Evaluating Google Maps’ Eco-Routes: A Metaheuristic-Driven Microsimulation Approach

Abstract

Eco-routing, as a key strategy for mitigating urban pollution, is gaining prominence due to the fact that minimizing travel time alone does not necessarily result in the lowest fuel consumption. This research focuses on the challenge of selecting environmentally friendly routes within an urban street network. Employing microsimulation modelling and a computer-generated mirror of a small traffic network, the study integrates real-world traffic patterns to enhance accuracy. The route selection process is informed by fuel consumption and emissions data from trajectory parameters obtained during simulation, utilizing the Comprehensive Modal Emission Model (CMEM) for emission estimation. A comprehensive analysis of specific origin–destination pairs was conducted to assess the methodology, with all vehicles adhering to routes recommended by Google Maps. The findings reveal a noteworthy disparity between microsimulation results and Google Maps recommendations for eco-friendly routes within the University of Pittsburgh Campus street network. This incongruence underscores the necessity for further investigations to validate the accuracy of Google Maps’ eco-route suggestions in urban settings. As urban areas increasingly grapple with pollution challenges, such research becomes pivotal for refining and optimizing eco-routing strategies to effectively contribute to sustainable urban mobility.

1. Introduction

The impact of the transportation sector as a global polluter is significant, accounting for 29% of total United States (U.S.) greenhouse gas emissions by the business sector [1]. Land transport plays a significant role, accounting for 18% of global CO₂ emissions [2]. Traffic engineers have been working on solutions that can reduce these CO₂ emissions. Some of those solutions include better coordination for traffic signals, fewer stops for heavy vehicles, optimization of route choice, and others [3].

Route choice problems involve finding a path between origin and destination points in a transportation network that satisfies predefined criteria. The main goal is usually to minimize travel time, but researchers have also optimized other criteria, such as avoiding congestion and improving driving comfort [4]. Google Maps, as one of the most widely used route-finding applications today, has an incredibly significant impact on travel decisions all around the world. Google Maps currently holds a significant share in the navigation app market due to its extensive real-time traffic data, wide user base, and integrated eco-routing options. However, apps such as Waze, which also operates under Google, excel in crowdsourced, real-time reporting, which can enhance route efficiency for users. HERE WeGo is known for its offline navigation capabilities and robust mapping data for urban and rural areas, though it may lack the depth of Google Maps in eco-routing. Apple Maps has improved in recent years, particularly in integrating features for user privacy and select eco-routing, yet still lags in adoption and may not offer the same level of traffic or emissions data.

Even minor improvements in Google Maps’ route selection could have a tremendous impact on the resulting driving decisions and environment. From October 2021 onward, Google Maps has been suggesting as one of the recommended routes an eco-friendly route that, according to Google, minimizes impact on the environment (or at least represents the best of the three offered solutions) [2]. However, although probably very well designed and calibrated, this method used by Google to find the optimal eco-route has not been tested by researchers yet.

In this paper, we address the problem of finding environmentally friendly routes in an urban street network by considering various geometric and traffic operations elements. Minimizing travel time on a route does not imply that environmental parameters, such as fuel consumption and emissions, are also optimized [5], which highlights the distinctive nature of the eco-routing problem compared to the standard route choice problem that primarily focuses on travel time minimization.

While eco-route selection has been a topic of research in previous studies [6], we have not identified any scientific studies that specifically test the accuracy of Google Maps’ eco-routing recommendations using microscopic simulation. In our study, we aim to address this gap by evaluating Google Maps’ eco-routes against routes generated through detailed simulation models. By using the Vissim 2022 software, which allows for high-resolution simulation of individual vehicles and traffic signal interactions, we model traffic operations at the street level for an urban network. This study aims to address this research gap by exploring how Google selected eco-routes compared to the ones for the same urban network modeled by highly sophisticated simulation tools: Vissim for traffic operations [7] and Comprehensive Modal Emission Model (CMEM) [8,9,10,11] to model the environmental impact of the routes. This approach enables a precise comparison between Google’s eco-routes and routes optimized for reduced fuel consumption and emissions in a controlled simulation environment.

