Integrating future trends and uncertainties in urban mobility design via data-driven personas and scenarios

Abstract

Urban mobility contributes significantly to greenhouse gas emissions and comes with negative social impacts for various groups, such as limited accessibility to opportunity or basic services. Transitions towards sustainable and people-centred urban mobility systems are paramount. Yet, this is accompanied by various challenges. Complex urban systems are accompanied by high uncertainties (e.g., technological progress, demographics, climate change) which are currently not well integrated. Possible solutions originate from design, policymaking, and innovation, with a widespread disconnection due to non-compatible methods. This paper presents a method to improve the ability to design future urban mobility systems by integrating different approaches for modelling what the future could be and who could be the users. The research question is how diverse future user needs can be integrated in design processes for urban mobility systems. The proposed scenario-based design and personas allows to create data-driven proto-personas—a set of archetypical users with assigned characteristics and behaviours—test their validity, derive distributions across geographical areas, and transform them for different 2030 scenarios. This serves as input to create full personas and synthetic populations as intermediary design objects for the collaboration of designers and simulation experts. The methodology is exemplarily applied in the context of Paris. It contributes to urban mobility solution design that is more aware of future uncertainty and diverse needs of users, therefore, better capable to respond to today’s challenges. The approach is replicable with open data and accessible source code: https://github.com/TjarkGall/proto-persona-clustering.

1 Introduction

Urban mobility comes with multiple intertwined challenges, such as rising greenhouse gas (GHG) emissions, unequal access to opportunity, or limited rare materials [1], Climate [2,3,4,5]. The United Nations SDG 11.2 emphasises that ‘by 2030, [we must] provide access to safe, affordable, accessible and sustainable transport systems for all’ [6]. Banister [7] describes the stakes of sustainable urban mobility as more than just the movement of people and goods within cities, but also the interplay between transportation systems, land use patterns, environmental considerations, and social equity. Sustainable urban mobility aims to balance between meeting the diverse mobility needs of urban populations while minimising negative environmental impacts and fostering inclusive and equitable access to transportation.

Achieving this balance can be described as a wicked problem with competing interests, no clear solution, and high levels of uncertainty [8]. Potential solutions are ranging from technologies and urban plans to behavioural nudges and public policies. Planners and decision-makers use a range of methods to better understand the status quo, the interactions of different components, as well as testing the potential impacts of future solutions, such as new mobility services or public transport expansions. Agent-based simulations are increasingly common to incorporate complex characteristics and interactions between agents, such as congestion, shared mobility use, or rebound effects [9]. Framed in people-centred design, it allows detailed system modelling of impacts on people across time and space [10]. However, two gaps remain: (1) The integration of variations of the future (e.g., population dynamics) in today’s design practices, and (2) the disconnection between qualitative and quantitative design, including the transfer of outputs. The resulting research question for this paper is how diverse future user needs can be integrated in design processes for urban mobility systems?

To respond to this question, this article starts by introducing key conceptual components used throughout, namely future scenarios (Sect. 2.1), personas (2.2), and synthetic populations (2.3), followed by an overview of related works. In Sect. 3, the method is described and applied to Greater Paris, followed by discussion and conclusions (Sects. 4–5). The case of Paris is chosen due to the significant role metropolitan areas play in global sustainable urban mobility, the existence of related open-source mobility planning and simulation cases and methods this paper aims to strengthen, and the availability of open data to ensure replicability.

2 Designing future people-centred mobility solutions

This section introduces three existing components that are used in qualitative and quantitative methods to design for future uncertainties (scenarios) and model urban mobility users in predominantly qualitative (personas) and statistically representative (synthetic populations) manner. This section serves the dual purpose of outlining existing works in the field and identifying research gaps, as well as introducing the core components of the methodology described in the following chapter.

