Exploring Urban Mobility Trends using Cellular Network Data

Abstract

The growth of urban areas intensifies the need for sustainable, efficient transportation infrastructure and mobility systems, driving initiatives to enhance infrastructure and public transport while reducing congestion and emissions. By utilizing real-world mobility data, a data-driven approach can provide crucial insights for planning and decision-making. This study explores the efficacy of leveraging telecoms data from cellular network signals for studying crowd movement patterns, focusing on Trondheim, Norway. It examines routing reports to understand the spatiotemporal dynamics of various transportation routes and modes. A data preprocessing and feature engineering framework was developed to process raw routing reports for historical analysis. This enabled the examination of geospatial trends and temporal patterns, including a comparative analysis of various transportation modes, along with public transit usage. Specific routes and areas were analyzed in-depth to compare their mobility patterns with the broader city context. The study highlights the potential of cellular network data as a resource for sha** urban transportation and mobility systems. By identifying deficiencies and potential improvements, city planners and stakeholders can foster more sustainable and effective transportation solutions.

1 Introduction

The expansion of urban centers such as Trondheim, Norway’s third-largest city with a 2023 population of approximately 206,000 [2], underscores the pressing need for infrastructure modernization to address challenges such as traffic congestion and pollution driven by rising populations and mobility demands. Trondheim’s growth—spurred by significant educational and research institutions like the Norwegian University of Science and Technology, SINTEF, and St. Olav’s University Hospital—places pressure on its transportation infrastructure and mobility system. This development has led to a shift towards sustainable mobility solutions, with initiatives like Miljøpakken and MobilitetsLab Stor-Trondheim (MoST) emphasizing public and active transportation. Achieving such goals necessitates a data-driven approach towards understanding and improving the spatiotemporal dynamics of urban mobility systems for enhanced efficiency, sustainability, and human well-being.

One significant hurdle faced by such mobility initiatives is the difficulty in precisely evaluating the effectiveness and necessity of projects and interventions in alignment with their mobility goals. This evaluation demands the collection of detailed mobility data across the entire system. Yet, traditional data collection methods such as traffic surveys, sensor networks, and traffic cameras not only incur significant costs when deployed at scale but also often provide insufficient coverage. These methods typically yield either highly detailed temporal data for narrow segments (e.g., specific areas or intersections) or overly generalized, aggregated data for the whole system. To enable truly effective, data-driven decision-making, there is a critical need for cost-effective methods that can deliver detailed temporal insights spanning the full geographical extent of the mobility system.

2 Cellular Network Data

Network data—derived from the cellular network activity of millions of subscribers across extensive areas—can serve as a source of rich mobility information. This data captures the movements and behaviors of large groups, offering insights into the spatial and temporal aspects of population flows within and across regions. The cellular network data explored in this research was obtained from Telia’s Crowd Insights platform.

2.1 Methodology

To transform cellular signals into mobility data, a network operator collects temporal data generated during routine mobility subscriber activities like calls, texts, and movements within network coverage, anonymizing and aggregating this data to respect privacy and meet GDPR standards. Spatial data from network cell coverage is used to estimate geospatial positions, facilitating movement pattern analysis without traditional triangulation. Such a dataset typically excludes sensitive, roaming, and inactive subscriptions, focusing on active handset-linked subscriptions and extrapolating this information to reflect the broader population. Advanced algorithms distinguish between stationary and moving signals, classifying them as dwells and transits, thus providing a foundation for understanding wider mobility trends via reports such as [3]:

• Activity Reports which shed light on the locations, times, and origins or destinations of groups.

• Trip Reports that offer an origin-destination matrix to understand cross-country movements.

• Routing Reports which provide insights into the most likely travel routes taken, encompassing various modes of transportation.

2.2 Advantages & Challenges

As discussed, a serious challenge in analyzing mobility trends is the acquisition of high-resolution temporal data that also spans extensive geographical areas. Cellular network data, collected during routine telecom operations, provides a rich source of mobility information which offers valuable insights into both real-time and historical patterns of population movement. This information is invaluable for researchers, policymakers, and urban planners in understanding traffic flows and congestion, facilitating informed, data-driven decisions in urban and transportation planning.