To accurately simulate field traffic conditions, we developed a Vissim microsimulation model of the street network on the University of Pittsburgh campus, with 50 intersections, 35 of them being signalized. This allowed us to properly model travel behavior, traffic demand patterns, and traffic signal settings. We utilized this urban network as a test bed to investigate the quality of Google’s approach in the most conservative/difficult scenarios. The urban network is particularly suitable for our purposes because Google’s method is supposed to be “more sensitive” in highly urban environments than in rural environments or highway sections. Such sensitivity arises from the presence of stop-and-go patterns at signalized intersections, which are difficult to predict without knowledge of signal timing plans. Furthermore, the network was tested under various levels of traffic saturation. The major arterials of our proposed network operate near saturation conditions, while the traffic on the minor arterials is undersaturated. We performed multiple simulation runs to quantify the advantages of utilizing real-time data from Google Maps and precise information about the traffic signal timing plans. The objective is to identify environmentally friendly routes using Vissim and CMEM and compare the findings with Google’s results.

Fuel consumption and emissions are obtained from CMEM [11] by generating vehicle trajectories from Vissim (for all predefined routes in the tested campus network) and finding relevant Fuel Consumption (FC) and other emission factors (e.g., CO₂, NO_x, HC, and CO).

To satisfy more specific objectives of our research, we tried to answer the following questions: (a) Are the fuel savings reported by Google Maps, for considered routes, consistent with fuel savings from the simulation approach? (b) How do eco-friendly routes suggested by Google Maps compare to those from the simulation approach? Are the results consistent, and if not, why?

The paper is organized as follows: After the brief introduction, the second section reviews relevant studies from literature. Then, a proposed methodology is presented, followed by results and discussion. The last section is dedicated to conclusions and future research.

2. Literature Review

The problem of eco-route choice belongs to the group of complex problems due to various factors involved in the multi-criteria decision process. To make an accurate assessment of environmentally friendly route choices, one must have access to specific information, such as data about the vehicle, engine types, road geometry, traffic signal timing parameters, weather conditions, and even prior engine operating time before inquiring about the route.

To find environmentally friendly routes, the Google team collaborated with the National Renewable Energy Laboratory (NREL) to develop a method for predicting fuel consumption. This method used two modules from previous relevant research, namely, FASTsim [12] and RouteE [13], which are linked to Google’s travel time prediction model (based on live traffic data, historical traffic patterns, and road closures). While FASTsim (Future Automotive System Technology Simulation) estimates vehicle’ performance based on vehicle type (light, medium, and heavy), model and manufacturer, engine type, fuel conversion, wheel types, battery drive for hybrid or electric vehicles, etc., RouteE [14] is a machine learning model trained with real-world driving data, including energy consumption behavior for trips around the world.

The baseline of the whole prediction process is energy consumption on a given link (road or path between two nodes). Then, Google counts travel time and fuel consumption data for the route and user. In that way, the trade-off between these two parameters eliminates routes with long travel times and high fuel consumption and suggests a set of reasonable routes. One of them is the most environmentally friendly route. In addition, the stop-and-go pattern at the intersections on the route is counted when estimating the travel time.