2.1 Future scenarios

Scenarios are a tool that permits working strategically with multiple futures. Scenarios are defined as having ‘a temporal property rooted in the future and reference external forces in that context […, to] be possible and plausible while taking the proper form of a story or narrative description; and […to] exist in sets that are systematically prepared to coexist as meaningful alternatives to one another’ ([11], p. 1). Explorative scenarios are a subset thereof showing different alternative futures to prepare for them today [12]. Future scenarios permit operationalising trends and uncertainties for design and decision-making [13]. Future scenarios have found ample application in urban mobility design (e.g., [14,15,16,17,18]. Most of them follow similar formats which can be simplified via scenario archetypes [15, 19, 20]. So far, few works combine quantitative simulations with explorative future scenarios. While Shaaban et al. [21] apply qualitative scenarios to conduct agent-based simulations, their application and tools are different as the focus is on agri-food systems. Agent-based simulations and scenarios are often mentioned together. However, scenarios are mostly used to define variations of mobility service or, e.g., modal shares, rather than being considered as distinct, possible alternative solutions for the futures. An exception is the paper by Schlenther et al. [22] on simulation-based investigation of transport scenarios for Hamburg. In this case, pre-existing scenarios are compared, focusing on changes of schedule, new mobility services, or, e.g., increased parking pressure simulated via restricted parking plots. Other promising approaches, such as a simulation approach for different potential future street layouts are evolving but with a focus on the spatial dimension [23]. The authors are not aware of an urban mobility simulation where the future population is adapted more than by scaling the population by the average annual growth rate.

2.2 Personas

Personas are a method used in people-centred design that allows considering diversity of people, needs, and lifestyles. They are fictitious characters that represent a homogenous class of users [24]. Future personas allow representing users in a prospective mode and are characterised by (1) a user model, (2) a communication tool; (3) a decision aid and prospective tool [25]. Usually, a persona is given a name, a visualisation, and a narrative that provides input on attitudes and behavioural traits [26]. This allows the persona to come to life and become embedded in the designer’s creative process [24, 27]. Methods to create personas can be grouped in three approaches [28]. First, fictional elements can be completed with qualitative data. For example, Goodman-Deane et al. [29] created personas by combining data from questionnaire and interviews. Second, personas can be created from fictional elements [30]. A mixed-method approach is used to create tangible futures and match qualitative mobility user personas with scenarios. Third, recent approaches are using big data leading to data-driven personas (see persona generator, [31]).

Personas create a foundation for discussion between designers and clients and prime them for deeper immersion in varying user profiles that differ from their own socio-economic profile, contributing to more inclusive design [27]. When designing mobility solutions, personas were used to feed the emotional design of autonomous shuttles [32] or the design of a ride-sharing service [33].

2.3 Synthetic populations

Different to personas, synthetic populations (SPs) aim to statistically reflect a real population [34, 35]. SPs represent reality via individuals, often grouped into households. Sociodemographic attributes are assigned to the individuals. Various approaches have been created to generate these. Some are based on statistical fitting algorithms [36, 37] and generate individuals so that specific attributes such as age or income classes are distributed in line with reference distributions to preserve spatial heterogeneity. Others use disaggregated, sparse data sets such as household travel surveys. These can be described, for example, by using Bayesian networks [38] or Hidden Markov Models [39]. Further, recent approaches join multiple sources of information through machine learning and deep generative modelling [40, 41].

SPs can be useful to model mobility patterns, e.g., to test potential future services or impacts of policy decisions by modifying the schedules of the agents [42], estimating changing GHG and noise emissions [43], or assessing policies [44].

2.4 Existing works

Some works exist that target structured design or simulation of future urban mobility users. For example, Al Maghraoui et al. [45] developed a qualitative approach to model the experience of mobility users. On the other hand, Kamel et al. [46] and Vosooghi et al. [47] created quantitative approaches to measure and integrate preferences, values, or other individual attitudes into data-driven approaches in the mobility context. Conceptual links between personas and synthetic populations have been identified in the past [48] but have not been applied jointly in the context of urban mobility to the authors’ knowledge. Building on the existing methodological components and the existing scientific contributions, some promising gaps have been identified. First, most qualitative design approaches are not made to adequately integrate quantitative characteristics. Additionally, many data-driven approaches lack systemic approaches to integrate qualitative trends due to the need to transform the information. Further, a challenge shared across persona and SP-based approaches, only few works have attempted to integrate future changes of users and populations into the approaches. These approaches are either addressing one specific dimension (e.g., population growth) or scaling up the present in some way. The authors are not aware of any work that integrates differences across different user groups in simulation. This paper proposes a mixed-method approach for urban mobility solution design incorporating future uncertainty, constituted of a data-driven persona generation process and a method to integrate qualitative attributes’ variation of agents into SPs.

