIRC+SWAN model in the appendix.
- This study focuses on the risk assessment and the evacuation path planning. In setting up of the coupled model, we adopted several existing parameterization schemes and obtained marginally satisfactory simulation results. Given that our focus is not on the numerical simulation of storm surges, we did not perform a parameter sensitivity analysis of the numerical model.
- We will enhance the conclusion section by providing a concise summary of the key findings and discussing the applicability of this work in practical decision-making processes.
  We hope these responses address the concerns raised. We are grateful for the opportunity to enhance our work based on the valuable suggestions received and look forward to the potential contribution of our study to the literature in this area of research.
Citation: https://doi.org/10.5194/egusphere-2023-2280-AC2
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2023-2280', Anonymous Referee #1, 01 May 2024
This study proposes a method for Storm Surge Evacuation Planning using a coupled Deep Q-Network (DQN) model, ADCIRC, and SWAN models. However, the research appears more like a report rather than a scientific study, as many parts are not clearly explained. Therefore, a major revision is required. I would like to suggest the following improvements for the manuscript:
-
Introduction:Â authors discuss the differences between traditional methods and the DQN method. However, it is difficult for me to understand the specific benefits of using Deep Reinforcement Learning (DRL) to improve evacuation planning. It would be helpful to provide more information on how DRL can enhance the evacuation process, highlighting the innovation of this work.
-
Figures: There are too many figures, and most of them could benefit from more detailed information, and enhance the layout of the figures. For example, Fig. 1 and Fig. 4 could be merged into a single figure. Additionally, adding more descriptive captions to the figures would be beneficial.
-
Validation: It is unclear where the authors validate the ADCIRC and SWAN models using real historical disaster events. Usually there would be some QQplot for real tide gauges. Please provide information on the validation process and the results obtained.
-
Methodology: Page 23, Lines 365-375: This section seems to belong to the methodology rather than the results. Please consider moving it to the appropriate section.
-
It would be helpful to explain the relationship between the Markov Decision Process and the DQN algorithm in the text. This would ensure a smoother transition and better understanding for the readers.
-
Results: The results regarding the optimal evacuation paths are presented in a simple and unclear manner. Please provide additional information and clarification to improve the quality of the results section. It is important to avoid giving the impression of being careless or sloppy.
-
Title: Please use the full name "Deep Reinforcement Learning (DRL)" in the manuscript title to provide a more accurate representation of the study.
I hope these suggestions help in improving the manuscript. Please ensure there are no grammatical errors in the revised version.
Citation: https://doi.org/10.5194/egusphere-2023-2280-RC1 -
AC1: 'Reply on RC1', Chuanfeng Liu, 22 Jun 2024
We would like to express our sincere gratitude for your professional advice. These comments help to improve rigor of our study. We have carefully considered your comments and have thoroughly revised the manuscript according to your suggestions. This study is a deep dive into the realm of interdisciplinary research, incorporating theories and methods from various disciplines. We acknowledge that while such interdisciplinary integration offers a broad perspective and enriches insights, facing challenges in articulating it systematically is inevitable. We regret any parts where we have not clearly explained. We are committed to further refine our article to ensure its presentation is cogent. We would like to show the details as follows:
- Traditional evacuation path planning methods are generally based on static scenarios and rely on manually drawn lines from a macro perspective, which makes it difficult to achieve road-level precision, lacking timeliness and availability. The benefits of using deep reinforcement learning to address path planning issues lies in its ability to automatically select safe roads and plan appropriate evacuation routes for affected individuals based on their limited surrounding environment. This approach can adapt to dynamically changing disaster environments, providing guidance for individuals during storm surge disasters. We will enhance the introduction of the benefits of using DRL, and highlight the innovation of this work.
- Considering your feedback on the excessive number of figures and the need for more detailed information, we decided to merge Figures 1 and 4 and added more descriptive captions to all figures. We believe these improvements will more effectively convey the information and enhance the overall quality of the article.
- We had actually used tidal observation data from hydrological stations to validate the ADCIRC+SWAN model. The reason why this part of the verification was not included in the main text is primarily because the focus of this study leans more towards evacuation path planning rather than storm surge simulation. The verification of tidal levels is not as crucial. We will include the validation results for the ADCIRC+SWAN model in the appendix.
- We will move the content you identified as belonging to the methodology section from the results section to ensure the logical flow and coherence of the sections.