Researchers have shown an interest in studying fuel consumption as a criterion for optimizing route choice. Bandeira, et al. [15] have used GPS-equipped vehicles to estimate the effects of emissions on route choice. While they found that faster intercity routes were more desirable in terms of fuel consumption and CO₂ emissions, the same routes increased other environmental factors (e.g., CO, NOx, and HC) by up to 150%. Yao and Song proposed an eco-route algorithm based on vehicle operation and emission data collected in Beijing, China, and the Dijkstra algorithm. The results showed that the proposed eco-route algorithm significantly reduces fuel consumption [16]. In Los Angeles, California, Boriboonsomsin et al. [17] examined the relationship between travel time and fuel consumption in route choice. They concluded that eco-friendly trips shorter than 10 miles save more fuel but take more time. Pereira et al. [18] compared the SmartDecision eco-route choice application with travel time optimization-based route choice on Google Maps. They found that the SmartDecision application provided routes with 15–30% less pollution and social cost than Google Maps. It should be noted that these experiments were conducted before Google introduced the eco-route choice. Bandeira et al. [19] have investigated the sensitivity of eco-route change during rush hour by collecting 222 h (about 1.5 weeks) of GPS (Global Positioning Systems) vehicle data. Although eco-routes reduce pollutant emissions by up to 60%, the data show that eco-routes are constant under different traffic volume levels, which could increase overall travel time. Alam et al. [20] investigated personal exposure to PM10 in route choice. In the case of Dublin, the results show that optimal route travel time did not correspond to the lowest values of PM10. Namoun et al. [21] addressed the problem of multimodal eco-route optimization. Results from the United Kingdom and Bulgaria show the applicability of a travel information system with eco-friendly route guidance. More recently, Fahmin et al. conducted a comprehensive literature review on eco-friendly routing [22]. Teng et al. introduced a path-ranking algorithm designed for a bi-objective eco-routing model focused on reducing both fuel consumption and travel time [23]. In the same year, Teng et al. developed a reliable pathfinding model that was proposed to minimize the fuel consumption budget [24].

Over the past few decades, a wide range of techniques and criteria have been employed to address ecologically friendly route choice problems. Microsimulation is a widely used technique in traffic engineering that aims to mimic actual traffic conditions and assess the effectiveness of specific control measures [25]. In this paper, we address the problem of eco-routing by using a specialized microsimulation tool for traffic control and developing digital twins for traffic networks.

3. Methodology

This section explains the data collection and analysis to identify environmentally friendly routes within urban street networks. Initially, a model of a subject street network was created in a microsimulation environment that depicts field-like traffic situations. To accomplish this, the authors used the network re-calibrated with Google Maps travel times. Then, the simulation data were systematically collected, on second-by-second trajectories, to obtain accurate fuel consumption and emissions data. The overall methodology used in this study is outlined in Figure 1.

3.1. Experimental Network

The proposed methodology for eco-friendly routing used a microsimulation approach. The testing was conducted with a field-like urban network situated in the area of the University of Pittsburgh campus in Pittsburgh. This network comprises 50 intersections, with 35 of them being signalized.

Traffic count collection was carried out by students on a few workdays during September 2022. The traffic signal timing was provided by the City of Pittsburgh. Travel times between intersections were also recorded during the same peak hours (AM) when the traffic counts were collected. All travel times were recorded on the same working day as the traffic counts, and multiple rounds of trips were conducted to obtain an average travel time as the final value.

To ensure that the Vissim model accurately represents the field network, an initial calibration process was undertaken. The aim of this calibration was to closely align the model with the field data. To perform the calibration properly, we executed simulations with five different random seeds. The warm-up time was set to 900 s, and the simulation duration was 1 h. The calibration results, including turning movement counts (TMCs) and travel times, are illustrated in Figure 2.

The calibration process involved the utilization of 286 data points pertaining to traffic turning movements, as well as 40 data points for travel times. The average R² values obtained for the calibration of turning movement demands and travel times were 0.9715 and 0.8059, respectively. These outcomes indicate a high level of effectiveness in calibrating the network, thus establishing it as a dependable model for evaluating the efficacy of eco-routes on the field-like campus street network.

While this calibration process ensured that the simulation model would reflect typical peak-hour demands and travel times on the campus, it was still not reasonable to expect that it could account for the exact traffic conditions at the time when experiments are performed or when a routing inquiry is made. Therefore, the authors needed to re-calibrate the network every time a new routing inquiry was made, and this was a major computational and methodological effort. Details about this re-calibration process are provided below in this section. It is worth mentioning that the process of network re-calibration begins with the states of the network that emerged from the initial calibration.

