ML for Forecasting and Analyzing Passenger Transport Behavior

Tamerlan Kabaloev27 May 2023

Problem

A metro analytics team in Moscow has to know how many passengers will enter each station tomorrow and each hour. Section closures and escalator repairs push people onto other transfers, and without a forecast the redistribution of flows is planned by hand. Forecasts are requested at short notice, so heavy custom models are not an option. The team's goal is passenger satisfaction rather than margin, so overloaded transfers are the cost that counts.

Example from the article: when the circular metro line is closed, passengers from "Kolomenskaya" to "Prospekt Mira" switch to other transfers, for example "Novokuznetskaya" to "Tretyakovskaya", and a minor collapse forms there.

Solution in brief

Turn fare-gate validations into daily and hourly flow series per station. Forecast them with low-code libraries: Prophet first, then PyCaret, which compares many models at once and did better on both granularities. Label stations as morning-peak, evening-peak or double-hump from their hourly profile.

Questions this case answers
  • How to forecast hourly and daily entry counts for each node of a network the day before, and classify each node by its shape of hourly load
  • How to choose a quick forecasting approach when requests arrive at short notice and heavy custom models are not an option
  • How to forecast daily and hourly passenger flow for each transit station
  • How to segment transit stations by their hourly flow profile

Tools and methods

  • PyCaret — compares several models at once and ranks features; tree-based models won most often
  • Prophet — trend, seasonality and holiday components; the first baseline
  • Hourly flow profile per station — splits stations into morning-peak, evening-peak and double-hump types
  • Fare-gate validations — transaction data per card ID, the source of flows
Continue reading — it's free

Google or a one-time email code. No password.