However, leveraging cellular data for mobility analysis comes with some drawbacks and challenges, including variable spatial resolution from differing network coverages and the potential loss of granularity due to anonymization and aggregation. Data processing assumptions, such as stationary signals and proximity to the nearest cell, may not always reflect the complexities of urban mobility. Furthermore, policies excluding data from smaller groups to protect privacy can lead to an underrepresentation of certain areas or times. These limitations necessitate a careful evaluation of the data’s utility against the backdrop of privacy considerations and the inherent characteristics of cellular networks.

3 Routing Reports for Trondheim Municipality

Routing Reports analyze peopleFlow—journey patterns of groups of people—inferring the most probable travel routes to gain insights into travel behaviors, highlighting preferences in routes and transportation modes over various regions and times. This analysis incorporates multiple transportation modes—such as road, rail, ferry, and pedestrian pathways—leveraging OpenStreetMap for map** probable trip paths over 1km. The routing reports employ an open trip planner to ascertain the quickest route between two points, considering the operator’s network coverage and penalizing unnecessary mode switches.

Figure 1: Satellite map of ways covered by the routing reports.

3.1 Data Description

The routing report dataset used in this study spans from 17th January 2019 to 30th November 2023, covering 1744 dates, with 35 dates omitted due to data gaps or poor signal quality. The raw dataset is aggregated into hourly and daily intervals, resulting in 82,798,311 and 3,792,160 data points respectively. As shown in Figure 1, the dataset geographically covers Trondheim municipality in its entirety but also includes some sections of the neighbouring Malvik municipality due to the location of the Trondheim Airport east of the municipality limits. The dataset contains six (or seven) attributes for the daily (or hourly) temporal aggregations, as follows:

• Date & Hour: The date and hour the mobility data was recorded. The Hour attribute is only present in the hourly temporal aggregation.

• wayID: This is a unique OpenStreetMap identifier assigned to each road segment. A stretch of road, rail or ferry route is thus composed of several connected ways (segments) each encoded as a LineString object consisting of the longitude and latitude coordinates of its nodes. There are 2210 and 2212 unique wayID values in the hourly and daily aggregations, respectively.

• tagKey: This refers to the transportation mode, one of three options: road, rail, or ferry.

• tagValue: This is a breakdown of each tagKey into its subcategories as defined by OpenStreetMap, such as primary/trunk/secondary for roads and expressboat/cruise for ferry, amongst others.

• Municipality: The municipality the wayID falls under based on the geographical coordinates of its nodes, either Trondheim or Malvik.

• peopleFlow: The estimated number of people passing through a specific wayID at the daily or hourly aggregation.

3.2 Data Preprocessing, Feature Engineering & Graph Augmentation

The raw dataset was preprocessed to rectify missing tagValue entries in the (hourly and daily) aggregations, supplementing them with data from OpenStreetMap. To facilitate direct comparisons, the daily dataset was constrained to wayIDs consistent with those in the hourly data. The excluded wayIDs from the hourly dataset typically had such low peopleFlow volumes in the daily data that the hourly numbers would not meet the minimum threshold of five individuals. In addition, some discrepancies were identified in the daily data, which were corrected to align with the total hourly peopleFlow values for the respective days.

Subsequently, the dataset was enriched with temporal attributes by extracting metadata from the Date feature—including Day, Month, WeekNumber, Year, and HolidayName—for subsequent analysis. Historical weather data [4] was incorporated to assess the impact of weather conditions on mobility patterns. Additionally, population statistics for Trondheim and Malvik municipalities were sourced from Statistisk Sentralbryå [2] to examine the relationship between mobility patterns and demographic trends over time.

Routing reports enable the analysis of peopleFlow trends by selecting origin and destination coordinates and calculating routes within a mobility network represented as a MultiDiGraph—a graph that allows multiple directed edges between the same pair of nodes. However, the geospatial graph associated with the routing reports has discontinuities, preventing some nodes from being reachable. To address such gaps, a fully-connected graph from OpenStreetMap, covering road, rail, and ferry modes, was obtained and adapted to prioritize ways in the routing reports while ensuring all nodes are accessible. Subsequent trend analyses was then confined to ways documented in the routing reports and associated peopleFlow data.

Figure 2: Yearly breakdown of peopleFlow volumes from January 2019 to November 2023.