- The Markov Decision Process is the mathematical foundation of the DQN algorithm. To leverage the advantages of deep reinforcement learning, we transform the path planning problem into a Markov Decision Process problem, and then employed the DQN model to address it. We apologize for not making this part clear in the article and we will enhance the explanation of the relationship between the Markov Decision Process and the DQN algorithm in the text.
- The optimal path is obtained through the method of exhaustive search. We define the optimal path as the route where the agent, starting from the initial cell, achieves the maximum cumulative reward, hence the exhaustive search method can be used to find the optimal path for a given start. This aspect was not detailed in the main text, making our work appear perplexing and sloppy. We will revise this section to include the relevant information.
- We will integrate the full name "Deep Reinforcement Learning (DRL)" into the manuscript's title to more accurately reflect the content of the study.
Thank you again for your suggestions, these will help in improving the manuscript, and we will ensure there are no grammatical errors in the revised manuscript.
Citation: https://doi.org/10.5194/egusphere-2023-2280-AC1
-
-
RC2: 'Comment on egusphere-2023-2280', Anonymous Referee #2, 13 May 2024
The study addresses an important problem in coastal areas prone to storm surges, such as the Daya Bay Petrochemical Industrial Zone. The coupling of risk assessment models with a road network and the utilization of DRL for path planning is well-motivated and logical. The authors have proposed several methods to address convergence challenges in their DRL model, such as masked state space, masked action space, and a tri-aspect reward mechanism which effectively enhance the model's convergence capabilities.
The analysis conducted using the coupled ADCIRC+SWAN model for simulating storm surges and evaluating risk assessments provides insights into inundation depths and extents during various typhoon scenarios. Additionally, the evaluation of optimal evacuation routes using DRL demonstrates promising results in terms of path similarity and distance to true destinations.
However, I would like to suggest some revisions that would strengthen the manuscript.
- In abstract, I suggest to add the study aim after the research gap is identified (However, in practical…), claiming the research objective that is tried to solve.
- In introduction, why ADCIRC+SWAN model employed for the study? The advantages of this model over other models need to be stated.
- In introduction, the state-of-the-art of DRL implications in emergency evacuation should be provided. If there are other studies employing DRL in emergency evacuation, the literature review should be conducted and claim the difference and originalities between the current study and previous studies. If this is the first application, please clearly claim in the introduction.
- Please clarify the specific objective of the study and its significance in addressing research gaps in storm surge evacuation in the introduction.
- Please clearly state the originalities and limitations of the current research in the introduction.
- Literature review is not very updated. Please update the literature review until recent studies in 2023 and 2024.
- Line 122-123, 11 major tidal components were included in the study, and please explain the reason of choosing these components.
- In part 4.1, please add ADCIRC+SWAN model calibration with historical data, enhancing the credibility of the flood simulation and risk assessment. After the model calibration, the typhoon scenarios and corresponding risk assessment could be further studied.
- Please include discussion of uncertainty or parameter sensitivity analysis associated with numerical models for storm surge simulations
- In conclusion, please add a section of concise summary of keys findings, and discuss how these findings can be applied in real-world decision-making process.
Â
Â
Citation: https://doi.org/10.5194/egusphere-2023-2280-RC2 -
AC2: 'Reply on RC2', Chuanfeng Liu, 22 Jun 2024
We deeply appreciate your pointing out the areas where our manuscript requires further substantiation, and your constructive suggestions for improvements. Your comments are invaluable in improving our manuscript and advancing research in the field. We are committed to addressing each of the concerns raised and have worked diligently to revise our manuscript accordingly. Taking the Daya Bay Petrochemical Industrial Zone as an example, this study integrates risk assessment with the road network and employs deep reinforcement learning algorithm for evacuation path planning. It can provide effective guidance for affected individuals based on the limited surrounding environment. This study focuses on the evacuation path planning for storm surges, and thus the description of the storm surge simulation part is rather brief, rendering the article somewhat sketchy and unclear. We will make revisions according to the suggestions provided, as follows:
- As suggested, we will revise the abstract to include the study's aim after the research gap. This addition clarifies our research objectives and enhances the abstract's coherence.
