Bus Dynamic Travel Time Prediction: Using a Deep Feature Extraction Framework Based on RNN and DNN

Round 1

Reviewer 1 Report

The paper is interesting and the methodology novel and well elaborated. My main concern is that it is only applied to two bus lines, one in each of the two cities under study. As suggested by the authos, in order to test the validity of the methodology this may be applied to a larger set of lines with different topologies, demand, characteristics, itineraries, schedules...

The use of continuous and discrete variables is well handled within their model. However, the authors use the speed of taxis as a proxy of the speed of traffic and the speed of buses is used as a proxy of the transit speed at the bus lane. If these assupmtions could be supported by a small analysis, especially in the first case, it would be of important added value to the paper. In addition, have the authors checked that taxis do not use the bus lane?

Minor comments:

• In the last column of table 1 it is presented the extension of the test size in terms of how many cities. Another metric  more objective sohuld be used, for example the number of bus lines, or the zie of the city, or the total of bus line kilometers studied.
• The authors use the term dynamic since the travel times present high variability during the day, but some times it is used together with the term real-time, which would be useless in the static case.
• In figure 6 there are 2 boxe swith historical data for transit time and dwell time, but in table 2 it is stated that average bus dwell time and travel time within 30 minutes are used, which is somehow real-time.
• Figure 10 should be improved, maybe using a secondary axes for the error time.

Author Response

Note: All changes and revisions made in the revised manuscript are highlighted for convenience.

 

Reviewer: 1

Comments:

1. The paper is interesting and the methodology novel and well elaborated. My main concern is that it is only applied to two bus lines, one in each of the two cities under study. As suggested by the authors, in order to test the validity of the methodology this may be applied to a larger set of lines with different topologies, demand, characteristics, itineraries, schedules...

 

Response: We appreciate your view much. Indeed, due to the limited data currently, we chose two main and representative roads in these two cities (Guangzhou and Shenzhen) testify our approach. The results show that the approach developed is able to apply for other route performances. Furthermore, as data collection conditions allowed to the public, we will conduct further in-depth tests in accordance with different topologies, requirements, features, routes, schedules, and so on.

 

2. The use of continuous and discrete variables is well handled within their model. However, the authors use the speed of taxis as a proxy of the speed of traffic and the speed of buses is used as a proxy of the transit speed at the bus lane. If these assumptions could be supported by a small analysis, especially in the first case, it would be of important added value to the paper. In addition, have the authors checked that taxis do not use the bus lane?

  

Response: Thanks for your comments. Following them, we added the explanation (highlighted) on L305-312.

 

 

Minor comments:

3. In the last column of table 1 it is presented the extension of the test size in terms of how many cities. Another metric more objective should be used, for example the number of bus lines, or the size of the city, or the total of bus line kilometers studied.

 

Response: Following your comment, this article supplements the size of test cities (highlighted) in Table 1.

 

4. In figure 6 there are 2 boxes with historical data for transit time and dwell time, but in table 2 it is stated that average bus dwell time and travel time within 30 minutes are used, which is somehow real-time.

                                                                                                                  

Response: Good spotting. Thanks. Historical data are used/analyzed as input of extraction of continuous features. Although the time-window given as 30 min, the fundamental data can be defined as historical data. Certainly, the existed studies have different definitions of real-time data. Nikolas Julio (2016) defined the dynamic travel time prediction as 10 minutes when studying the use of traffic shock waves and machine learning algorithms to predict bus speed in real time. Qichongb (2013) Predicts bus real-time travel time basing on both GPS and RFID data based on the assumption that the traffic flow keeps the same level in an interval of 30 min although he collects GPS data every 30 seconds. Hans (2015) forecasts real-time bus route state using particle filter and mesoscopic modeling with four loop detectors installed along the same corridor. The data collected provides access to volume and occupancy information collected approximately every minute. In order to predict the dynamic bus travel time, this paper adds the real-time GPS speed data of the bus every 20 seconds to the feature for dynamic bus travel time prediction, which effectively improves the prediction accuracy.

 

5. Figure 10 should be improved, maybe using a secondary axes for the error time.

 

Response: Thanks. You are right. Figure 10 was changed already and also highlighted.

 

 

Thanks for your reviewing effort !

 

Reviewer 2 Report

This paper deals with an interesting topic in Transportation Engineering and Intelligent Transportation System research and applications, that is travel time prediction. The paper proposes a real-time and dynamic travel time prediction model driven by big traffic data. These data relate to spatio-temporal features handled by using deep learning models. Thus, the authors propose a dynamic model for travel time prediction in real-time condition, based on a deep learning feature extraction framework and data fusion. The authors propose some heterogeneous features to improve the prediction, and in order to clarify the spatio-temporal features and to visualize the connections of the related characteristics in travel time evolution. The paper presents some tests based on real datasets of two bus lines  in two large urban center.

The paper is interesting for the theme it proposes and for the modeling aspects that it presents and applies. My comments are mainly about the clarity of the presentation, both for the writing style of the text and for the editing.

In general, I recommend a review, to improve readability and to make the presentation of the results obtained by applying the modeling framework proposed by the authors more clear and effective.

