An Intelligent Transportation System to control air pollution and road traffic in cities integrating CEP and Colored Petri Nets

The increase in vehicles in road traffic is a characteristic phenomenon of today’s world, which means that related problems such as traffic accidents, pollution (air and noise), long travel times, etc., are increasing in the same way. Numerous reports have been published in order to determine the extent of the problem, and in particular this has led to the development of a new area of study, Intelligent Transportation Systems (ITS) [1]. ITS has emerged as an important element for both improving human life and the modern economy [2], with the main objective of optimizing road traffic by managing the capacity of the roads, improving driver safety, reducing energy consumption and improving the quality of the environment, among many others things. Moreover, an increase is expected in the development of ITS, integrating concepts such as big data, thus generating the new concept of Internet of Vehicles (IoV), as Xu et al. propose in [3], where a survey of applications of IoV and big data in autonomous vehicles is presented.

A key component for the study and development of ITS is traffic modeling, which provides a framework to better investigate and test the state of the road in real time and accurately predict future traffic. In general, a desirable model must meet the following requirements:In this context, we focus on traffic control in cities, taking into account the levels of environmental pollution according to the air quality levels accepted by the international recommendations [4]. Thus, we are tackling a common problem in large cities, where traffic restrictions must be applied due to pollution.

The methodology we use to design ITS is Complex Event Processing (CEP) [5] in combination with formal methods to model and test the proposed solutions [6]. CEP provides users with facilities for analyzing and correlating large volumes of data in the form of events with the aim of detecting relevant or critical situations for a particular domain in real time. To meet this objective, the conditions describing the situations of interest to be detected must be specified as event patterns. Patterns are implemented by using the languages provided by CEP engines, the so-called Event Processing Languages (EPLs), and once the patterns are defined, they can be deployed in the CEP engine in question [7].

Additionally, Petri Nets (PNs) [8] are a formalism which provides mathematical rigor and a graphical representation of the model, offering a better comprehension from a visual model and a mathematical underlying model in order to obtain important results about all its possible behaviors. Furthermore, Petri Nets are supported by tools, which allow us to simulate and analyze the behavior of a given system in a suitable manner.

The main aim of our proposal is to combine the use of CEP and Petri Nets to model and test ITS and, specifically, the city traffic flow, taking into account air pollution conditions. Thus, the contributions of this study are:The structure of the paper is as follows. Section 2 presents the motivation of this work. Section 3 provides an overview of the CEP technology and the specific model of Petri Nets we use: CPNs. The city road model using CPNs is presented in Sect. 4 and the air quality and road traffic event patterns in Sect. 5. Section 6 presents the whole ITS system, integrating both the CPN model and the EPL patterns. This model is then applied to a real case study in Sect. 7, taking as reference the division into districts of Madrid, the capital city of Spain. Section 8 presents the related works, and finally, Sect. 9 presents our conclusions and lines of future work.

In urban environments, road traffic can be a significant environmental problem due to the disproportionate exposure of citizens to environmental toxics, thus representing a public health problem. Air pollution remains one of the main factors related to preventable diseases and premature mortality in the EU. In 2010, it was estimated that air pollution in the EU caused more than 400,000 premature deaths. It was also the cause of preventable diseases, including respiratory conditions such as asthma, and exacerbated cardiovascular problems [9].

The greatest impact on human health occurs in urban areas, where air pollution levels are highest. Of particular concern is the health impact of exposure to atmospheric particulate matter of 2.5 micrometers (PM\(_{2.5}\)) and ozone (O\(_3\)). However, nitrogen dioxide (NO\(_2\)) and sulfur dioxide (SO\(_2\)) are also a concern, both on their own and as ozone precursors. Across Europe, it is estimated that 20–30% of the urban population is exposed to PM\(_{2.5}\) with levels above the EU reference values, and 91–96% are exposed to more stringent levels than those of the World Health Organization [9]. At international level, the Organization for Economic Cooperation and Development (OECD) states that: “Unless we clean the air, by the middle of the century one person will die prematurely every 5 s from outdoor air pollution” [10].

