Our Client is one of the leading global providers of systems and components for freight and transit railcars and locomotives. They provide live freight car locations to the end clients.
The client wanted to improve customer service by developing an automated solution to predict the estimated arrival time of freight containers through near real-time estimates on arrival/estimated delay as the rail container moves on its route based on historical data.
- No pre-defined data on the train origin and the final destination
- No defined railway routes between stations
- Only 2 months of data available for modeling
- Input data for the model included Geolocation pings from various trips containing tracking Device ID, Latitude/Longitude, Geofence/City, timestamp, motion status, and Geofence type
- Identified origin and destination pairs based on geofence type. Grouped containers from the same train into one trip by offsetting time stamp
- Engineered 55 features including seasonal variables, current trip variables, and historical variables
- Identified key input features -distance left, historical travel & dwell time, origin time & day, and developed model leveraging XGBoost regression model
- Productionized model with an automated process to predict ETA for live trips. Model handled all possible exceptions such that the module is break-free. Logged every step for error-handling
- Predicted the ETAs for 95% of the cases within +/- 60 minutes
- Accurate predictions enabled the users to plan logistics better
- Predictions helped the railcar manufacturer to improve its customer service