Bottleneck Based Gridlock Prediction in an Urban Road Network Using Long Short-Term Memory

Round 1

Reviewer 1 Report

This study is for multiple predictions of link speed using LSTM. The study is necessary in ITS information. Despite this necessity, it can be seen that there are many things to successfully be addressed.

1. “Bottleneck” is not suitable in this research, I recommend using “Gridlock congestion”.
2. The test data, i.e., link speed data were generated by a microscopic traffic simulator. Training and validation should be conducted by a typical validation process (e.g., cross validation widely used in machine learning). For example, the data can be divided into the dataset. Dataset 1 can be used as the past dataset to training the neural network the model parameters, and then the dataset 2 (i.e., current data) can be used for forecasting using the trained neural network. Finally, the trained parameters should be presented.
3. Temporal evolution of signalized traffic speed shows wide and intensive variations. This is because of the repetitive interruption of traffic signal. Due to this, forecasting traffic speed especially for turning traffic flow is deeply related to the increased number of uncertainties. And then this is deeply related to the famous forecasting failure especially in the case when the time length for data aggregation. So, the temporal evolution of traffic speed should be demonstrated, and observe and predicted time-series states should be compared. In addition, the time length of data for dynamic provision of traffic flow information is 5 min.
4. To show the robustness of this approach, the method should be compared against widely used data-driven approach (i.e., KNN) and time-series analysis technique (e.g. Seasonal ARIAM and Kalman filtering), and season average technique which are widely used in ITS.
5. There is a critical problem on collecting speed data. The link speed is collected from the entry time on the upstream of a road section to the exit time on the downstream. And then the speed data includes time lag especially when traffic states are congested or jammed. So the forecasting problem is that “generate link speed (t+i) with time-series data (t-l, t-l-1, t-l-2, …., t-l-d)”, where l = time lag, and d = embedding size for time series. In addition, the size of time series is flexible in this case. But this definition of forecasting problem was not included in the manuscript. And the time-lag problem should be considered.
6. The delayed time which is required for data communications from information ends (i.e., vehicle GPS) to data centre should be included into “definition of forecasting problem.” This is fundamental for real-life application in ITSs.
7. Above time-lag problem, unavoidably occurs in the process of real-time data collection based on section-based detection approach such as vehicle GPS systems, should be successfully addressed for the real-life applications. If this problem is not successfully addressed, then the approach proposed in this study is only offline furcating. That is, the method cannot be used in the real-world information systems.
8. Observations and predictions should be compared with a scattergram which is one of clear approach to show the accuracy performances of the proposed method. In addition, mean absolute error (MAE) and mean average percentage error (MAPE) should be used as performance measures for readers. Furthermore, the RMSE used in this study does not has a reasonable comparison in itself. If the forecasting horizon is multi steps, then the comparison results for multi-time steps (e.g., t+1, t+2, …., t+n) should be presented.
9. In real-world, the penetration of a vehicle GPS system is low (i.e., less then 3~5%), and then the reliability of collected link speed dramatically decreases. Especially, in the case of the traffic jam, the collected simples also decrease. So, this practical and real problem, unavoidably occurs in the real-time data collection, should be addressed effectively and successfully.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The paper proposed LSTM method to predict intersection bottleneck and detect gridlocks based on the identified bottleneck. The topic is interesting. The authors presents their simulation based method well and discussed the results. The paper can be improved with the consideration of the following main recommendations:

1. section 3.3. the authors demonstrated the effects of of sample size with different GPS vehicle percentage. I have a few main concerns: 1) regarding to GPS vehicles, what are the main information collected by them? speed and locations? if so, how did the authors took GPS errors into the considerations? in general, GPS signal have 30 to 50 meters error. 2) the authors have 5 gridlock labels, how did you determine those labels for all the samples? 3) better explains regarding to relative low DT for DR_1 TO 4 are needed. At least, with the current status, I have no confidence on the detection rate with the model even only focus on DR_5 which is the best performers.
2. Figure 14, and 15 shows no clear trending relationship between sample size and error/computation time. The authors need further work to explain how to determine robust sample size.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This study quantifies the intersection bottleneck and gridlock with the deep learning and GPS trajectory data. A mathematical definition of gridlock is developed based on multiple connected links and neighboring intersections. The LSTM model allows a better understanding of temporal correlations.

However, I have two major concerns:

1. over your definition of congested intersection and definition of gridlock.

the congested intersections are those any of which two links are with lower speed than 5 km/h. However, for two directions of one link, we may find one direction is congested but another one is not congested. Thus, the intersection is not congested as you define in the study. In actual, the intersection may form gridlock.

Thus, it is better to count the link pairs that are congested and count the congested intersections for gridlock identification. The product of all those indicator variables may filter out many gridlock cases. 

    2. over your deep learning framework for gridlock prediction.

It is obvious that the gridlock consists of multiple neighboring intersections thus has spatial correlations. LSTM is modeled from the perspective of temporal correlations. 

Thus, I recommend to extend the deep learning framework to a spatial-temporal one.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

This article presented by the authors investigate on GPS vehicles that have passed through the links in the simulated gridlock-looped intersection area using the Long Short-Term Memory (LSTM) neural network has been applied to predict gridlock based on bottleneck states of intersections in the simulated urban road network. However, I will mention aspects to improve this article:
-The authors use acronyms in a bad way, so their correct way of using it, is to write the first letter in capital letters that represents the letter of the acronym, such as, "Simulation of Urban MObility (SUMO)". This problem must be corrected throughout the document.
-I recommend the authors to separate the Introduction because it is very extensive, this separation should consist of an Introduction Section and a Related Works Section.
-I recommend placing other references about the prediction of traffic in an urban environment such as:
=> Zambrano-Martinez, J. L., Calafate, C. T., Soler, D., Cano, J. C., & Manzoni, P. (2018). Modeling and characterization of traffic flows in urban environments. Sensors, 18 (7), 2020.
=> Guo, J .; Huang, W .; Williams, B.M. Adaptive Kalman filter approach for stochastic short-term traffic flow
rate prediction and uncertainty quantification. Transp. Res. C Emerg. Technol. 2014, 43, 50–64.
-Who provided the authors with the GPS data?
-What is the reason that the authors choose the data from 6am to 9am? To support their arguments, a graph should be placed that explains why they chose that time slot.
-The Legend of Figure 6 must be a little bigger to be able to read it without difficulty.
-The authors only use a single scenario, don't they compare with another scenario to apply the same study?
-At the moment that this study predicts traffic, how does it provide a solution to not congest the same sector with traffic anymore?
-What is the minimum credibility percentage for predicting traffic at an intersection?
-Figure 12, 13, 14, 15 is too small to understand.
-The conclusions are too brief or a few, which should expand it. In addition, they must include future works.
-Abbreviations are not necessary, since the authors mention the acronyms used throughout the text of the article.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Thanks for including my concerns in the first round. Currently, the manuscript has a better presentation.

Reviewer 4 Report

Thanks a lot to the authors for performing the suggested changes to improve the article.