MRTS FLOW PREDICTION

Introduction

Mass Rapid Transit Systems (MRTS), also referred to metro system, generate a lot of data. These data have a potential to give useful visualization. Analyzing it, we can come up with applications and solutions to improve and benefit these systems. One of such applications is of predicting passenger ridership between any given two stations in the metro network. The historical data visualization helps to design the system infrastructure. Predicting ridership at a station helps on many fronts and enables real time decision making with a potential to optimize resources. Some useful actions for the metro management with the daily ridership prediction are:

• Scheduling more trains on busy stretches automatically.
• Helping advertising companies do bidding for ad spots in real-time (on digital displays)
• Aiding managers to regulate staffing more effectively

Methodology

This work consists of analyzing the Automatic Fare Collection (AFC) data of Kochi Metro Rail Limited (KMRL). Useful visualizations based on the data have been done. Importantly, prediction of passenger flow between a pair of stations has been demonstrated using a novel method outlined by the paper: <https://www.hindawi.com/journals/jat/2018/5942763/

Reason for choosing this paper:

• The metro systems the study (paper) was based on, have very similar characteristics to Kochi Metro.

• The goal of the paper matches with ours.

• Not many papers are available doing analysis of data generated by metro systems.

Data Description

KMRL operates in the heartland of Cochin City. At the time of this work, the Kochi Metro network was linear, stretching from Maharajas College to Aluva.

Figure 1

The data available is of Automatic Fare Collection (AFC) system. This data is collected by the electronic gates installed at every station, through which every passenger has to go through for entering or exiting a station by submitting their tickets.

The AFC data is of the years June 2017 to December 2018. The data is stored in form of Excel sheets.

Important fields of the data are:

• Station: Specifies which of the 16 stations the data entry is associated with.

• Fare product: Specifies for what purpose the person enters the station e.g. Single journey trip, Round journey trip, station access card (for staff) etc.

• Ticket/Card Number: A unique hashed alphanumeric sequence given to identify each row entry and thus each bought ticket uniquely.

• Transaction Type: Takes values Entry or Exit to ascertain whether the person entered the station or exited.

• Date and Time: Give the date and time of the entry or exit of the person for a station.

Some Visualizations

4.1 Station wise ridership: The net inflow and outflow from stations was done to understand the travel demand at different stations.

Explanation about how the data is processed:

Data was loaded into a pandas dataframe. Data was filtered by date (over here December 2017) was taken. Data was separated by Transaction type being 'Entry' or 'Exit'. Afterwards, the filtered data was grouped by stations. Then bar plots are made based on entry and exit data. The passenger entering different station is shown in Fig 2 and the passengers exiting different stations are shown in Fig 3.

Figure 2

Figure 3

4.2 Travel pattern during the day: The inflow and outflow of passengers at different stations was done to give the flow of the passengers during different hours of the day.

Explanation about how the data is processed:

A function was built to give passenger inflow/outflow for all station for a day or a range of dates. Taking inputs as the range of dates to generate the output for and whether to give outflow or inflow volume. Data is aggregated based on transaction time with intervals of 1 hour (option given in function). Then the data is filtered based on entry or exit, after which a line plot is done. In figure 4 the visualization for a single day in December 2017 of outflow volume of 16 stations is presented.

Figure 4

5 Ridership Prediction

5.1 For predicting the passenger ridership, Edapally is chosen as the origin and M.G Road as the destination station. The model chosen will also consider adjacent stations of this OD (Origin Destination) pair. The stations adjacent to Edapally stations are Changampuzha and Pathadipalam while the stations adjacent to M.G Road are Lissie and Maharajas College.

5.2 Explanation of Origin-Destination (OD) Model

Instead of only considering past values of the flow between Edappaly and M.G Road station and predicting based on such information, [the model takes into consideration the effect which the traffic of adjacent stations is having on the flow of our concerned OD.]{.ul}

Justification: We assume that the traffic of the adjacent OD pairs and that of our concerned OD are positively correlated. We do this because of the intuition that if adjacent ODs are busy then that indicates that the target OD (Edappaly and M.G Road) is busy too.

Thus we are also taking the past data of adjacent OD pairs and including them in our feature vector. There will be thus nine entries corresponding to each OD pair (because there are three origins Edapally, Changampuzha, Pathadipalam and three destinations M.G Road, Lissie and Maharajas College).