3.2. Identification of O-D Pairs and Period for Analysis

Nine origin–destination (O-D) pairs, connecting various entrance and exit points (shown in Figure 3) of the University of Pittsburgh campus network simulation model, were used in the analysis. All routes in Figure 3 were generated using Google Maps. Most of the routes connecting these O-D pairs had travel times between 6 and 12 min. Considering the size and shape of the network, the authors applied their best effort to define a number of the O-D pairs and their geographical distribution. These efforts included trial-and-error investigation of the O-D pairs to remove those O-D pairs (and corresponding routes) that did not produce enough variability in Google Maps’ results (e.g., if Google Maps always identified only one route for a certain O-D pair). It should be noted here that similar eco-routing papers (16) generally examine up to half a dozen routes, whereas there are some that have studied about a dozen routes [19].

The deliberate exclusion of origin–destination (O-D) pairs with limited route variability in our investigation into Google Maps’ eco-routing recommendations is a strategic choice to elevate the robustness and significance of our analysis. By omitting pairs displaying constrained route variations, we narrowed our focus to O-D pairs more likely to offer diverse eco-routing options, facilitating a comprehensive understanding of the application’s responsiveness to dynamic traffic conditions. Importantly, this exclusion does not imply that the exclusive route for those pairs simultaneously fulfils multiple objectives such as minimizing travel time, distance, and fuel consumption. Rather, it signifies that routes generated by Google Maps for those specific O-D pairs lack significant variation, hinting at a constrained impact of parameter alterations on the recommended routes. This deliberate selection ensures that our study concentrates on scenarios where changes in traffic conditions or other factors substantially influence Google Maps’ eco-routing suggestions, contributing to a more accurate evaluation of the application’s overall performance. It is crucial to underscore that our primary goal is to identify and comprehend patterns of variability in eco-routing suggestions, eschewing an exclusive focus on routes meeting multiple criteria concurrently.

The purpose of looking at more than a few O-D pairs was to examine how sensitive Google Maps is in detecting alterations in traffic demand and to observe whether such potential alterations have an impact on the results of Google Maps’ eco-routing recommendations.

For each analyzed hour, we performed multiple iterations of this analysis, based on 15 min time steps. The 15 min interval was selected because it is short enough to quickly capture any major changes in traffic demands and long enough to avoid meaningless repetition of the procedure when Google Maps provides the same outputs. If the relevant conditions change during a 15 min interval (e.g., travel times on the routes, suggested routes, eco-route recommendations), the re-calibration step is repeated to account for the real field data. In this way, each 15 min interval of changes is treated as a separate simulation task.

After the initial calibration and identification of origin–destination (O-D) paths, we propose the following algorithm to obtain travel times for selected routes and find eco-friendly routes within the simulation.

Every 15 min of simulation, perform the following steps for each link in the network:

Step 1. Collect travel times from Google.

Step 2. Re-calibrate the links based on Google data.

Step 3. Calculate fuel consumption and emissions.

Step 4. Use multi-criteria ranking to select eco-friendly routes.

Each of these steps will be explained in detail in the following subsections.

3.3. Google Maps Eco-Route Data Collection

For each 15 min period (of the analyzed one hour) and each of the O-D pairs, we performed inquiries on Google Maps (e.g., personal cell phones were used for this purpose) to obtain the best route between the given O-D pair. Google Maps would then usually report three different routes (with their respective travel times), one of which would be identified as the Eco-route. The inquiries were made on 23 June 2023 from 4 pm to 5 pm.

Each of the recommended routes was recorded, and they were given appropriate names for easier data handling (Figure 3). The eco-route, as recognized by Google Maps, was marked as such, and all its data were time-stamped for easier post-processing. At the same time when the inquiry is made, the researchers took a screenshot of the Google Map to have a record of traffic congestion as displayed in color-coded links of the Google Map (Figure 4). These color-coded links were later used for re-calibration of the simulation conditions.