According to [9], air pollutants can be classified as primary (emitted directly into the atmosphere) or secondary (formed in the atmosphere of precursor pollutants). The main primary air pollutants include primary PM, BC, sulfur oxides SO\(_x\), NO\(_x\) (which includes NO and NO\(_2\)), NH\(_3\), CO, methane (CH\(_4\)), benzopyrene (BaP) and hydrocarbons. Secondary air pollutants include secondary PM, O\(_3\) and NO\(_2\). The AQI index [11] reports daily air quality on the basis of five of these major pollutants.

Air pollutants may have a natural, anthropogenic or mixed origin, depending on their sources or the sources of their precursors. Emissions from motor vehicles contribute to air pollution in urban areas, and in many cities ensuring adequate air quality is a major problem. Road transport is the main source of air pollution in urban areas, and, therefore, there is a growing need to control current and future flow emissions as accurately as possible. As a result, a series of emissions models and emission factor databases have recently been developed. For instance, Yuan et al. [12] describe the complex way in which air pollution dispersion occurs in high density cities, such as Hong Kong.

Moreover, there is a relationship between the state of the traffic and the level of emission of pollutants. Borge et al. [13] presented a detailed study for the city of Madrid (Spain). This study was conducted by analyzing hourly emissions from nearly 15,000 road segments distributed in 9 management areas covering Madrid City and surroundings. Traffic status was evaluated in four levels: free flow, heavy, saturated and stop and go. Significant quantitative information can be derived from this work, such as the relationship between the average speed vs the speed limit in order to establish the traffic level. Table 1 contains the traffic level ratios for a trunk road/primary city proposed in Borge et al.’s work.

This section explains the background for both the CEP technology used for defining air quality and road traffic event patterns, and the CPN formalism.

Following the example depicted in Fig. 4, a city will be divided into zones, which are represented by places, labeled with the zone names: Zonei, \(i=1,2,\ldots ,n\), where n is the number of zones. Our goal is to obtain different routes to travel from Zonea to Zoneb. Transitions will represent movements from one zone to another by traversing some streets. Thus, transition tritoj captures the movement from Zonei to Zonej. These transitions have a Boolean guard, which will not allow the same zone to be traversed again to avoid cycles.

As an illustration, we show the main page of a simple city map CPN model in Fig. 5, which is an extension of the city map shown in Fig. 4, by allowing the reverse movements as well. The color set associated with places Zonei is defined as follows:

which consists of a Cartesian product, where the first component corresponds to the target zone, the second component represents the route followed by the token to reach this place, as a list of traversed zones, from its starting zone to the current one, and the third component is the time spent using this route, expressed in minutes. As an example, see the marking of Zone1 in Fig. 5: (6,[1],0). This represents a car traveling from Zone1 to Zone6, which is currently in Zone1, so no time has been spent yet. Variables x, y labeling the PT-arcs are of color set \({ Zo}\), so they have the three components described above. TP-arcs are labeled with expressions that allow us to add the new step in the followed route and increase the time spent by the corresponding amount. Notice the transition guards transit(x,i), which avoids our crossing the same zone twice and only allows a movement when a car has not reached its destination. Thus, transit(x,i) is only true when a car in Zone x can move to Zone i. This guard is defined as follows:

where function noReturn is used to avoid cycles and isDest checks whether the car has reached its destination. These functions are defined as follows:

The operator ”#i x” in both functions extracts the ith element of tuple x, that is, given the tuple (a,b,c), #1(a,b,c) evaluates to a. Function mem(l,i) is a Boolean function that checks whether element i is in the list l. Thus, noReturn checks whether i is in the second field of x, which contains the list of traversed zones. Function isDest reverses the list (function rev) and checks whether the head (function hd) is equal to the destination (first field of x).

