The study presents the methodology and system architecture of an integrated urban road network travel time prediction framework based on low-frequency probe vehicle data. Intended applications include real-time network traffic management, vehicle routing and information provision. The framework integrates methods for receiving a stream of probe vehicle data, map matching and path inference, link travel time estimation, calibration of prediction model parameters and network travel time prediction in real time. The system design satisfies three crucial aspects: computational efficiency of prediction, internal consistency between components and robustness against noisy and missing data. Prediction is based on a multivariate hybrid method of probabilistic principal component analysis, which captures global correlation patterns between links and time intervals, and local smoothing, which considers local correlations among neighbouring links. Computational experiments for the road network of Stockholm, Sweden and probe data from taxis show that the system provides high accuracy for both peak and off-peak traffic conditions. The computational efficiency of the framework makes it capable of real-time prediction for large-scale networks. For links with large speed variations between days, prediction significantly outperforms the historical mean. Furthermore, prediction is reliable also for links with high proportions of missing data.

The paper explores the potential of spatiotemporal network partitioning for travel time prediction accuracy and computational costs in the context of large-scale urban road networks (including motorways/freeways, arterials and urban streets). Forecasting in this context is challenging due to the complexity, heterogeneity, noisy data, unexpected events and the size of the traffic network. The proposed spatio-temporal network partitioning methodology is versatile, and can be applied for any source of travel time data and multivariate travel time prediction method. A case study of Stockholm, Sweden considers a network exceeding 11,000 links and uses taxi probe data as the source of travel times data. To predict the travel times the Probabilistic Principal Component Analysis (PPCA) is used. Results show that the spatio-temporal network partitioning provides a more appropriate bias-variance tradeoff, and that prediction accuracy and computational costs are improved by considering the proper number of clusters towards robust large-scale travel time prediction.

City-wide travel time prediction in real-time is an important enabler for efficient use of the road network. It can be used in traveler information to enable more efficient routing of individual vehicles as well as decision support for traffic management applications such as directed information campaigns or incident management. 3D speed maps have been shown to be a promising methodology for revealing day-to-day regularities of city-level travel times and possibly also for short-term prediction. In this paper, we aim to further evaluate and benchmark the use of 3D speed maps for short-term travel time prediction and to enable scenario-based evaluation of traffic management actions we also evaluate the framework for traffic flow prediction. The 3D speed map methodology is adapted to short-term prediction and benchmarked against historical mean as well as against Probabilistic Principal Component Analysis (PPCA). The benchmarking and analysis are made using one year of travel time and traffic flow data for the city of Stockholm, Sweden. The result of the case study shows very promising results of the 3D speed map methodology for short-term prediction of both travel times and traffic flows. The modified version of the 3D speed map prediction outperforms the historical mean prediction as well as the PPCA method. Further work includes an extended evaluation of the method for different conditions in terms of underlying sensor infrastructure, preprocessing and spatio-temporal aggregation as well as benchmarking against other prediction methods.

Recognition of spatio-temporal traffic patterns at the network-wide level plays an important role in data-driven intelligent transport systems (ITS) and is a basis for applications such as short-term prediction and scenario-based traffic management. Common practice in the transport literature is to rely on well-known general unsupervised machine-learning methods (e.g., k-means, hierarchical, spectral, DBSCAN) to select the most representative structure and number of day-types based solely on internal evaluation indices. These are easy to calculate but are limited since they only use information in the clustered dataset itself. In addition, the quality of clustering should ideally be demonstrated by external validation criteria, by expert assessment or the performance in its intended application. The main contribution of this paper is to test and compare the common practice of internal validation with external validation criteria represented by the application to short-term prediction, which also serves as a proxy for more general traffic management applications. When compared to external evaluation using short-term prediction, internal evaluation methods have a tendency to underestimate the number of representative day-types needed for the application. Additionally, the paper investigates the impact of using dimensionality reduction. By using just 0.1\% of the original dataset dimensions, very similar clustering and prediction performance can be achieved, with up to 20 times lower computational costs, depending on the clustering method. K-means and agglomerative clustering may be the most scalable methods, using up to 60 times fewer computational resources for very similar prediction performance to the p-median clustering.

Prediction of future traffic states is an essential part of traffic management and intelligent transportation systems. Previous work has shown that spatio-temporal clustering of traffic data such as flows or speeds into network day-types improves both the performance and the robustness of traffic predictions. Since some data types may not be available at a network-wide level, or only for certain periods, this paper investigates how similar such representative day-types are if based on different data types. The similarity of day-type clusters is evaluated with qualitative calendar visualization and two quantitative metrics, the Adjusted Mutual Information (AMI) which considers day-to-cluster assignments, and a new proposed Centroids Similarity Score (CSS) which compares centroids. The paper also explores the impact on flow and speed prediction performance of substituting one data type for the other in the clustering or classification phases. Using microwave sensor data from the Stockholm motorway network, our findings show that clusterings based on flows and speeds and across a range of clustering methods have reasonably high similarity. CSS is found to be a more relevant similarity indicator than AMI in the prediction application context. By capturing more relevant traffic state information, flow-based clustering and classification are robust for both flow and speed predictions, while speed-based clustering significantly degrades flow prediction performance.