3 Persona clustering and reweighting method

This section introduces the case study, describes the proposed method, followed by a description of data, and the application of the steps.

3.1 Case study of Paris-Saclay

The method is applied to the Communauté d’Agglomeration Paris-Saclay (CPS), an inter-council partnership of 27 municipalities (Fig. 1) south of Paris. It contains an urban hub (Massy), two regional train lines, a planned new metro, and two plateaus, with a linear sub-urban development along one of the regional railways. It borders the Metropolis of Greater Paris (MGP; blue in Fig. 1). The functional mix consists of agricultural uses, low-density mixed-used developments, as well as science and technology clusters. The focus lies on mobility design of CPS for the medium-term future of 2030. However, as a significant relationship exists with Paris, the MGP, and the surrounding areas, such mobility studies should be conducted on the regional scale. The underlying data is the recent 2019 census [49] with information on individual and household level and the national mobility survey [50] from 2018 to 2019.

Fig. 1

Four geographical areas used to categorise census data

3.2 Description of 10-step methodology

The method starts with building and validating a set of proto-personas (PPs) based on census data and mobility surveys for the present (Step 1–5). PPs are defined as a pre-stage for personas. They contain a set of quantitative information that are archetypical for mobility behaviours of a part of the population [51]. In the process, PPs are created independently of age and area to create personas that can be used across age groups and living locations. Steps 6–8 compile trends and uncertainties and match them with scenarios, followed by a proportional redistribution of PPs (Step 9). Lastly, the focus is on transforming the generated data into design objects for qualitative and quantitative approaches (Step 10). Figure 2 shows the steps, including the methods and tools used for each step, and the input data. The core steps of the method are fully reproducible with open data and a public repository.¹ In the next part, each step is described in detail. Intermediary results are shown throughout, while final outputs are in the results section.

Fig. 2

Combined development process of data-driven personas and synthetic populations for future scenarios

3.2.1 Step 1: Choosing cluster attributes

Three categories of census data attributes are selected from the available ones: (1) Attributes for clustering (e.g., household size), (2) Attributes for categorisation (i.e., age, area), (3) Attributes to supplement PPs. For the selection of variables, three considerations are important. Some variables can be categorised based on literature or inherent requirements of the process. Second, subjective choices inform the clustering in the context of mobility behaviours and preferences. Third, a verification for partial dependence of several variables supports the choice of most suitable variables. 25 variables are the basis for the following data preparation and clustering.²

3.2.2 Step 2: Data preparation

The selected attributes are organised, their data types transformed, and individual entries scaled stochastically by their statistical weight to a total of nearly 14 M individuals. Attributes that are not considered as relevant are deleted. To account for the geographical focus and mobility flows between the focus area CPS and its surroundings, geographical codes are assigned for each record. To avoid overrepresentation and complexification, a statistical analysis is performed to test for correlations between numerical variables aiming to reduce the number of attributes if some of them are significantly correlated. Statistically significant (p < 0.05) and high degree correlations exist between the number of children under 17 (NE17FR) on the one side and children under 5 (NE5FR) as well as household size (NPRR) on the other. Consequently, NE17FR is excluded from the clustering. For the remaining combinations, no variable can be identified which explains another variable strongly enough that exclusion would be justified.