For instance, let us consider the marking 1‘(6,[1,2],3) in Zone2, which captures a car traveling from Zone1 to Zone6, currently positioned in Zone2 and having spent 3 min. To continue, two transitions could then be considered: tr2to1 and tr2to3. However, transition tr2to1 cannot be fired, since Zone1 already belongs to the route list. Instead, transition tr2to3 is enabled, since 3 does not belong to the list and the car has not reached its destination. Firing tr2to3 removes the token (6,[1,2],3) from Zone2 and generates a new token (6,[1,2,3],7) on Zone3 (assuming \(c23=4\)). Notice the fusion tags on the left corner of places Zonei, which identify the shared places (fusion places) with the Destination CPN page.

Figure 6 depicts the Destination CPN page that completes the model, where the main element is place Destination. This is a sink place, where tokens come to finish the routes. It captures the information about the route followed and the total time spent on it. The color set of Destination is defined as follows:

which consists of a list (route to reach its destination) and an integer (total transit time).

This section explains the event patterns modeled and automatically implemented, by using our MEdit4CEP-CPN editor [6], to detect situations of interest in two domain applications: air quality and road traffic.

CEP technology provides us with air pollution and traffic condition levels, which can then be used in the city map CPN model presented in Sect. 4 to make decisions about traffic regulations. In this scenario, we use synthetic data to simulate this information flow, which is assumed to be provided by a CEP engine. The integration of both technologies, thus allowing the CPN to provide traffic regulations conclusions in real time, will be a matter of future research, as indicated in the conclusions section, and also following some ideas about the learning of the proposed system that were presented by Li et al. in [20].

We start from the city map CPN model presented in Fig. 5, which is now enriched in order to manage the two situations of interest indicated in the previous section. The intention is to make decisions about traffic control by closing and opening the connections between zones according to the levels of traffic and pollution. Closing an area may seem an extreme decision, especially if it is an access to the city. However, it is a measure that can be adopted in large cities, as occurs in the case of Madrid. According to Article 35 of the Sustainable Mobility Ordinance, approved by the Madrid City Council, extraordinary measures to restrict traffic and parking vehicles on urban roads may be enacted either during episodes of high air pollution or for reasons of road safety and severe traffic congestion [21]. Taking this ordinance as reference, connections will be closed when one of the following conditions occurs:These specific conditions are encoded in our CPN model by the Boolean function close:

Parameter a represents the Air Quality level, obtained by the AirQualityLevel pattern (see Fig. 11), p stands for the number of peaks, obtained by the Threshold4PM2_5 pattern (see Fig. 2) and l is the traffic level, obtained by patterns FreeflowTraffic, HeavyTraffic, SatturatedTraffic and StopAndGoTraffic (see Fig. 13 for SaturatedTraffic level pattern).

Furthermore, traffic levels are computed taking into account the average speeds of cars in relation to the limits, so we can now use the current traffic level to compute the average times for the transits between adjacent zones, using the following function:

Parameter l is the traffic level and t the average time for a transit between two adjacent zones. This function takes into account the ratios between average and limit speeds provided in Table 1, establishing how the transit time is increased for levels 2, 3 and 4. These times are actually increased in an inverse proportion of 80%, 50% and 10%, respectively. These percentages capture the average proportion reduction in traffic speed according to Table 1.

Function genAQLlevel generates Air Quality Levels, genPeaks generates the peaks and genTraffic the traffic levels for transits between Zonei to Zonej and vice versa. Function genAQLlevel is based on a custom distribution, where levels 2 and 3 have the highest likelihood (40% and 35%), levels 1 and 4 are almost the same (9% and 10%), level 5 is rare (5%), and level 6 is very rare (1%). It is then defined as follows:

where function discrete(1,100) generates a random value between 1 and 100, which is then used to produce the final result. In a similar way, function genPeaks produces the peaks, which are evenly distributed between 1 and 20 occurrences, so it is defined as follows:

The generation of traffic levels is also based on a custom distribution, where levels 1 and 2 occur often (40% and 50%), 3 is rare (8%), and 4 hardly occurs (2%):

Function genLevels is used in the new city map CPN model (see Fig. 14) to feed places Levelij by the firing of a new transition init_g. Place pinit is initially marked with one single token, so init_g only fires once. Its firing generates one token on each place Levelij, with the 4 values previously mentioned, namely Air Quality level, Peaks and Traffic level in both directions.

