Estimation of origin-destination (OD) vehicle flows and link capacity calibration are fundamental processes in transportation science, especially in the context of transport modelling and simulation. They ensure that transportation models accurately reflect real-world travel behaviour and network conditions.

This thesis develops a simulation-based optimization algorithm for network-wide link capacity calibration. To address the high dimensionality of large-scale networks, the algorithm is integrated with partial least squares (PLS) regression, which reduces the number of variables and enhances computational efficiency. The algorithm is evaluated on an urban road network in Stockholm, Sweden, where it demonstrates feasibility and higher efficiency compared to the simultaneous perturbation stochastic approximation (SPSA) method.

For large-scale OD estimation, this thesis advances the field in two main directions. First, it develops a data fusion framework that integrates multiple heterogeneous data sources, including mobile network data, link count observations, and turning proportion data. Second, it proposes several methods to enhance the computational efficiency of OD estimation in large urban networks. These include: (i) implementing data-driven network assignment (DDNA) using GPS data to construct a fixed OD–to–link mapping, thereby eliminating the need for iterative assignment within a bi-level optimization structure; (ii) applying non-negative matrix factorization (NNMF) for dimensionality reduction, which simplifies the optimization by reducing the number of variables; and (iii) developing a numerical solver based on an interior-point method that exploits structural properties of the assignment matrix, such as sparsity and linearity, to enhance computational performance. The proposed OD estimation methods are evaluated on real-world networks in central Stockholm and Norrköping, Sweden, demonstrating accurate and stable OD and link flow estimates with substantial gains in computational efficiency compared to solving the OD estimation problem without dimensionality reduction techniques or numerical solver improvement.

Origin-Destination (OD) matrices are essential inputs to both traffic planning and management. To enable traffic management decisions, the OD matrix needs to be estimated in the order of minutes, which is very challenging in many situations. However, new large-scale mobility data, such as vehicle probe data, in combination with computationally efficient estimation methods make it possible to estimate OD matrices sufficiently fast to enable traffic management decisions. In this paper, we propose a computationally efficient method that uses link count data and vehicle probe data, within a data-driven network assignment (DDNA) framework in combination with non-negative matrix factorization (NNMF), to estimate OD demand in urban networks. Historical data are used to construct a low-dimensional description of the OD estimation problem using non-negative matrix factorization. The method and the quality of the OD matrix estimated using the low-dimensional representation are evaluated on empirical data for central Stockholm, Sweden. The results demonstrate that the method provides a computationally fast technique to find an OD matrix that provides link flow estimates close to the link flow observations for both training and test sets.

Origin-Destination (OD) and link flows are crucial input to several traffic planning and management problems. In this paper we evaluate a data-driven network assignment method that first estimates the assignment process using GPS probe data, then performs a simultaneous and consistent OD and link flow estimation. The method is evaluated on empirical data for Stockholm, Sweden, showing that the method explains a significant part of the variance in link flow observations for both a training and test set of link flow observations. The results also show the importance of calibrating both the weight parameters in the OD estimation step and the Logit parameter in the route choice model. Using a too small weight on the a priori OD matrix will lead to large variance and over-fitting of the estimated OD against link counts, while a too large weight will cause unnecessary bias in the results.

Capacity of a road in an urban network is defined as the maximum number of vehicles which could pass this road in a unit time. Since travel delays occur when the travel demand exceeds the capacity, performance of congestion estimation strongly depends on capacity values. In order to obtain correct values of capacities in a network, one needs to implement calibration, which is a process to modify parameters to make the model outputs match observed data. A calibration problem is also known as an inverse problem in mathematics.

In this thesis, a new calibration method for road capacities in urban networks is presented. The method relies on Partial Least Squares (PLS) regression, which combines calibration and dimensionality reduction capabilities. A sampling strategy in determining which training data should be used is implemented to further im-prove the calibration efficiency and accuracy. Moreover, influence of different parameters such as wiggling amplitude in sensitivity analysis and number of loading vectors on calibration results are investigated. This method is demonstrated to be feasible and efficient in an urban road network (Stockholm, Sweden).

Besides, this method does not require any other constraints (such as non-negativity) in the optimization part and no additional terms need to be added to guarantee the closeness between initial guess of capacities and estimated values of those, which can be regarded as an indicator that this calibration method does not strongly rely on initial guess of input variables. It is a promising method since it can not only be used in capacity calibration but also has a potential on origin-destination (O-D) calibration problems which share a very similar structure: more unknown variables than measurements and similar data structures in input and output spaces. Even more generally, this method has the potential to be applied on most of high-dimensional inverse problems.