3.2.3 Step 3: Proto-persona (PP) clustering

Next, census entries are clustered based on selected attributes. The mix of categorical (e.g., socio-professional category) and numerical info (e.g., number of cars) excludes common clustering approaches. The k-prototype method is used which was developed to cluster including categorical data [52]. The method searches most suitable central values for a set number of clusters. These central values are adapted over n iterations to improve the model fit. The quality of the outcome can be defined by the distance between the values assigned to each cluster and the central values. For example, if the number of cars per household of one cluster is one but some data entries with two cars are still assigned to the cluster due to a high fit in other categories, this difference is defined as distance to be minimised.

To cluster, the number of clusters is defined. This can either be a subjective choice or the result of a preceding test. For the latter, a scree-test can be performed [52]. The objective function is made up from the ‘within-ness’,the ‘sum over all clusters within distances to the prototypes for each cluster [52], p. 202). By running the clustering process for a variety of cluster sizes (k = 2, 4, […] 50) with 100 iterations, the scores can be compared. This process is conducted with a 5% sample of the overall scaled population data (n = 683 k). While more clusters are usually producing more accurate results, the goal is finding a balance between a number as low and an accuracy as high as possible. For this, the elbow method can be applied in which the number of clusters is set at the point where the curve becomes less steep (Fig. 3). Here, 16 clusters are chosen.

Fig. 3

Scree test graph showing total within-ness (y-axis) for different number of clusters (x-axis)

Next, all remaining variables are used to cluster the scaled population (n = 13,658,311). This is done with 100 iterations and including non-applicable variable entries (e.g., no indication where job is located as person goes to school). This results in the distribution of entries by cluster shown in Fig. 4. The individuals are grouped in clusters of relatively homogenous characters, between 400 k and 1.2 M people per cluster.

Fig. 4

Distribution of entries by cluster

The clusters represent PPs. A selection of the central values of the cluster outcomes are shown in Table 1. The Table 1 shows the simplified meaning behind the majority of attributes that have been used for clustering, including information concerning the household (e.g., number of children and household members), individual socio-demographics (e.g., degree, activity location, origin), as well as supplementary information (e.g., parking spot availability or housing type). To test the accuracy of the clusters, the most frequent clusters are chosen to compare its cluster central values and values of all entries assigned to the cluster. The numerical variables show a high accuracy across the variables. For the categorical data, some have very high accuracies and indicate that nearly 100% of data entries share the central value of the PP. In others, values under 50% appear. However, these are primarily in variables with many categories and all values are sufficiently high to evaluate the cluster results as solid and, with the limitation inherent to clustering over 13 million entries into 16 PPs, representative.

Table 1 Selected central values for 16 clusters / proto-personas

These 16 clusters represent groups of people with shared mix of attributes linked to the way of life and mobility behaviours. As an example, PP16 (n = 736 k) is a widowed woman living alone in a 3-room, 40–60 sqm apartment in Paris. She owns her apartment in a multi-unit building but does not own a car. She holds a vocational degree and is retired. Originally from another metropolitan region in France, she moved to Paris. This sample PP shows that with the performed process, comprehensible descriptive profiles can be generated.³

3.2.4 Step 4: Proto-personas per age group and area

The clusters represent PPs which do not appear equally often across age groups and area. The objective was developing personas independently of age and area to focus on socio-economic characteristics and introduce these values via scenario variation. However, these variables can be re-associated and analysed. The distribution is organised by cluster, age group (0–14 years, 15–29 years, 30–44 years, 45–59 years, 60 + years) and geographical areas (Fig. 5). This provides an idea in how far all personas can describe sufficient varying classes and represent groups such as children. For example, for the age group of 0 to 14 years, ten clusters have occurrences, with PP8 being the strongest with occurrences between 3 and 5%. While less personas appear in the under 15 and over 60 years group, sufficient PPs are included in both to represent a diversity across the age group.

