A utility-based traffic sensor network optimization for data-driven traffic prediction

Abstract

Traffic sensors collect real-time data on current traffic conditions, and much research has focused on determining optimal sensor locations for various monitoring and information-inference tasks. Meanwhile, data-driven models for time-series traffic prediction increasingly rely on these sensor inputs. However, existing sensor-location studies rarely account for the specific role of sensor data in data-driven prediction models. This work aims to reduce the size of the sensor network with only a minor impact on model performance. We proposed a genetic algorithm method for traffic sensor network optimization, with prediction accuracy and quantified epistemic uncertainty as the objectives. The results show that sensors distributed in various locations have different utilities for prediction performance. In the case study of Amsterdam highway ring roads (A10), we can obtain a reduction of 60% of the total sensor budget while preserving adequate prediction model performance.