Index	Origin	Destination	Number of passengers	travelling(on a date)
0	Changampuzha Park	Lissie	53
1	Changampuzha Park	M.G Road	152
2	Changampuzha Park	Maharajas College	326
3	Edapally	Lissie	751
4	Edapally	M.G Road	1091
5	Edapally	Maharajas College	2570
6	Pathadipalam	Lissie	42
7	Pathadipalam	M.G Road	98
8	Pathadipalam	Maharajas College	201

Table 1

Our feature vector would be holding values like the column with the heading "Some Date" and the value we're interested in predicting is of 5^th^ row i.e. Edapally to M.G Road.

After processing the data and making training samples according to different dates we'll feed the processed data to a Kneighborsregressor algorithm to predict the OD flow for Edapally to M.G Road for the next date.

5.3 Processing the Data

Steps:

1. Data is filtered by date.

2. Two dataframes (suppose A and B) are made based on 'Transaction type' being entry or exit.

3. All unnecessary columns are dropped.

4. The dataframes A and B are merged (natural join) into a new dataframe (suppose C) based on 'Ticket/Card number' ( as this serves as an identification for a single passenger).

5. All duplicates rows in C are dropped.

6. Data in C is grouped by stations of A and station of B.

7. Data is then filtered on the condition that stations of A should be in ['Edapally','Pathadipalam','Changampuzha Park'] and station of B should be in [' M.G Road ',' Lissie ',' Maharajas College '].

8. These steps are done for every date (available from the original data).

9. Weekends are ignored as we assume they have a different pattern compared to working days.

Kindly find the data processing implementation in kmrl1.ipynb (https://github.com/aye21874/Metro_Flow_Prediction-Code-)

5.4 Making the input data

Our feature vector contains data for target OD (Edapally to M.G Road) as well as neighboring ODs. In addition to this each our input data consists of 5 consecutive dates. Thus our feature vector becomes a feature matrix. This matrix is made by concatenating the last column of Table 1 for 5 consecutive dates. The feature matrix is depicted in Table 2.

Index	2017/10/03	2017/10/04	2017/10/05	2017/10/06	2017/10/09
0	20	30	32	49	29
1	42	105	87	107	81
2	154	283	258	274	184
3	189	289	336	308	257
4	361	529	497	523	411
5	794	1146	1126	1212	989
6	27	27	44	37	40
7	30	70	64	61	68
8	81	134	140	103	110

Table 2

The feature matrix is again converted into a single column feature vector by appending one column below the other hence generating a feature vector with 45 (9*5) rows. The desired output is passenger flow value of target OD for the next date. Over here in Table 2 we would be interested in finding value of row with index 4 for date 2017/10/10. We would be interested in index 4 as it corresponds to our target OD as shown in Table 1. We do this for every date and get 320 samples from the available data.

5.5 Training the Model

• After generating the input data, we append target output to it as the K-nearest Neighbors Regression Algorithm is a Supervised learning Algorithm.

• After doing this we split the data into training and testing (80:20) and apply Principal Component Analysis (PCA) on the training samples.

• We apply PCA because the feature vector has a very high dimension (45) and most probably would contain redundant information, increasing the uncertainty of the prediction. Thus applying PCA captures the variations in lower dimensions.

• We take the first 10 principal components having the highest sample variance.

• We train our model on this training data to predict the Target OD flow.

5.6 Model Evaluation

The model is evaluated based on Mean absolute percentage Error (MAPE). The Trained model is given test set data and the difference between its output and the actual target value is compared by MAPE.

After running the model on the test set we got an MAPE value of ~15.

Kindly find the implementation with the quoted result in kmrl2.ipynb (https://github.com/aye21874/Metro_Flow_Prediction-Code-)

This means our prediction is off by 15% of the true value on average. Thus the model has a high potential of utilization in real world scenario and can be used by the Metro administrator.

6 Coursera Skills displayed and Conclusion:

The coursera skills displayed are as under:

• Mathematics for machine learning course helped me greatly in understanding machine learning concepts in general and the significance of PCA.

• After going through the IBM python for data visualization, I was able to create meaningful visualizations that provide insight and help making decisions.

The project demonstrated the working of a novel model where adjacent ODs data is also analyzed in making accurate prediction. The project applied a machine learning algorithm on the data generated by the Kochi Metro System and got an agreeable error rate on the predicted value comparable to the paper it referred. The model slightly under fits the data, mainly because the data generated is less. If more data is available better accuracy could be achieved.

Still better results would be achieved in dense metro networks such as Delhi Metro, Bengaluru Metro etc as the model relies on correlations between adjacent ODs and considering these Metro's huge traffic, relations could be spotted more easily by the ML algorithms in those cases.