Routes were recorded automatically by using the Tasker application designed for Android smartphones. By taking advantage of Android’s accessibility features, Tasker can launch apps, take screenshots, and even integrate with a plugin called “AutoInput” that allows users to place input and automatically populate text fields. The Tasker provides a graphical user interface and a way to automate the execution of the scripts. The authors set Tasker to run a script every 15 min that automates screenshots of a Google Maps route.

3.4. Re-Calibration of Routes in Simulation

To make a fair comparison (between routes identified by Google Maps as eco-routes and those obtained through simulation analysis), we had to ensure that the traffic in the simulation was similar to traffic conditions in the field, as recognized by Google Maps, at the time when an inquiry for eco-routing was made. For this purpose, the authors used the Genetic Algorithm (GA) technique, a well-known metaheuristic approach.

GA is an optimization technique inspired by the principles of natural selection and genetics, commonly used to solve complex problems that are difficult to address with traditional methods. Genetic algorithms operate by generating a population of potential solutions, called individuals, which evolve over successive iterations (generations) to improve performance. Each individual is evaluated using a fitness function, which measures how well it meets the desired objectives. The fittest individuals are selected for reproduction, where they undergo processes analogous to biological crossover (combining parts of two individuals) and mutation (introducing small random changes). These processes generate new individuals that inherit traits from previous generations, encouraging diversity and exploration within the solution space. Over time, the genetic algorithm converges toward optimal or near-optimal solutions.

Before the GA re-calibration, we manually imposed customary desired speed decisions (an element in Vissim) on some of the secondary streets. This was needed due to ongoing traffic conditions resulting from poor pavement conditions, work zones, street parking, stop signs, speed bumps, etc.

Before re-calibration, we also consider the analyzed network as a group of segments. Herein, we define a segment or a link as a directional group of traffic lanes between two signalized intersections. This was important, when comparing field and simulated travel times, to avoid the situation where the total travel times of a route are similar but the segment travel times are very different, thus causing situations where two errors cancel each other. Thus, the travel time for a given route is the sum of the travel times of the segments that cover the route.

A mathematical formulation of the re-calibration process is presented as follows (Equations (1) and (2)):

Minimize

$$\sum _{s=1}^{\left|S\right|}\left|{GTT}_s-{STT}_s\right|$$

(1)

subject to

$$q_{ij min}\le q_{ij}\le q_{ij max},\forall i,\forall j\in N$$

(2)

where

• s—the index of a segment, where s = 1, 2, 3… |S|;
• i—the index of a vehicle input, where i = 1, 2, 3… |I|;
• j—the index of a vehicle input category, where j = 1, 2, 3… |J|;
• GTTs—Google travel time at the s-th segment;
• STTs—Simulation travel time at the s-th segment;
• q_ij—number of vehicles per hour at the i-th input of j-the category of the network;
• q_{ij min}—minimum values for q_ij;
• q_{ij max}—maximum values for q_ij;
• N—number of vehicle inputs in the network, where n = 1, 2, 3… |N|.

The fitness function (Equation (1)) that needs to be minimized presents the difference between the travel times of all segments provided by Google Maps and the simulation for the one hour of analysis. The constraints in Equation (2) give the interval of feasible vehicle inputs for the network.

Let us set the number of vehicle input categories to three (|J| = 3). The categorization is performed according to the intensity of vehicle volumes in the network. The minimum values for categories 1, 2, and 3 are 100 veh/h, 250 veh/h, and 600 veh/h, respectively, while the maximum values for categories 1, 2, and 3 are 500 veh/h, 1250 veh/h, and 2500 veh/h, respectively. For the subject network, the number of different vehicle inputs (which represents the number of variables) is |N| = 27.