For instance, we could obtain one token 1‘(3,15,2,3) as initial marking for a place Levelij, which corresponds to an AQI 3 level, 15 peaks and traffic levels 2 (heavy) for the transit Zonei to Zonej and 3 (Saturated) for the opposite direction.

The city map CPN model is then modified as indicated in Fig. 15. Transit transitions (tr2to3 and its reverse tr3to2 in the CPN piece shown in the figure) now use the level information provided by init_g, so guards isOpen(3,x,a,p,l1) and isOpen(2,y,a,p,l2) are used to check whether the transit is open or closed, taking into account pollutants and traffic conditions.

These guards are defined as follows:

As an illustration, consider the markings shown in Fig. 15 for places Zone2, Zone3 and Level23. Transition tr3to2 is closed, since the Boolean expression (a=3 andalso l2> 2) is true and therefore close returns true. However, transition tr2to3 is open, since none of the close conditions is satisfied.

Notice also the use of function traverseSection in the arcs from transit transitions to their destination zones (see Fig. 15). This function is implemented by using the traffic_time function commented above, as follows:

which adds the current leg⁴ to the path followed and increases the total time by the amount computed by function traffic_time. For instance, taking again the marking in Fig. 15, we have obtained \({ traverseSection}(3,x,c23,l1)=(6,[1,2,3],8)\).

Figure 16 indicates the changes for the connections with the Destination place. In this figure, we have considered 100 tokens in the Zone1 place traveling from Zone1 to Zone6. This choice of 100 tokens for the simulations has turned out to be an appropriate value to obtain the minimal route, i.e., some tokens have been able to reach the Zone6 place. Notice the loop arcs between the Zonei places and the end_i transitions, which are now introduced so as to obtain the minimal route in the Destination place, instead of all the possible routes from Zone1 to Zone6. Thus, only one initial token is now required on Destination for the route under study, which initially has a very high value (tmax) as total transit time (second field). As new tokens arrive at Destination, this value is updated with the minimum value between the old one and that of the token that has just arrived. For this purpose, function mini is defined, which compares the time stored on the token on Destination with the time obtained from a route, taking the minimum one:

Function isDestiny extends function isDest introduced in Sect. 4. This function enables transition end_i when the token represented by parameter y has reached its destination (isDest), as previously mentioned. Furthermore, isDestiny checks whether the initial and target zones are the same in y and z.

Should we need to compute a different route, we would only change the initial marking on Destination to the corresponding one 1‘([i,j],tmax), writing the following initial marking on place Zonei: 100‘(j,[i],0). Moreover, we can compute several routes at the same time by marking the Zonei places and the Destination place with the corresponding tokens. For instance, we can compute a route from Zone1 to Zone3 and a route from Zone3 to Zone6 writing the marking 1‘([1,3],tmax)++1‘([3,6],tmax) at Destination, token 100‘(3,[1],0) at Zone1 and token 100‘(6,[3],0) at Zone3.

Figures 17 and 18 show the pollution and traffic levels obtained in a simulation and the corresponding result obtained in place Destination, with the fastest route from Zone1 to Zone6. According to the levels indicated in Fig. 17, transitions tr1to2, tr1to3 and tr1to5 would be closed, so the only route available is that indicated on place Destination in Fig. 18. Notice that this is route b in Table 2, but now the average time is different. In Table 2, this route took 11 min and now, due to heavy traffic conditions, it takes 14 min.