4 Analyses and Discussion

This section details our investigation into the dynamics of Trondheim’s mobility system, utilizing historical peopleFlow volumes obtained from the routing reports. Our analysis encompasses geospatial trends, temporal patterns, the impact of external factors, and some identified areas of interest in the city.

4.1 Geospatial Trends

Figure 2 presents the average (mean) peopleFlow volumes derived from routing reports between 7th January 2019 and 30th November 2023, juxtaposed with annual population data. This reveals a consistent increase in peopleFlow volumes, mirroring the average yearly population growth of approximately 3000 residents. The Covid-19 pandemic’s effect is pronounced, showing a dip and subsequent recovery in peopleFlow volumes from March 2020 to March 2021. Excluding this period, seasonal declines in peopleFlow are evident during Easter, Summer, Christmas, and New Year holidays. The rationale for using mean peopleFlow rather than total sums is due to the overlap** volumes across different routes. For instance, a single individual traversing a road segmented into three ways would be counted in the peopleFlow for each segment, thereby overestimating the total mobility if sums were used.

Figure 3 breaks down the annual peopleFlow volumes by municipality, offering insight into the mobility patterns to and from Trondheim. Notably, much of the data from Malvik municipality pertains to road and rail traffic on the E6 trunk highway east of Trondheim which extend to Trondheim Airport. This analysis reveals that Malvik’s traffic volumes generally remain lower than Trondheim’s, except during significant holiday seasons when there is increased travel in and out of Trondheim. Furthermore, analyzing the yearly volumes by transportation mode (tagKey) in Figure 4 highlights periods of increased volume in one mode compared to others. Such fluctuations may signal shifts in the transportation preferences of the city’s inhabitants, prompting further investigations.

Figure 3: Detailed overview of peopleFlow volumes across municipalities.

Figure 4: Detailed overview of peopleFlow volumes across transportation modes.

4.2 Temporal (Normalized) Patterns

The temporal analysis of the routing reports encompasses hourly, daily, and weekly levels across different transportation modes, focusing on data from 1st March 2022 to 30th November 2023, with the Covid-19 period analyzed separately in Section 4.3. Due to significant volume disparities among the transportation modes, normalization within the range $[0,1]$ was applied to each mode’s data for a fair comparison of flow trends. Figure 5 shows that the hourly patterns for rail and road traffic are similar, while ferry traffic tends to increase during road and rail’s off-peak hours. Similarly, Figure 6 reveals that ferry traffic peaks on Sundays, contrasting with the weekend decline in road and rail flows. Meanwhile, Figure 7 indicates a less pronounced correlation among the three modes, though road and rail still exhibit parallel trends. Notably, during summer holidays, while road and rail traffic decreases, ferry traffic remains consistent. The monthly trends, not shown here, offer a broader view which smoothens out the weekly fluctuations.

Figure 5: Normalized hourly variation of peopleFlow volumes across transportation modes.

Figure 6: Normalized daily variation of peopleFlow volumes across transportation modes.

Figure 7: Normalized weekly variation of peopleFlow volumes across transportation modes.

4.3 External Factors

This section delves into how external factors like the COVID-19 pandemic, weather conditions, and road attributes (speed limits and lane counts) affect peopleFlow volumes across Trondheim.

4.3.1 Covid-19 Pandemic

The pandemic’s impact on peopleFlow volumes is discernible across Figures 2, 3, and 4. The correlation between the Norwegian government’s Covid-19 measures (from 13th March 2020 to 1st March 2022) and peopleFlow variations is visible in Figure 8, which presents a 7-day rolling average of the peopleFlow volumes. The clear correlation between policy changes and mobility patterns justifies the exclusion of this timeframe from earlier analyses.

Figure 8: Effect of Covid-19 measures on peopleFlow volumes.

4.3.2 Weather Conditions

Analysis incorporating historical weather data shows a slight impact of weather on peopleFlow trends. For precipitation types, Figure 9 indicates that flow volumes are highest in clear weather, followed by rain, snow, and mixed rain/snow conditions. On the other hand, the lowest flows are associated with rainy, snowy, or cloudy weather conditions.

Figure 9: Variation of peopleFlow volumes with weather conditions.