Below my comments by points:

1. A revision of the whole text is recommended to make the meaning of the acronyms clearer. As is well known, these must be explicitly declared when an expression that is to be abbreviated is introduced for the first time in the progression of the text. For greater clarity, the abbreviations already proposed in the abstract should be repeated in full when they appear for the first time in the full text of the paper. These situations are not always followed up in the paper.
2. Table 1 has non-optimized layout for clear reading. It is recommended to improve its readability.
3. Variables and formulas appear in the test with non-homogeneous style, font, size, line spacing (see page 7 to page 11). It is advisable to improve the readability of the paper especially in relation to editing.
4. It is advisable to improve the readability of the list in line 280 with the algorithm code, also in consideration of what is indicated in the previous point.
5. it is recommended to improve the quality of the images in figures 7 and 8.
6. it is advisable to homogenize in equations 8, 9 and 10 the reference to the number of samples, which appears both as N and as n.
7. Paragraph 5.3.1 needs a revision, as it is not clear which methods are being compared and what are the results. It is necessary to better clarify both the text and the data reported in the table with respect to the analysis scenarios and the methods investigated (control models and proposed models).
8. Paragraph 5.3.2 needs a clearer presentation, especially in the consistency between the data in figure 5 and what is reported in the text (MAE in seconds or MAPE in%?).
9. The information that figure 10 aims to provide is not fully appreciable due to the scale of the representation. The representation of the time error could be separated from that of the progressive ones (Predicted and target), in order to allow a graphing in an adequate scale to appreciate the extent of the error in seconds.
10. As for the conclusions, a revision of the text is recommended to make the comments on the results more explicit and clear. In fact, in the conclusions there are not sufficient references to the results obtained in the tests carried out. In particular, it is not clear what evidence can be drawn on the comparison between the different methods from Figure 10, as reported in lines 515-517.

 

Author Response

Note: All changes and revisions made in the revised manuscript are highlighted for convenience.

 

Reviewer: 2

1. Comments: A revision of the whole text is recommended to make the meaning of the acronyms clearer. As is well known, these must be explicitly declared when an expression that is to be abbreviated is introduced for the first time in the progression of the text. For greater clarity, the abbreviations already proposed in the abstract should be repeated in full when they appear for the first time in the full text of the paper. These situations are not always followed up in the paper.

 

Response: A good point. Thanks. Following your advice, I have revised the entire text in accordance with your request to make the meaning of the acronyms clearer. All changes are highlighted also.

 

2. Table 1 has non-optimized layout for clear reading. It is recommended to improve its readability.

 

Response: Thanks for your advice. We changed Table 1 with highlight.

 

3. Variables and formulas appear in the test with non-homogeneous style, font, size, line spacing (see page 7 to page 11). It is advisable to improve the readability of the paper especially in relation to editing.

 

Response: Thanks for your reminding. The irregular line spacing is due to the formula software but will be improved much based on the professional help from Editors. I will ask for the Great-help of composing to check and ensure it before the final edition.

 

4. It is advisable to improve the readability of the list in line 280 with the algorithm code, also in consideration of what is indicated in the previous point.

 

Response: Thanks again. The overall algorithm procedure will be attached into the paper as the supplements but is too long to assign into the context of the paper. Please allow us to insert the procedures into the followed file.

 

5. It is recommended to improve the quality of the images in figures 7 and 8.

 

Response: Thanks for your good advice, Figures 7 and 8 indeed are compressed, which is resulted from insertion into the Word-software. However, we will provide the original pictures as the individual files to ensure the quality of the images.

 

6. It is advisable to homogenize in equations 8, 9 and 10 the reference to the number of samples, which appears both as N and as n.

 

Response: Thanks for your good suggestion. We corrected them with highlight on the top of Page 14.

 

7. Paragraph 5.3.1 needs a revision, as it is not clear which methods are being compared and what are the results. It is necessary to better clarify both the text and the data reported in the table with respect to the analysis scenarios and the methods investigated (control models and proposed models).

 

Response: Following your advice, we added the explanation to make it clearer. It is also highlighted in Section 5.3.1.

 

8. Paragraph 5.3.2 needs a clearer presentation, especially in the consistency between the data in figure 5 and what is reported in the text (MAE in seconds or MAPE in%?).

 

Response: Your help is appreciated much; we improved Section 5.3.2 (highlighted).

 

9. The information that figure 10 aims to provide is not fully appreciable due to the scale of the representation. The representation of the time error could be separated from that of the progressive ones (Predicted and target), in order to allow a graphing in an adequate scale to appreciate the extent of the error in seconds.

Response: Good spotting. Thanks. Figure 10 was changed to make more sense with highlight.

 

10. As for the conclusions, a revision of the text is recommended to make the comments on the results more explicit and clear. In fact, in the conclusions there are not sufficient references to the results obtained in the tests carried out. In particular, it is not clear what evidence can be drawn on the comparison between the different methods from Figure 10, as reported in lines 515-517.

 

Response: Thanks for your professional suggestion. We changed the conclusion and summarized the four important findings to make it clearer. All changes were highlighted

 

 

Thanks for your reviewing effort !

 

 

Round 2

Reviewer 1 Report

All my comments have been tackled except the last one. As I see it in the pdf file figure 10 remains as in the original version.

 

5. Figure 10 should be improved, maybe using a secondary axes for the error time.

 

Response: Thanks. You are right. Figure 10 was changed already and also highlighted.

Author Response

I am sorry for the mistake of providing the revised edition that did not replace the new Figure 10. We all authors apologize for the fault of submitting the file when handling the submission system, seriously. 

Please check the final and revised manuscript. Thanks for your support and effect.