Finally, notice that the state graph can now be constructed in an even lower time, because of the restrictions introduced. Some transitions can now be closed, so the number of possible routes can be reduced significantly, and consequently, the graph size can also be reduced. Figure 19 depicts state 6 of the state graph that is obtained with the same simulated conditions of pollutant and traffic levels, which corresponds to the fastest route.

In this section, the city map CPN model is applied to a real scenario, which is based on the division into districts of Madrid, the capital city of Spain. Madrid consists of 21 districts.⁵ In addition, 24 air quality monitoring stations are spread throughout the city (Table 4), but not in all districts. Hourly data gathered from these stations can be read at the URL http://​www.​mambiente.​munimadrid.​es/​sica/​scripts/​, where we can also obtain data logs.

Figure 20 shows the 21 districts and the air quality monitoring stations, which are indicated in either yellow or green. We decided to join districts 10, 19 and 20—in which there are no measuring stations—to some adjacent districts, in order to establish our city map model, so as to have an air quality monitoring station in each zone. Thus, we consider a map that consists of 18 zones, and in each zone, we have a station that provides us with the air quality information. We call \({ Zonei}\) the zone corresponding to district i and \({ Zonei\_j}\) the zone obtained by joining districts i and j.

From Table 4, we can see that only \({ NO}_2\) is measured in all stations, so this will be the only pollutant considered in the scenario. In fact, NO\(_2\) is possibly the most significant pollutant in the case of Madrid, due to the peak values it reaches during the episodes of high air pollution. This pollutant is mainly produced by diesel engines, so the city council applies traffic restrictions during these episodes, banning vehicles from driving in the city center.

We have several traffic monitoring stations in each zone, located in the main streets of each district. Thus, we selected one traffic monitoring station in each zone in order to gather the estimated traffic in the zone, so these stations provide us with the general traffic conditions in each zone. The information gathered from these stations can be obtained at https://​bit.​ly/​2Oobzzy. In this case, the information is updated every 5 min and can be rendered using either Google Maps\(^\circledR \)  or Google Earth\(^\circledR \)  apps, as shown in Fig. 21.

We then constructed the city map CPN with 18 zones (see Fig. 22), and the closing conditions were adapted to only one pollutant (NO\(_2\)), as follows:

The pollution and traffic information used to restrict the movements from one zone to another is always based on the destination zone, as shown in Fig. 23, where we can see that transits from Zone2 to Zone3 (transition tr2to3) use the levels taken from Zone3, and conversely for the movements from Zone3 to Zone2.

As an illustration, we have chosen a high pollution scenario that occurred in Madrid on January 23rd, 2018, at 21:00, which is shown in Table 5. Columns H18 to H22 indicate the NO\(_2\) level measured from 18:00 to 22:00, and \({ P21} and { P22}\) the average increase in the last 3 h, but only if it is positive, otherwise it is zero. Values greater than 15 for these two last columns are highlighted in bold.

In this table, NO\(_2\) levels are reported in ppb, but the information provided by the stations is expressed in \(\upmu\,{\text{g/m}}^3\), so a measurement conversion is required. For this conversion, NO\(_2\) pollutant is considered to behave as an ideal gas, so the ideal gas equation, \(PV=nRT\), is applied for the conversion, taking as temperature T the average value obtained in Madrid on January 23rd, 2018, 283.55 K and taking as pressure (P) the average value in Madrid, 0.9114 atm. In this equation, V is the volume to be obtained, n the number of moles and R the gas constant, \(R=0.082\) atm  l/(mol K).

We used these data to obtain the levels indicated in Fig. 24, and then according to our closing conditions, the access to both Zone5 and Zone8 is closed at 21:00, because the NO\(_2\) pollution level is Unhealthy for Sensitive Groups (3); in the last 3 h the increase is always positive, and the average increase is higher than 15 ppb. One hour later, at 22:00, access to Zone5 is open, Zone8 is still closed, as are Zone4 and Zone7.

