Speed prediction in large and dynamic traffic sensor networks

Highlights

• Dynamic traffic sensor networks bring challenges in the context of urban mobility

• We evaluate three approaches for speed prediction over large/dynamic sensor networks

• The global and cluster-based approaches provide accurate and robust prediction models

• The global approach solves the cold start problem

• We provide a large dataset and assess the effectiveness of the three approaches

Abstract

Smart cities are nowadays equipped with pervasive networks of sensors that monitor traffic in real-time and record huge volumes of traffic data. These datasets constitute a rich source of information that can be used to extract knowledge useful for municipalities and citizens. In this paper we are interested in exploiting such data to estimate future speed in traffic sensor networks, as accurate predictions have the potential to enhance decision making capabilities of traffic management systems. Building effective speed prediction models in large cities poses important challenges that stem from the complexity of traffic patterns, the number of traffic sensors typically deployed, and the evolving nature of sensor networks. Indeed, sensors are frequently added to monitor new road segments or replaced/removed due to different reasons (e.g., maintenance). Exploiting a large number of sensors for effective speed prediction thus requires smart solutions to collect vast volumes of data and train effective prediction models. Furthermore, the dynamic nature of real-world sensor networks calls for solutions that are resilient not only to changes in traffic behavior, but also to changes in the network structure, where the cold start problem represents an important challenge. We study three different approaches in the context of large and dynamic sensor networks: local, global, and cluster-based. The local approach builds a specific prediction model for each sensor of the network. Conversely, the global approach builds a single prediction model for the whole sensor network. Finally, the cluster-based approach groups sensors into homogeneous clusters and generates a model for each cluster. We provide a large dataset, generated from $\sim$1.3 billion records collected by up to 272 sensors deployed in Fortaleza, Brazil, and use it to experimentally assess the effectiveness and resilience of prediction models built according to the three aforementioned approaches. The results show that the global and cluster-based approaches provide very accurate prediction models that prove to be robust to changes in traffic behavior and in the structure of sensor networks.

Introduction

Highly populated cities increasingly face mobility challenges caused by transport and traffic. The huge volume of data collected by real-time traffic monitoring sensors provides new opportunities to develop models and algorithms that enhance transportation services towards intelligent transportation systems, in particular those dealing with traffic predictions. Vehicle speeds on road networks are determined by complex traffic processes governed by stochastic and non-linear interactions between individual drivers [1], hence predicting the speed of vehicles is as complex as predicting the underlying traffic processes. Short-term traffic prediction techniques have been investigated and exploited since some time [2]. However, the emergence of smart cities, where urban areas are covered by massive amounts of sensors, combined with the development of transportation technologies, requires traffic prediction techniques that are fast, scalable, and suitable for complex and heterogeneous sensors networks like those deployed in smart cities. Many different traffic sensor technologies are currently used to monitor road networks, such as those based on inductive-loop detectors, magnetometers, video image processors, microwave radar sensors, laser radar sensors, passive infrared sensors, ultrasonic sensors, passive acoustic sensors, and devices exploiting combinations of the aforementioned technologies [3].

In this work we focus on sensors capable of capturing the speed of vehicles traveling over large and dynamic road networks, where sensors can be added or removed from the network for various reasons, and address the problem of training accurate prediction models that are capable of maintaining their accuracy over time – we call this the model aging problem – and cope with structural changes affecting sensor networks — we call this the network dynamicity problem. In this context we assume that sensors collect their observations in the form of textual data and periodically send such information to a centralized entity. We also assume that some centralized entity is in charge of training prediction models according to the available sensor observations.

We address these challenges by proposing and analyzing three different approaches that can be used to train machine-learned prediction functions: local, global, and cluster-based. The local approach is the solution commonly used in the literature, where each sensor is considered separately from others to train a specific predictive function. This approach suffers the cold start problem and therefore hardly applies to dynamic sensor networks, where sensors may be continuously added and removed on a daily basis. Moreover, in large and dynamic sensor networks the local approach requires to train and maintain a large amount of different prediction models. To overcome these issues we propose the global and cluster-based approaches, where models are trained on data coming from all the sensors in the network (or groups of similar sensors, in the cluster-based case) to build resilient predictive functions. The global approach provides substantial benefits in terms of reduced complexity and costs. Furthermore, by relying on a single prediction function that is independent from specific sensors, the global approach naturally solves the cold start problem. Moreover, the global approach is expected to be robust with respect to structural changes occurring in sensor networks, thus also addressing the dynamicity problem.

We also tested a cluster-based approach to prove its potential in representing a viable compromise between the local and global approaches. Specifically, the cluster-based approach trains distinct predictive functions for groups of similar sensors, where sensors are clustered according to some similarity metric; depending on the number of clusters, the behavior of this approach resembles the one of the local approach (when a high number of clusters is used) or the behavior of the global one (when few clusters are used). From the experimental evaluation we cannot conclude yet that this approach indeed represents a good compromise, since results are discordant and further work is needed. The contributions of this paper can be summarized as follows:

• we propose the global and cluster-based approaches for learning vehicle speed prediction functions in large and dynamic sensor networks.

• driven by three experimental questions, we provide a comprehensive evaluation to assess the effectiveness of the predictive models trained according to the three approaches. The training is conducted by using different state-of-the-art machine learning algorithms on a large, real-world sensors dataset. The dataset covers a time span of 12 months, during which 130 (145) sensors were added (removed) to (from) the network. The evaluation shows that the models created using the global approach represent good solutions when dealing with dynamic sensor networks, as they prove to be accurate and resilient both to model aging and to structural changes in the sensor infrastructure (which, in turn, includes the cold start problem).

• We release to the scientific community the real-world dataset used to assess our proposals. The dataset originates from $\sim$1.3 billion records collected during the whole 2014 by 272 different road traffic sensors deployed in the city of Fortaleza, Brazil. Due to privacy concerns we do not release the original raw data, but a dataset obtained after an aggregation and cleaning process. To the best of our knowledge, this is the largest and richest dataset made publicly available for research on speed prediction in dynamic sensor networks.

The paper is structured as follows: Section 2 reports an overview of the related works dealing with the traffic prediction problem. Section 3 defines our prediction problem and discusses three approaches to solve the problem. Section 4 presents the dataset used in our experiments, as well as the pre-processing steps used to transform the data into a format suitable for speed prediction. Section 5 details the experimental evaluation and discusses the results. Finally, Section 6 draws the final conclusions and sketches potential lines of future research.