Real time tracking and prediction of public transport transit speed in Pays de Gex

Fernando González Cortés

fernando.gonzalez@geomati.co

http://fergonco.org/ | http://geomati.co/

## Real time tracking and prediction of public transport transit speed in Pays de Gex

4am: /GetAllNextDepartures.xml


                        <?xml version="1.0" encoding="UTF-8"?>
                        <nextDepartures>
                            <timestamp>2017-10-20T04:12:09+0200</timestamp>
                            <stop>
                                <stopCode>VATH</stopCode>
                                <stopName>Val-Thoiry</stopName>
                            </stop>
                            <departures>
                                <departure>
                                    <departureCode>216940</departureCode>
                                    <timestamp>2017-10-20T06:30:09+0200</timestamp>
                                    <reliability>T</reliability>
                                </departure>
                                <departure>
                                    <departureCode>216806</departureCode>
                                    <timestamp>2017-10-20T06:50:09+0200</timestamp>
                                    <reliability>F</reliability>
                                </departure>
                                <departure>
                                    <departureCode>217160</departureCode>
                                    <timestamp>2017-10-20T07:10:09+0200</timestamp>
                                    <reliability>T</reliability>
                                </departure>
                            </departures>
                        </nextDepartures>

6.30am: /GetThermometer.xml


                        <?xml version="1.0" encoding="UTF-8" standalone="yes"?>
                        <thermometer>
                            <timestamp>2017-10-20T04:12:21+0200</timestamp>
                            <stop>
                                <stopCode>VATH</stopCode>
                                <stopName>Val-Thoiry</stopName>
                            </stop>
                            <lineCode>Y</lineCode>
                            <destinationName>Ferney-Voltaire</destinationName>
                            <destinationCode>FERNEY-VOLTAIRE</destinationCode>
                            <steps>
                                <step>
                                    <departureCode>216940</departureCode>
                                    <timestamp>2017-10-20T06:30:00+0200</timestamp>
                                    <stop>
                                        <stopCode>VATH</stopCode>
                                        <stopName>Val-Thoiry</stopName>
                                    </stop>
                                </step>
                                <step>
                                    <departureCode>216941</departureCode>
                                    <timestamp>2017-10-20T06:32:56+0200</timestamp>
                                    <stop>
                                        <stopCode>THGA</stopCode>
                                        <stopName>Thoiry-Gare</stopName>
                                    </stop>
                                </step>
                                <step>
                                    <departureCode>216942</departureCode>
                                    <timestamp>2017-10-20T06:34:20+0200</timestamp>
                                    <stop>
                                    ...

Store event in Shift table:

TPGStopRoute:

Model selection:

✓ Linear regression
✓ Polynomial regression
✗ Time series
✗ K-Nearest neighbours

speed = β₁·weekday + β₂·schoolfr + β₃·minutesDay^16 + β₄·weekday·minutesDay

Model generation

For each segment in the network:
- Prepare the data: filter errors out, calculate speed, generate variables: school in france, day of the week, etc.
- Make R build a model
```
glm(data = speeds, 
  speed ~ weekday + 
          schoolfr + 
          poly(minutesDay, 16) + 
          weekday * minutesDay)
```
- Store the model in the database

Forecast generation

Each 15 minutes:
- for each segment in the network:
  - Get the model from the database
  - Gather the current variables: weekday, school, etc.
  - Make R produce a forecast
```
predict.glm(model, newdata = forecastDataset)
```
  - Store the forecast in the database

Forecasted speeds


        geom           |  id  |     draw_timestamp     | speed   
-----------------------+------+------------------------+--------
 LINESTRING(6.14137... | 8133 | 2017-10-20 15:15:00+00 |      9  
 LINESTRING(6.14085... | 8215 | 2017-10-20 17:15:00+00 |     10  
 LINESTRING(6.14108... | 8210 | 2017-10-20 16:00:00+00 |      9  
 LINESTRING(6.14132... | 8170 | 2017-10-21 05:45:00+00 |     12  
 LINESTRING(6.14090... | 8205 | 2017-10-20 14:45:00+00 |      9  
 LINESTRING(6.14123... | 8154 | 2017-10-20 20:30:00+00 |     10  
 LINESTRING(6.14126... | 8132 | 2017-10-20 15:15:00+00 |      9  
 LINESTRING(6.14174... | 8126 | 2017-10-20 13:30:00+00 |     10  
 LINESTRING(6.14173... | 8196 | 2017-10-21 12:15:00+00 |     12  
 LINESTRING(6.14120... | 8144 | 2017-10-20 18:00:00+00 |     10

Performance considerations

Model generation: ~2h
Forecast generation: ~2m

Forecasted speeds


        geom           |  id  |     draw_timestamp     | speed   
-----------------------+------+------------------------+--------
 LINESTRING(6.14137... | 8133 | 2017-10-20 15:15:00+00 |      9  
 LINESTRING(6.14085... | 8215 | 2017-10-20 17:15:00+00 |     10  
 LINESTRING(6.14108... | 8210 | 2017-10-20 16:00:00+00 |      9  
 LINESTRING(6.14132... | 8170 | 2017-10-21 05:45:00+00 |     12  
 LINESTRING(6.14090... | 8205 | 2017-10-20 14:45:00+00 |      9  
 LINESTRING(6.14123... | 8154 | 2017-10-20 20:30:00+00 |     10  
 LINESTRING(6.14126... | 8132 | 2017-10-20 15:15:00+00 |      9  
 LINESTRING(6.14174... | 8126 | 2017-10-20 13:30:00+00 |     10  
 LINESTRING(6.14173... | 8196 | 2017-10-21 12:15:00+00 |     12  
 LINESTRING(6.14120... | 8144 | 2017-10-20 18:00:00+00 |     10

Gathered speeds: Shift

TPGStopRoute

OSMSegment

And...

Performance considerations

Model generation: ~2h
Forecast generation: ~2m
Data transformation made with PostgreSQL materialized views: ~1m

Study area

```
  4000 OSM segments
```
```
x   96 timestamps
```
```
------
```
```
384000 records to draw
```
Bidimensional data!

Thanks!

Slides --> http://fergonco.org/slides/foss4gbe17/

More details --> http://fergonco.org/blog.html