With these levels (at 21:00), we obtained the simulation results shown in Fig. 25 for a movement from Zone1 to Zone20_21, using 1000 tokens. This starting value of 1000 tokens was obtained after a number of simulations, with increasing values, until a stable value was obtained. The shortest path obtained for these conditions is indicated by the first field on the token on place Destination:

which took 54 min (second field).

The results of the state space analysis are shown in Table 6 for the same route used in the simulation. The state space has been constructed for three different scenarios. The first, No Restrictions, corresponds to the ideal situation, free flow traffic in which all transits are open, and thus, it produces the largest graph, with the maximum number of both nodes and arcs. The other two scenarios correspond to the data obtained at 21:00 and at 22:00, respectively. In order to explore the state graph produced and thus obtain the shortest path, we use the SearchAllNodes(pred,eval,start,comb) function, provided by CPN tools. This traverses all the nodes of the state space, evaluating the eval function for the nodes for which predicate pred is true. This evaluation starts with the initial value indicated by the expression start, and the result is obtained by combining (function comb) the value obtained with the previous execution of eval and the new value computed for the current node.

Thus, the specific use of this function for our case study is as follows:

where start and Shorter are defined as follows:

Function eval is then executed for all nodes of the state space in order to obtain the token in the Destination place of page EndDestination, comparing the second field (time spent) with the previous value so as to obtain the shortest one. The resulting value after evaluating this expression for the 21:00 scenario is 1‘([1, 2, 13, 1819, 2021], 54), which coincides with the value that was obtained by simulation.

Two strategies have also been applied to reduce the exploration. The first, called Branching Pruning, uses the Branching and Stop options of CPN tools to prune the exploration when a path has a cost greater than another one already computed. In this case, the exploration following that path is stopped, so a significant reduction is obtained, as we can see from the table in the three scenarios. The predicate used for the Branching Pruning is the following:

In this predicate, every if statement checks whether the Zonei and Destination places contain any tokens and every then clause returns whether the travel time of the token at Zonei is lower than those already explored that were obtained with the expression VALUE. Thus, the state space exploration of a certain branch is pruned if the travel time obtained is higher to this value.

In fact, from the analysis of the graph we concluded that the best route in all scenarios was that obtained by simulations, since none of the routes in the graph obtained by using the Simulation Pruning strategy could reach its destination (all were worse and so were stopped). The predicate for the Simulation Pruning is similar to the previous predicate but substituting the VALUE expression with the values obtained by simulation (40, 54 and 54, respectively).

In recent years, several works have proposed ITS models based on PN formalisms. In order to enhance them with the functionalities provided by the CEP technology, we proposed an unprecedented ITS model [22], which integrates both PNs and the CEP technology to analyze traffic regulations only based on the CO pollutant level imposed by the EPA.

Regarding other works on modeling of ITS systems by using PNs, Cavone et al. [23] present a survey on freight logistics and transportation systems based on PN formalisms together with applications to analysis, simulation, optimization and control. Junior et al. [24] propose an analytical model based on the Stochastic PN (SPN) theory for evaluating Vehicular Ad-Hoc Networks (VANETs) infrastructures, where expolynomial distributions are used to represent roadside unit service rates. They study the overall system performance taking into account parameters such as vehicular density, message frequency and RSU (RoadSide Unit) radius. Qi et al. [25] make use of deterministic and stochastic PNs to design an emergency traffic-light control system for intersections prone to accidents. Reachability analysis techniques are then used to prevent deadlocks and livelocks. Júlvez and Boel [26] propose a dynamic model based on continuous PNs to model the macroscopic behavior of traffic systems. The proposed traffic model provides a predictive control strategy on traffic systems taking into account that traffic conditions may vary quickly. The authors focus on the minimization of the total delay (waiting time) of the cars in the system. Hübner et al. [27] propose a vehicle formation model for traffic optimization by means of a consensus algorithm and a condition event PN, which is used to model the topology of the vehicles’ positions. The goal formation is characterized by maximum vehicle density and limited interactions. Riouali et al. [28] use Generalized non-deterministic batch Petri Nets (GNBPN), which are an extension of hybrid PNs. They model discrete and continuous aspects of traffic flow dynamics, taking into account state dependencies based on external rules, such as stop sign or priority roads. ČapkoviČ [29] uses three different models of PNs to model segments of a transport network, namely P/T Petri Nets, Timed Petri Nets and Hybrid Petri Nets. P/T Petri Nets are used to find the safe and unambiguous structure of the traffic-light controller. Then, Timed Petri Nets are used to analyze the time relations between the traffic lights, and finally, Hybrid Petri Nets are used to find flows of vehicles within the bounds of possibility determined by the traffic lights. Bonnefoi et al. [30] propose a specification methodology based on a set of UML diagrams to generate an analyzable PN formal model of an ITS. The system requirements are then expressed as LTL or CTL properties, which are verified by using a PN model checker. Aitouche et al. [31] present a multiagent model using CPNs for traffic regulation of an automated highway. Their goal is to find solutions for the problem of congestion in Automated Highway Systems (AHS), taking into account the departure and arrival time of vehicles, ensuring their correct routing.