Fig. 5

Cluster occurrences per area and age group in percentage (100% for each area, showing values from 0.5% upwards)

3.2.5 Step 5: PP validity test for mobility

A challenging element in persona development is validating their accuracy. Even when working with a PP without qualitative attributes, it must be validated that choices of attributes and clustering approach can lead to personas that represent variations in urban mobility. Two approaches can be used for this. First, a subjective visual approach allows to validate the representativeness of the underlying data. If considered relevant, specific personas can be added, for example, underrepresented personas such as people with reduced mobility. Second, mobility surveys can be used to test if statistically significant differences in mobility behaviours across the PPs can be identified. In this paper, this is tested with the mobility of people survey [50] which details what mobility-related resources are available (e.g., driving license, car) and what trips are done where and when on a given day. All entries are considered where both origin and destination of trips were inside the Île-de-France region. The minimum, maximum, sum, and average values for variables (aggregated from 0 to 8 trips) are calculated. Table 2 shows the input variables, their description, and the performed calculations. For this paper, three variables are chosen.

Table 2 Information from mobility survey [50]

To match mobility data with the PP, shared variables are used.⁴ The results show that differences between categories are in line with expectations. To exemplify this analysis, a random data entry that is matched to PP16 is chosen. A selection of the assigned information is: 1 person, female, single, retired, 1 car, 2 trips, first starting at 11:45:00, average duration 15 min, with a two-hour activity in between, and a 1 km distance. This fits the prior description of PP16. After repeating this process multiple times, it can be concluded that the PPs describe successfully varying mobility behaviours. However, the matching process on few shared variables, as well as the exclusion of under 16-year-olds due to lack of data bear potential for future improvements.

3.2.6 Step 6: Uncertainty/trend compilation

The last steps described preparatory steps with today’s data. To integrate possible futures via scenarios, uncertainties and trends can be used to redistribute PPs. Trends are developments that are assumed to materialise. For example, a continuing ageing of the society, a stable male/female distribution, or a population increase for an area. Uncertainties are contextually relevant attributes that are assumed to change but for which it is impossible to be certain of the direction of change. This might be the level of private car ownership or the willingness to take public transport. Each of them can vary. For example, the car ownership might range between 0.8 and 1 per person based on socio-economic and political developments (currently 1 on average). Some key trends (Table 3) and critical uncertainties (Table 4) are chosen for this study.

Table 3 Compilation of trends

Table 4 Compilation of uncertainties

3.2.7 Step 7: Future trend integration

These key trends are integrated depending on PP, age-group, and timescale. An anticipated annual population growth rate can be applied to each of the PPs via their respective occurrence rate (weight) that results from the clustering. On the other hand, an ageing society would lead to a proportional shifting of PPs by age-group. Based on the population and ageing forecasts from the statistical service of France [53], growth rates (Table 5) are calculated which are equally applied to all PP within one area. The growth rate is calculated as combined value of both trends.

Table 5 Annual growth rates in percent for each category, based on [53, 56] (Attention: Values do not refer to Ile-de-France but IDF without Paris, MGP and CPS)

3.2.8 Step 8: Uncertainties’ values per scenario

The key difference to step 7 is that uncertainties are not applied equally but dependent on an underlying scenario. A number of different possible scenarios (usually 4–7) is defined for a particular year in the future, in this case 2030 (Table 6). Each scenario can be complemented with a name, narrative, and set of associated, application-dependent information. For the use case of this paper, three uncertainties and a numeric adaptation for each is defined. For their adaptation, scenario archetypes can be used [19, 54], in particular those of the mobility sector [15]. Building on these scenarios, values for each uncertainty are set to create coherent scenarios and a wider distribution across the variables (Table 7).

Table 6 Four archetypical scenarios [15]

Table 7 Trends and uncertainties quantified and adapted by authors for each scenario

3.2.9 Step 9: Redistribute PP by scenario

Three options exist to adapt the PP for each future scenario. The distribution of the PP can be adapted to best match the values provided for the uncertainties. Second and if necessary, certain attributes of the present PP can be adapted to better allow the matching of future uncertainties. Third, some or all of PPs could be adapted differently for the scenarios. The latter would lead to a much higher number of personas, defeating its purpose. The focus of this paper is on the former option as it is deemed the optimal case for the methodology. Two steps need to be conducted. First, the growth and demographic change scales the overall distribution. The growth rates (Table 5) are applied to scale the PPs without changing the proportion. This constitutes the basis for all scenarios. Next, changes of certain impacts shall be reflected as accurate as possible by changing the distribution of personas. This is done by optimising the initial population weights so that uncertainty-based target values are reached in the future scenarios. The target scenarios are characterised by target mean values of numerical attributes across the overall population and anticipated target shares of persons with specific attribute values for nominal attributes. The following formula outlines the inputs and basis for the fitting process:

