Exploring the nonlinear relationships between human travel and road traffic congestions using taxi trajectory data

Abstract

Urban road traffic congestion remains challenging due to global urbanisation and has caused travel delays, energy consumption, and detrimental emissions. Therefore, exploring the potential dominant factors associated with traffic congestion generation is necessary to mitigate traffic congestion. The built environment around congested areas is the core factor in the generation of traffic congestion, however, only a few considered the impact of human travel features on congested roads. We divided human travel factors into purpose- and movement-related factors and explored the nonlinear relationship between human travel factors and traffic congestion. The results from taxi travel in Wuhan show that travel purpose factors mostly impact traffic congestion on low-grade inner-city short roads, while movement factors mainly impact the periphery ring or high-grade long roads. Movement-dominant congestions are widespread but not severe. Severe traffic congestion occurs mainly due to purpose-dominant travel. For purpose-dominant congestions, all excessive POI visits may worsen traffic congestion, and higher POI mixing degree has a positive effect on reducing congestion. For movement-dominant congestions, the detour rate and congestion level show a positive dependence, and the whole travel distance and travel accomplished rate indicate a U-shaped nonlinear relationship with congestion. This study provides detailed partial dependence plots of how congestion varies with human travel factors, providing insights and locational indications for traffic participants and urban designers to reduce congestion and improve urban mobility.

Appendices

Appendix 1: Relative interpretability difference of linear and nonlinear model

Table

Table 5 Frequency (%) of dominant human travel factors with maximal relative interpretability

5 lists the frequency of dominant human travel factors with maximum relative interpretability to traffic congestion in the nonlinear and linear models. The relative interpretability of each factor can be expressed based on the standardisation coefficient to represent the impact of travel factors on congestion in the OLS model (Xu et al. (Xu et al. 2021)). Normalisation was conducted to eliminate dimensional differences of all independent and dependent variables to ensure a relative interpretability index comparably in the OLS model. Notably, this study only explored the dominant possibilities among 13 variables of congestion; therefore, we did not necessarily cover all real leading factors, and other unaccounted potential dominant factors may be spread over these 13 variables.

The RF model showed that the most dominant category of traffic congestion was movement-related, while the OLS model restrictively consider that can predominantly account for the variation in the traffic congestion index. In fact, travel distance has significant nonlinear relationships to the traffic system organisation (He et al. (He et al. 2022)), which was critically underrated by the OLS model. The RF model could capture an implicit nonlinear relationship between the travel distance and the curvature of traffic congestions to a large extent, indicating that movement-related factors dominate at least half of all traffic congestions in the dataset. In contrast, the linear OLS model attempts to fit the data into a straight line in the case of a nonlinear relationship, thereby obscuring the nonlinear impact of movement-related factors. Though the RF model similarly identifies purpose mixing as a more dominant impact than a single travel purpose, the OLS model narrowly considers the purpose mixing degree with linear relationships as the most dominant driving factor. Conformably, both models exhibit similar dominant ranks of single POI travel purposes to traffic congestions, illustrating purpose-related factors present a linear relationship do not mislead linear model.

Appendix 2: Threshold effects of human travel factors in peak/off-peak perspective

Figure 

Fig. 11

Partial dependence plots of human travel factor during peak/off-peak periods

11 shows the partial dependence plots of human travel factors as dominant variables from a peak and off-peak perspective. It is evident that congestion risk is consistently higher during peaks compared to off-peaks. The partial dependence plots at different periods provide specific congestion mitigation strategies. During workday peak, travel purpose optimization can reduce congestion index on 40.58% of roads (combined with Table 4); however, achieving congestion levels below 1.5 (considered basically smooth) proves challenging. Travel purpose optimization is intricately tied to urban design, necessitating long-term efforts to introduce more accessible destinations placed at corresponding origins of related roads. Interestingly, the congestion impact is less severe on road dominated by POI mixing, and the optimal functional mixing degree of destinations from these roads can be precisely designed according to Fig. 10d. Conversely, travel movement-related optimization emerges as a more effective approach to reduce congestion on an additional 59.42% of roads, easily bringing the congestion index below 1.4. During Workday off-peak, purpose-dominant and movement-dominant congestions tend to be less intense, aligning with common expectations. Shopping and lodging are identified as the most dangerous contributors on their dominant congestions, indicating a notable role in urban planning and renewal. Furthermore, modeling holidays as an entire period offsets dependence on travel purpose factors and loses potential indications.