Huang et al. [32] use Synchronized Timed Petri Nets in a methodology to design and analyze an urban traffic network system in a modular way, showing the clear presentation and readability of such design. Qi et al. [33] use Timed Petri Nets to design a two-level strategy at signalized intersections for preventing incident-based urban traffic congestion by adopting additional traffic warning lights.

In this paper, we propose an ITS model for traffic control considering both air pollutant and traffic levels with the aim of alleviating not only pollution but also other traffic problems. The proposed model conforms to the three aspects that an ITS should satisfy: (1) consistent with traffic flow, (2) flexible to characterize a dynamic flow and (3) simple, but rich enough to allow us to draw conclusions about traffic regulations.

Air quality conditions and traffic flow data are assumed to be determined by a set of sensor stations. The CEP technology is then used to process the information gathered from these sensors and determine the AQI and traffic levels. Specifically, we use the MEdit4CEP-CPN tool to model the event patterns regarding the PM\(_{2.5}\) pollutant as well as the traffic level event patterns. A CPN modeling of a city map is then defined to compute the optimal available routes taking into account that some connections may be closed due to air quality or traffic conditions. In addition, an important feature of the CPN model is that transit times between zones are computed taking into account that traffic density is likely to affect them.

The analysis techniques of CPN tools are used to obtain these optimal routes. One of these techniques is state exploration based on constructing the state space graph, which allows us to obtain all possible routes and thus obtain the optimal one. However, the state space graph can generally be quite large, and in some cases it could even be impossible to generate. In these cases, simulation techniques are to be applied, which quickly provide suboptimal solutions. The simulation technique applied in this work is based on generating a number of identical tokens on the starting zone, so as to cover as many routes as possible, and return the fastest one as the suboptimal solution. State space exploration can benefit from the branching and stop options of CPN tools so as not to construct all the state space, pruning the branches that have a higher cost than others previously computed. We apply this technique, showing the benefits it provides in both time and graph size. We also propose a combination of both simulation and state exploration, using the results obtained by simulation in order to apply the branching and stop exploration.

As future work, the CPN model could be enriched by improving the initial transit times between zones, taking into account, for instance, the size of the regions and the starting and ending points inside each region. Other key information to compute the routes could be to consider traffic-light information or accidents that could cause delays or closing connections.

Other aspects could also be considered, such as closing to traffic on the basis of type of fuel and emissions of vehicles.

A final step would be to integrate this CEP-based solution with an Enterprise Service Bus (ESB) to test this approach in a real scenario, in which sensing data will be produced by heterogeneous and ubiquitous sources.

Note: All the CPN models presented in the paper, the reports obtained, and the predicates used for the application of branching and stop options are available via the link: https://​doi.​org/​10.​17632/​cbjxbhzn43.​1.