Let us denote by r the route index, where r = 1, 2, 3, … |R|. After re-calibration, Simulation Travel Time (STT) on the r-th route can be calculated as (Equation (3)):

$${STT}_r=\sum _{s=1}^{\left|S\right|}{STT}_s·{\theta }_{sr}$$

(3)

where

• ${\theta }_{sr}=\left\{\begin{array}{c}1\mathrm{i}\mathrm{f}s-\mathrm{t}\mathrm{h}\mathrm{s}\mathrm{e}\mathrm{g}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{b}\mathrm{e}\mathrm{l}\mathrm{o}\mathrm{n}\mathrm{g}\mathrm{s}\mathrm{t}\mathrm{o}r-\mathrm{t}\mathrm{h}\mathrm{r}\mathrm{o}\mathrm{u}\mathrm{t}\mathrm{e},\\ 0\mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e};\end{array}\right.$

We used Google Maps to find the GPS coordinates of the origin and destination points of all segments. Then, by using the Google Maps Directions API, we obtained the travel time for each s-th segment (GTTs).

3.5. Parameters of Genetic Algorithm and Initial Population

In the GA procedure, we used the following values for the parameters of its major operators:

• Population size—20,
• Number of generations—140,
• Mutation probability—2%,
• Crossover probability—60%.

Let us assume that the initial population (P_ini) consist of |P| individuals and that p is the index of each individual, where p = 1, 2, 3, … |P|. It is already decided that |P| = 20. If U_p is the p-th individual in P_ini, the following equation defines the initial population (Equation (4)):

$$P_{ini}=\left[\begin{array}{c}{U}_1\\ U_2\\ ⋮\\ U_p\\ ⋮\\ U_{20}\end{array}\right]$$

(4)

Each U_p consists of |N| q_ij-th variables. Due to that, the initial population P_ini can be described in the following way (Equation (5)):

$$P_{ini}=\left[\begin{array}{cccccc}{q}_{ij1 1}& q_{ij 1 2}& \dots & q_{ij 1 n}& \dots & q_{ij 1 27}\\ q_{ij 2 1}& q_{ij 2 2}& \dots & q_{ij 2 n}& \dots & q_{ij 2 27}\\ ⋮& ⋮& ⋮& ⋮& ⋮& ⋮\\ q_{ij p 1}& q_{ij p 2}& ⋮& q_{ij p n}& ⋮& q_{ij p 27}\\ ⋮& ⋮& ⋮& ⋮& ⋮& ⋮\\ q_{ij 20 1}& q_{ij 20 2}& \dots & q_{ij 20 n}& \dots & q_{ij 20 27}\end{array}\right]$$

(5)

To provide genetic algorithms with a sensible initial population and speed up the search process, we propose the following algorithm (Algorithm 1):

Algorithm 1. Genetic algorithm

Step 1. U1 is given and defined by 27 qij vehicle inputs collected from the field. Step 2. for p = 2 to |P|      for n = 1 to |N|           max $\left\{{int(q}_{ij 1\ast n}*k_{pn}),q_{ij min} \right\}\le q_{ij p n}\le min\left\{{int(q}_{ij 1 n}/k_{pn}),q_{ij max} \right\}$ where kpn are random numbers between 0.8 and 1.

3.6. Re-Calibration Results

To achieve fair comparison conditions, the authors re-calibrated the Vissim simulation model (by increasing and decreasing relevant traffic demands) based on the traffic congestion of individual links, as captured by the screenshots of the Google Maps (taken at the time of each eco-routing inquiry). Figure 4 shows an example of such a comparison for the first 15 min of re-calibration. Vissim’s “link coloring” feature was used to display aggregated average link speeds.

To the authors’ best knowledge, Google does not publicly provide speed ranges (or thresholds) for the different colors, ranging from green to dark red, that are used to indicate the current traffic conditions on the streets (left part of Figure 4). Therefore, we have relied on information from various Google Maps users. However, these sources may be unreliable, and our observations provided the best guess about the meaning of the speed colors (Legend of Figure 4). These color ranges are included in the simulation and present visual outputs of our re-calibration process (right part of Figure 4). However, the authors also suspect that the speed ranges used by Google Maps for coloring differ between highways and urban streets. For this reason, the ranges used in this paper may not represent the universal values for all road types.

