FORECASTING CONGESTION ON CALIFORNIA FREEWAYS: A STATISTICAL LEARNING APPROACH

Abstract: Congestion forecasting is a critical input for operational improvement project programming following asset management concepts. As the most populous state in the United States, California experiences high levels of traffic congestion that impacts commuters, businesses, and overall quality of life. Many factors such as population growth, economic development, limited roadway capacity, and increasing travel demand are contributing to congestion growth. Despite its importance, limited research has been conducted to estimate a single congestion growth rate for California freeways. This paper aims to establish a new methodology that combines statistical and machine learning approaches to forecast congestion and its growth rate on California freeways. The methodology utilizes monthly vehicle hours of delay data collected through the Performance Measurement System (PeMS), operated by the California Department of Transportation (Caltrans). In this method, stratified random sampling technique was applied to collect data from each of the Caltrans districts and the results were combined to establish a time series for the delay across the entire state. A Long Short-Term Memory (LSTM) neural network along with a linear regression model were used to forecast delay for the upcoming months based on historical patterns and estimate its growth rate. The results show that congestion grows at a rate of 8.5 percent per year for all the freeways covered by PeMS.