Persona variables and constants

• $K\in {\mathbb{N}}:\mathrm{Number}\,\mathrm{of}\,\mathrm{personas}$

• ${{\omega }{\prime}}_{k}\in \left({0,1}\right):\, \mathrm{Initial}\,\mathrm{population}\,\mathrm{share}\,\mathrm{per}\,\mathrm{persona }k \left(\mathrm{baseline}\right)$

• ${\omega }_{k}\in \left({0,1}\right):\, \mathrm{Population}\,\mathrm{share}\,\mathrm{per}\,\mathrm{persona}\,\mathrm{k } (\mathrm{to}\,\mathrm{be}\,\mathrm{chosen})$

Attribute constants

• $\mathcal{A}:\, \mathrm{Set}\,\mathrm{ of}\,\mathrm{attributes}\,\mathrm{with}\,\mathrm{ a}\,\mathrm{mean}\,\mathrm{target}\,\mathrm{ value}$

• ${a}_{k}\in {\mathbb{R}}:\mathrm{Value}\,\mathrm{of}\,\mathrm{attribute}\,\mathrm{A}\,\mathrm{in}\,\mathrm{persona}\,\mathrm{k}$

• $\widehat{A}\in {\mathbb{R}}: \mathrm{Target}\,\mathrm{mean}\,\mathrm{value}\,\mathrm{for}\,\mathrm{attribute}\,\mathrm{A}$

• $\mathcal{B}:\, \mathrm{Set}\,\mathrm{of}\,\mathrm{attributes}\,\mathrm{with}\,\mathrm{a}\,\mathrm{target}\,\mathrm{share}\,\mathrm{per}\,\mathrm{value}$

• ${\mathcal{V}}_{B}: \mathrm{Set}\,\mathrm{ of}\,\mathrm{values}\,\mathrm{that}\,\mathrm{are}\,\mathrm{permissible}\,\mathrm{ for}\, \mathrm{attribute}\,\mathrm{ B}$

• ${b}_{k,v}\in \left\{{0,1}\right\}:\,\mathrm{Defines}\,\mathrm{whether}\,\mathrm{attribute }\,\mathrm{B} \,\mathrm{in}\, \mathrm{persona} \,\mathrm{k}\, \mathrm{has}\,\mathrm{ value}\,\mathrm{ v}$

• ${\widehat{B}}_{v}\in \left\{{0,1}\right\}:\, \mathrm{ Target}\, \mathrm{share}\,\mathrm{for}\,\mathrm{value}\,\mathrm{v}\,\mathrm{of}\,\mathrm{attribute}\,\mathrm{B}$

The target share values need to fulfil:

$$\sum_{{v\in \mathcal{V}}_{B}}{B}_{v} \le 1. \forall B \in B$$

If target shares for all possible values are given, equality must hold.

Optimisation constants

• ${\xi} \mathrm{AB}:\, \mathrm{Objective}\,\mathrm{ weight}\,\mathrm{for}\,\mathrm{attribute }\, AB$

• $\alpha \boldsymbol{ }:\mathrm{Regularisation}\,\mathrm{weight}$

Problem formulation

$$\begin{aligned} & \mathop {{\text{min}}}\limits_{{\omega_{k} }} \mathop \sum \limits_{{A \in {\mathcal{A}}}} \xi_{A } \left( {\hat{A} - \mathop \sum \limits_{k} \omega_{k} a_{k} } \right)^{2} \\ & \quad + \mathop \sum \limits_{{B \in {\mathcal{B}}}} \mathop \sum \limits_{{v \in V_{B} }} \xi_{B } \left( {\hat{B}_{v} - \mathop \sum \limits_{k} \omega_{k} b_{k,v} } \right)^{2} \\ & \quad + \alpha \mathop \sum \limits_{k} \left( {\omega_{k} - \omega_{k}^{\prime } } \right)^{2} \\ & {\text{s.t.}}\quad \sum \omega_{k} > 0 \\ \end{aligned}$$

