Estimating city wide hourly bicycle flow using a hybrid LSTM MDN

Abstract

This study proposes a novel method to improve the estimation of hourly bicycle flow. The specific model employed is a Long Short-Term Memory Mixture Density Network (LSTMMND). This model presents the additional upside approximates a distribution of cycling flow conditional on the input data and hence addresses potentially unobserved heterogeneity. In a case of city-wide bicycle flow in Copenhagen, the LSTMMDN yielded ∼ 75% more accurate bicycle flow estimates than the calibration-based approach currenty used by transport agencies. The paper further quantifies the improvement of accident analyses brought on by the improved bicycle volume measures. The LSTMMDN estimates result in an improved model fit in a crash model, compared to other estimates for bicycle exposure, with all other variables unchanged. Overall, the results strongly indicate that investing in more advanced methods for bicycle volume estimation will benefit the quality of performance measures related to bicycle issues computed by transport agencies.