Integrating heterogeneous data sources for traffic flow prediction through extreme learning machine

Abstract

Traffic flow prediction is important to transportation policy and management. Time-series models based on historical data can hardly reflect non-recurrent events and the impact of multiple factors. This paper proposes Internet search engine data and environmental data as supplemental data sources for capturing the latest changes in multiple factors of transportation. The time series model and extreme learning machine are built for San Francisco traffic flow prediction based on real heterogeneous data sources. First, the time series model based on historical traffic data is used as a benchmark to predict traffic flow. Then, we integrate heterogeneous data with the time series model and extreme learning machine for comparison. We find that the ELM with heterogeneous data improves on the linear model in both level and directional accuracy. These results emphasize that the heterogeneous data is more informative than the original traffic data used by other researchers and demonstrate the utility of the nonlinear approach for addressing this data selection problem in the context of prediction using search trend data, leading to an improvement in traffic flow prediction accuracy.