(1)

The first term in the objective calculates the mean of an attribute a given the updated weights and compares it to the required target value. Similarly, for nominal attribute values where a target value has been defined, the second term compares the obtained share of all PPs that have this specific value with the requested share. Each objective (representing an attribute or an attribute value) can be prioritised through a weighting factor. For attributes with higher attribute weights, matching their target value will be enforced with higher importance than for other attributes. Finally, a regularisation term is introduced which allows to give importance to not deviating too strongly from the initial weights of each persona. The regularisation weight defines whether the weights will be adapted in an aggressive way (increasing few personas strongly) or more uniform (increasing many personas carefully), but also has an impact on how well the posed target values can be attained. In the present use case, the number of cars has a weight of 0.4, household size 0.3, work/study location 0.3 and a weight of 0.1 was given to the regularisation term to keep the variation between initial and final weights as small as possible. The outcomes are provided in the results section.

3.2.10 Step 10: Create personas and synthetic populations (SPs)

Finally, present and future sets of PPs and their distributions are used to create full personas as intermediary design objects. Simultaneously, PPs distribution by area, in combination with travel surveys, and registries of residential, work, commercial, and leisure activities can be used to generate one present, and a set of future SPs. Building on the additional information assigned from the mobility survey to the personas (e.g., most common activity chains), qualitative persona creation methods allow building a set of personas (incl., for example, name, photo, location [30]).⁵ This is described by way of example in the results section. The difference to other processes is that the starting point (age, characteristics, location, activity chain) is not defined by the designer but instead comes from the underlying evidence-based approach and the supplementary information. Further, building on Vallet et al. [48], we can use the generated distributions and profiles for the generation of multiple synthetic populations. The only component that needs to be changed in the synthesis process are the initial weights in the census information. As, by using the clustering algorithm, a persona can be assigned to each record in the census data, their weights can be adapted according to the reweighting process. Applying this technique and analysing simulation outcomes based on these future SPs has been conducted separately in parallel research.

4 Scenario-specific sets of personas and synthetic populations

The first output is a quantitative distribution of PPs across scenarios. This data set remains linked to the initial census data and can aid the process of scaling up and redistributing the census entries to result in new quantitative insights on varying possible futures, as well as feed quantitative design approaches. Figure 6 shows the distributions in Paris and CPS.⁶ The distribution for the 2019 base and the four archetypical scenarios is provided. We can analyse two information from the figure. First, all four scenarios have a higher population, clearly visible for Paris and slightly for CPS. Second, each PP is represented by a colour and changes its occurrence from one scenario to another.

Fig. 6

Resulting distribution of PP (in 1000 s) for 2019 (census base year) and 2030 scenarios using outputs from reweighting and population growth rates per area

The information from the PPs, enriched with information from the mobility survey, can feed the creation process of data-driven qualitative personas. An example persona for the PP10 (Annex, Fig. 7) has been created by adapting an existing persona template [55].

Fig. 7

Exemplary persona card for PP16. Everything written in italic is coming from underlying data. All other were deduced from it when and where possible. Persona format adapted from template by xtensio [55]

However, some limitations prevail. This paper focuses on the individual. Some data exists on household level and could find integration in the process. Inherent to working with futures, subjective inputs were used. While this can be supported via literature and expert validation, accuracy remains an uncertainty. The authors argue that the chance to be closer to the actual future is higher when working with several futures compared to working with a singular one or simply use the assumptions of present users and population as done predominantly in previous works.

