Investigating micro-simulation error in activity-based travel demand forecasting: a case study of the FEATHERS framework

Activity-based models of travel demand have received considerable attention in transportation planning and forecasting in recent years. However, in most cases they use a micro-simulation approach, thereby inevitably including a stochastic error that is caused by the statistical distributions of random components. As a consequence, running a transport micro-simulation model several times with the same input will generate different outputs, which baffles practitioners in applying such a model and in interpreting the results. A common approach is therefore to run the model multiple times and to use the average value of the results. The question then becomes: what is the minimum number of model runs required to reach a stable result? In this paper, systematic experiments are carried out using Forecasting Evolutionary Activity-Travel of Households and their Environmental RepercussionS (FEATHERS), an activity-based micro-simulation modelling framework currently implemented for the Flanders region of Belgium. Six levels of geographic detail are taken into account. Three travel indices – average daily activities per person, average daily trips per person and average daily distance travelled per person, as well as their corresponding segmentations – are calculated by running the model 100 times. The results show that the more disaggregated the level, the larger the number of model runs is needed to ensure confidence. Furthermore, based on the time-dependent origin-destination table derived from the model output, traffic assignment is performed by loading it onto the Flemish road network, and the total vehicle kilometres travelled in the whole Flanders are subsequently computed. The stable results at the Flanders level provides model users with confidence that application of FEATHERS at an aggregated level requires only limited model runs.     

Investigating the predictive performance of computational process activity-based transportation models

The aim of this paper is to achieve a better understanding of computational process activity-based models, by identifying factors that influence the predictive performance of A Learning-based Transportation Oriented Simulation System model. Therefore, the work activity process model, which includes six decision steps, is investigated. The manner of execution in the process model contains two features, activation dependency and attribute interdependence. Activation dependency branches the execution of the simulation while attribute interdependence involves the inclusion of the decision outcome of a decision step as an attribute to subsequent decision steps. The model is experimented with by running the simulation in four settings. The performance of the models is assessed at three validation levels: the classifier or decision step level, the activity pattern sequence level and the origin–destination matrix level. The results of the validation analysis reveal more understanding of the model. Benchmarks and factors affecting the predictive performance of computational activity-based models are identified.     

A data-driven approach for origin–destination matrix construction from cellular network signalling data: a case study of Lyon region (France)

Transportation - Spatiotemporal data, and more specifically origin–destination matrices, are critical inputs to mobility studies for transportation planning and urban management purposes....     

Exploratory analysis of Zipf’s universal power law in activity schedules

People’s behavior is governed by extremely complex, multidimensional processes. This fact is well-established in the transportation research community, which has been working on travel behavior (travel demand) models for many years. The number of degrees of freedom in a person’s activity schedule is enormous. However, the frequency of occurrence of day-long activity schedules obeys a remarkably simple, scale-free distribution. This particular distribution has been observed in many natural and social processes and is commonly referred to as Zipf’s law, a power law distribution. This research provides evidence that activity schedules from various study areas exhibit a universal power law distribution. To this end, an elaborate analysis using 13 household travel surveys from diverse study areas discusses the effect of proportional outlier removal on the power law’s exponent value. Statistical evidence is provided for the hypothesis that activity schedules in all these datasets exhibit a power law distribution with a common exponent value. The study proposes that a Zipf power law could be used as an additional dimension within a travel demand model’s validation process. Contrary to other validation methods, no new data is required. The observation of a Zipf power law distribution in the generated schedules appears to be a necessary condition. Additionally, the universal activity schedule distribution might enable the full integration of activity schedules in models based on universal mobility patterns.

     

The usefulness of the Sequence Alignment Methods in validating rule-based activity-based forecasting models

This research paper aims at achieving a better understanding of rule-based activity-based models, by proposing a new level of validation at the process model level in the A Learning-based Transportation Oriented Simulation System (ALBATROSS) model. To that effect, the work activity process model, which includes six decision steps, has been investigated. Each decision step is evaluated during the prediction of the individuals?? schedules. There are specific decision steps that affect the execution pattern of the work activity process model. So, the comportment of execution in the process model contains activation dependency. This branches the execution and evaluation of each agent under examination. Sequence Alignment Methods (SAM) can be used to evaluate how similar/dissimilar the predicted and observed decision sequences are on an agent level. The original Chi-squared Automatic Interaction Detector decision trees at each decision step utilized in ALBATROSS are compared with other well known induction methods chosen to appraise the purpose of the analyses. The models are validated at four levels: the classifier or decision step level whereby confusion matrix statistics are used; The work activity trips Origin?CDestination matrix level; the time of day work activity start time level, using a correlation coefficient; and the process model level, using SAM. The results of validation on the proposed process model level show conformity to all validation levels. In addition, the results provide additional information in better understanding the process model??s behavior. Hence, introducing a new level of validation incur new knowledge and assess the predictive performance of rule-based activity-based models. And assist in identifying critical decision steps in the work activity process model.