Daily bicycle traffic volume estimation: Comparison of historical average and count models

Abstract

This research tackles the problem of missing cycling counts at permanent count stations that experience frequent sensor malfunctions during the year. A simple approach that has been widely applied among practitioners is to use historical average values to fill in data gaps. A systematic assessment of this approach has not yet been undertaken. A more advanced approach that has been recently proposed by many researchers is to use regression count models that relate hourly/daily cycling counts to weather-specific variables such as temperature, precipitation, and wind speed, among others. Although most previous studies have focused on an explanatory analysis of variables' coefficients, an assessment of the estimation accuracy of the cycling count models is generally limited. The objective of this paper is to evaluate the use of the two approaches for estimating daily bicycle counts and undertake a fair quantitative comparison that highlights the accuracy of each. To benchmark the results, a comparison was also made between these two methods and a more sophisticated deep-learning autoencoder neural network model, which was developed in previous research. The current study made use of a large data set of about 13,000 daily bicycle volumes from the city of Vancouver, Canada. The data were collected between 2009 and 2011 at 22 different count stations. The results of the comparison showed poor performance of historical averages models, in which the mean absolute percentage error ranged between 36.9 and 59.4% in most cases. The application of count regression models led to improved estimation accuracy, in which the error was in the range of 20.0-34.4% with a weighted average of 25.8%. Further error analysis by month of the year showed that the average estimation errors of weekday daily bicycle volumes in July and August were relatively low: 15-17%. Despite the improved estimation accuracy of the count models, they were significantly outperformed by the more complex autoencoder neural network model. The results of this paper suggest the inappropriateness of using historical average data to compensate for missing counts. The implications of using different estimation methods are discussed.