5 Discussion and conclusion

This paper asked how diverse future user needs can be integrated into design processes for urban mobility systems. Future scenarios and different ways of modelling users were introduced that can support service design and simulation. Next, a mixed-method approach was proposed using existing qualitative (i.e., persona creation, scenario planning) and quantitative (i.e., k-proto clustering and agent-based simulation) methods to inform today's design decisions of urban mobility systems. Using these methods permits to (1) address pre-existing weaknesses (e.g., dominating qualitative approaches for persona creation, lack of user uncertainty integration in simulation), (2) build a basis for interdisciplinary collaboration between designers and data scientists, and (3) propose new components that can be used to adapt personas and SPs in a structured manner to future scenarios.

The method was used to generate validated data-driven proto-personas and reweigh them across future scenarios to allow the integration of trends and uncertainties in qualitative design and quantitative simulation. This results in a set of personas and SPs per scenario that enables the use of established people-centred design processes (i.e., personas) in a future context. In parallel, the proposed method results in a set of SPs that feed agent-based simulations of future scenarios. Aside from its field-specific contributions, the combined approach enables designers and data scientists to collaborate via shared and aligned artefacts.

5.1 Discussion

Existing works discussed in this article used extensively personas, SPs, and scenarios. Conceptualised by Grudin and Pruitt [27], personas have found ample application in mobility service design (e.g., [32, 33]. Vallet et al. [30] introduced future personas to consider future mobility user needs, while Al Maghraoui et al. [45] proposed methods to integrate future user experiences. While qualitative approaches dominate persona creation, Stevenson and Mattson [31] presented a quantitative method to generate personas.

SPs have found significant attention in mobility simulation and design, including replicable generation thereof [42] and various use cases in agent-based simulation (e.g., [43, 44]. Kamel et al. (2019) and Vosooghi et al. (2019) set a foundation for quantitative integration of diverse users, while Vallet et al. [48] proposed a conceptual framing of personas and SPs to advance people-centred urban mobility system design.

While each can play a crucial role in designing sustainable and people-centred urban mobility futures, the use of the analysed approaches is limited to specific fields and so far without consideration of multiple futures. Working with such scenarios can be critical to addressing rising uncertainty originating, among others, from the pace of technological advances, the climate crisis, and behavioural evolutions.

This paper proposes a method that uses these already established concepts to advance the work of designing with a people-centred perspective for multiple futures while permitting the collaboration of experts from different disciplines. It provides a novel generation approach for data-driven and realistic personas of today and tomorrow in the urban mobility context and responds to the lack of adapted future simulations. The outputs of this method have been used successfully to generate alternative SPs, test potential mobility solutions across future scenarios, and analyse impacts across PP groups in different geographical contexts.

Aside from its contributions, a range of limitations persist: Most notably the qualitative approach for defining future trends and uncertainties parameters. However, this challenge is inherent to working with futures and can be partially mitigated by basing assumptions on expert knowledge and existing research. Otherwise, a set of methodological choices has been made throughout the process. While each was made to respond to the identified challenges and needs, other methods exist which might be more adequate and should be tested as well.

5.2 Conclusion

In conclusion, the wicked problems of planning sustainable and people-centred urban mobility solutions today require a detailed consideration of the future impacts of socio-demographic and technological changes on solutions, users, and associated needs. The presented methodology comes with limitations inherent to its ambition and the underlying concepts. However, the authors argue that it makes a relevant contribution to people-centred urban mobility solution design and simulation processes, as well as creating artefacts for structured exchange between experts from qualitative and quantitative disciplines. The existing uptake across different use cases with different profiles of collaborators so far has shown its potential utility. The continuous extension and improvement, as well as wider testing and validation are nevertheless seen as crucial. This publication, supported by the use of open data and provided open-source code shall enable this process and ultimately permit designing more sustainable and people-centred solutions for the urban populations and mobility users of tomorrow.

Annex

See Tables

Table 8 Census information chosen for clustering (translated by first author)

8,

Table 9 Codes of used variables.

9,

Table 10 Descriptive statistics for numeric variables

10,

Table 11 Descriptive statistics for categorical data after population scaling, part 1

11,

Table 12 Descriptive statistics for categorical data after population scaling, part 2

12, and

Table 13 Initial, target, and actual proportional distributions of PP across scenarios

13; Fig. 7.