An Open-Source Framework to Implement Kalman Filter Bus Arrival Predictions

Abstract

The accuracy of vehicle arrival predictions affects every aspect of transit performance including ridership, reliability, and operating costs. Kalman Filter algorithms have been shown to provide more accurate predictions than simple regression. This paper presents a scalable framework to implement Kalman Filters on an entire bus network running live. A novel architecture to cache the data and weight inputs based on current operating conditions is presented. All the necessary features to support Kalman Filter predictions are described and implemented in TheTransitClock, an open-source prediction tool. TheTransitClock was deployed on the Metro Transit bus network in the Minneapolis-St. Paul region for one month. The Kalman Filter algorithm predicted the arrivals of over 900 buses on 167 routes. The accuracy and sensitivity of the method was compared to a schedule-based prediction method used in practice. The Kalman Filter was found to provide more accurate and stable predictions, especially in times and places that are difficult to predict with conventional methods.

Bus Single-Trip Time Prediction Based on Ensemble Learning

“…(2) Kalman filtering technique, which is formed by introducing the state space into modern control theory, has been applied to short-term traffic demand and travel time prediction of highways [ 14 ]. Since the Kalman can fully adapt to irregular changes (Kalman gain), it is more suitable for single-step predicting, but its prediction accuracy decreases significantly in multi-step predicting [ 15 ]. (3) Artificial neural network is a model that explores the nervous system functions of the human brain by modeling and linking neurons (i.e., the basic units of the human brain) to simulate the functions of the human brain.…”