4.3.3 Speed Limits & Lane Counts

OpenStreetMap data, linked via wayIDs from the routing reports, provided information on speed limits and lane counts for each way. The analysis, depicted in Figure 10, reveals a positive correlation between peopleFlow volumes and both higher speed limits and greater lane counts. This trend might be influenced by the routing algorithm’s preference for quicker routes. Notably, the effect of speed limits on peopleFlow plateaus beyond 80km/h due to the reduced number of routes supporting such speeds.

Figure 10: Combined effect of speed limits and lane count on peopleFlow volumes.

4.4 Targeted Areas of Interest

This section provides a temporal analysis of peopleFlow trends along specific routes and areas identified as relevant to Miljøpakken and MoST initiatives.

4.4.1 Comparison with Public Transit

Leveraging Automated Passenger Counting (APC) data from AtB, Trondheim’s public transport authority, enabled a comparison between public transit and overall peopleFlow volumes along identical routes. The AtB APC data—covering the period from 1st May 2020 to 30th November 2023—included 1,112,221 unique bus trips over 1,295 days, spanning 6 lines and 204 stops. For the analysis in Figure 11, both AtB and routing report volumes were smoothed using a 7-day rolling average and subsequently normalized to address the considerable scale differences between them. While a strong correlation exists between AtB and Telia volumes, the instances where AtB volumes surpass Telia volumes stand out and require further investigation, suggesting public transit might be experiencing traffic increases not reflected in the broader mobility system.

Figure 11: Comparison of public transit and peopleFlow volumes along selected bus routes.

4.4.2 Miljøpakken Bromstadruta Project

Figure 12: Bromstadruta cycle path (left) and Trondheim city center (right).

Miljøpakken is develo** a 3.2km cycle path and sidewalk project in Trondheim named Bromstadruta [1], depicted in Figure 12 (left) with the cycle path outlined in red and the corresponding routing report ways in blue. Such infrastructure projects stand to benefit from a thorough analysis of mobility patterns along the proposed route early in the planning phase. For example, Figure 13 shows normalized peopleFlow trends along Bromstadruta versus the city at large, uncovering unique dynamics that could be crucial for planning and achieving project goals.

Figure 13: Analysis of normalized peopleFlow trends along the planned Bromstadruta cycle path in Trondheim.

4.4.3 Trondheim City Center

A similar analysis for specific urban areas can be conducted, in this case focusing on the city center as shown in Figure 12 (right) with the target area highlighted in grey. Figure 14 reveals the city center’s pronounced daily (unnormalized) peopleFlow volumes compared to other areas. An in-depth examination of these central urban mobility patterns could provide valuable insights for optimizing traffic flow, managing congestion, and allocating public transportation resources more effectively.

Figure 14: Analysis of peopleFlow trends within Trondheim’s city center.

5 Conclusion and Future Work

This paper explores the utility of cellular network data to support efficient and sustainable mobility initiatives like Miljøpakken and MobilitetsLab Stor-Trondheim, offering a cost-effective, detailed, and wide-ranging source of mobility information. Through the analysis of Trondheim, Norway’s routing reports, the potential of such data for gaining insights into population flows within and across regions becomes evident, highlighting:

1. Geospatial trends of the mobility dynamics of specific areas and routes within and across municipalities.

2. Temporal patterns across different transportation modes from hourly to monthly scales.

3. The impact of external factors like pandemics, weather, public transit, and transportation infrastructure on mobility.

These insights are crucial for data-driven decision-making in urban planning and policy formulation. However, the inherent nature of cellular network signals and practical realities such as privacy concerns bake some assumptions into the mobility data which require careful consideration in its application. Furthermore, data preprocessing and other enhancements are required to fully capitalize on its potential.

This study is part of a larger initiative to develop a digital twin of Trondheim’s transportation infrastructure and mobility system, enriched with comprehensive data-driven historical insights and predictive modelling capabilities. Future research will focus on a thorough spatiotemporal analysis of the mobility network—extending to other modes of transport such as cycling and walking—to pinpoint key ways critical to the peopleFlow dynamics across the network.

Acknowledgements

This research received funding from the PERSEUS project, a European Union’s Horizon 2020 research and innovation program under the Marie Skłodowska-Curie grant agreement No. 101034240. The authors also acknowledge MobilitetsLab Stor-Trondheim (MoST) for their financial contribution and AtB for providing mobility data, which has been instrumental in our research.