- In the introduction, we will provide a detailed rationale for employing the ADCIRC+SWAN model, highlighting its advantages over other models. The ADCIRC+SWAN model is currently a more mature model used for storm surge simulations, incorporating both wave and current factors, with more accurate simulations of storm surges. Additionally, the two models share a same grid for synchronous coupling, simplifying usage.
- There are already some studies that employ DRL methods for evacuation path planning. However, this study represents the first application of DRL to incorporate storm surge risk assessment within large-scale raster environments for this purpose. We will add details of other related works in the introduction, summarize the differences between our research and others, and clarify the innovations of our study.
- In the introduction, we will delineate the specific objectives and highlight the significance of our study in addressing research gaps in storm surge evacuation. By summarizing our works, we aim to offer readers a concise overview of the unique contributions of our research.
- We will update the Introduction section to explicitly state the original contributions and limitations of our current research, ensuring transparency and setting the stage for future work in this area.
- The initial draft of this article was actually completed in 2022, and therefore does not include related research in 2023 and 2024. We will update the literature review to include the latest advancements and discussions in the field.
- In the South China Sea, the predominant tidal components are four diurnal and four semidiurnal components. We have also included three additional components, which actually have a minor impact on the simulation results. Given that our research concentrates on storm surge risk assessment and evacuation path planning, the precision requirements for storm surge simulation are less stringent. Consequently, we employed these 11 tidal components.
- We had calibrated and validated the ADCIRC+SWAN model using tidal observation data from hydrological stations. The exclusion of this calibration part from the main text is primarily because our study is focused more on evacuation path planning than on storm surge simulation. Therefore, the verification of tidal levels was considered less critical. We will include the calibration results for the ADCIRC+SWAN model in the appendix.
- This study focuses on the risk assessment and the evacuation path planning. In setting up of the coupled model, we adopted several existing parameterization schemes and obtained marginally satisfactory simulation results. Given that our focus is not on the numerical simulation of storm surges, we did not perform a parameter sensitivity analysis of the numerical model.
- We will enhance the conclusion section by providing a concise summary of the key findings and discussing the applicability of this work in practical decision-making processes.
  We hope these responses address the concerns raised. We are grateful for the opportunity to enhance our work based on the valuable suggestions received and look forward to the potential contribution of our study to the literature in this area of research.
Citation: https://doi.org/10.5194/egusphere-2023-2280-AC2
Peer review completion
Journal article(s) based on this preprint
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 907 | 365 | 48 | 1,320 | 71 | 84 |
- HTML: 907
- PDF: 365
- XML: 48
- Total: 1,320
- BibTeX: 71
- EndNote: 84
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Chuanfeng Liu
Yan Li
Wenjuan Li
Hao Qin
Darong Liu
Kai Zhou
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(64321 KB) - Metadata XML
This study proposes a method for Storm Surge Evacuation Planning using a coupled Deep Q-Network (DQN) model, ADCIRC, and SWAN models. However, the research appears more like a report rather than a scientific study, as many parts are not clearly explained. Therefore, a major revision is required. I would like to suggest the following improvements for the manuscript:
Introduction:Â authors discuss the differences between traditional methods and the DQN method. However, it is difficult for me to understand the specific benefits of using Deep Reinforcement Learning (DRL) to improve evacuation planning. It would be helpful to provide more information on how DRL can enhance the evacuation process, highlighting the innovation of this work.
Figures: There are too many figures, and most of them could benefit from more detailed information, and enhance the layout of the figures. For example, Fig. 1 and Fig. 4 could be merged into a single figure. Additionally, adding more descriptive captions to the figures would be beneficial.
Validation: It is unclear where the authors validate the ADCIRC and SWAN models using real historical disaster events. Usually there would be some QQplot for real tide gauges. Please provide information on the validation process and the results obtained.
Methodology: Page 23, Lines 365-375: This section seems to belong to the methodology rather than the results. Please consider moving it to the appropriate section.
It would be helpful to explain the relationship between the Markov Decision Process and the DQN algorithm in the text. This would ensure a smoother transition and better understanding for the readers.
Results: The results regarding the optimal evacuation paths are presented in a simple and unclear manner. Please provide additional information and clarification to improve the quality of the results section. It is important to avoid giving the impression of being careless or sloppy.
Title: Please use the full name "Deep Reinforcement Learning (DRL)" in the manuscript title to provide a more accurate representation of the study.
I hope these suggestions help in improving the manuscript. Please ensure there are no grammatical errors in the revised version.