The purpose of this re-calibration procedure was to create, in the simulation, link-congestion conditions similar to those from the field. This was performed to try to ensure that a simulated vehicle traveling on an O-D route would face similar conditions (e.g., being stopped at multiple intersections) to those that are perceived by Google Maps at the time of the eco-routing inquiry.

A successful re-calibration, from a macro-scale perspective, consisted of iterative efforts to adjust traffic flows until a good match between color-coded congestion links (in the simulation and the field) was achieved. This means that the travel time differences between Google Maps and Vissim travel times were less than 10% (which is the threshold). The results of the re-calibration for each of the 15 min periods of Vissim analysis (with 5 min of warm-up time) are shown in Figure 5.

In the re-calibration process, 74 travel time data points were used on the same number of segments covering all routes of the studied network. The average R² values obtained for the re-calibration of travel times for each of the 15 min time intervals were 0.8012, 0.8412, 0.8364, and 0.8261, respectively. These results indicate satisfactory effectiveness in re-calibrating the network.

3.7. Emission and Fuel Consumption Model

Once the simulation model was re-calibrated, the authors collected relevant vehicular data for the recommended routes identified by Google Maps for each of the eco-routing inquiries. These data were collected from at least 30 vehicles (per hour) that traveled each of the indicated routes (to provide a large sample accounting for a variety of driving behaviors, vehicle types, etc.). The detailed, second-by-second data from the vehicle trajectories were saved in a format suitable for post-processing in CMEM [26,27,28,29,30,31,32]. Finally, vehicle data from the Vissim trajectories were exported into CMEM, which then calculated relevant fuel consumption, CO₂, and other emissions (HC, CO, and NOx,) for each of the identified vehicles (30 for each route). The connection between Vissim and CMEM is shown in Figure 6.

Comparison of Eco-Routes Identified by Google and Simulation

Once the CMEM results were available for all of the routes and all of the analyzed vehicles (30 per route), the overall fuel consumption and CO₂ quantities were summarized and compared. The routes that had the lowest FC and CO₂ for that 15 min period were named FC Eco-Route and CO₂ Eco-Route, respectively. Most of the time, it is expected that the same route (in simulation) will be identified as both FC and CO₂ Eco-Route (as FC and CO₂ have a quite linear relationship) but some exceptions are possible.

Choosing an eco-route among all those obtained after CMEM tests belongs to the group of problems called multi-criteria decision-making [33]. This problem occurs when a decision maker has to choose one of several alternatives, where each alternative is described by more than one criterion with the associated weighting. A few techniques have been used to solve multi-criteria decision-making problems, including Technique for Order Preference by Similarity to Ideal Solution (TOPSIS), Elimination and Choice Expressing Reality (ELECTRE), Analytic Hierarchy Process (AHP), Preference Ranking Organization Method for Enrichment Evaluation (PROMETEE), and others. In this paper, we use the Simple Additive Weighting (SAW) method, which is the most widely used in practice [34].

In our case, the alternatives (A_i) are i recommended routes in an O-D pair, while the j criteria (C_j) are Fuel Consumption (FC) and emission parameters (CO₂, NOx, CO, and HC). The criteria’s weights are adopted based on our perception and marked by w_j. This constellation can be represented by the following matrix X (Equation (6)):

$$\begin{gathered} {\mathrm{A}}_{\mathrm{i}}/{\mathrm{C}}_{\mathrm{j}} \mathrm{FC} {\mathrm{CO}}_2{\mathrm{NO}}_{\mathrm{X}} \mathrm{CO} \mathrm{HC} \\ X=\begin{array}{c}Route 1\\ \dots \\ Route i\\ \dots \\ Route n\end{array}\left[\begin{array}{ccccc}{x}_{11}& x_{12}& x_{13}& x_{14}& x_{15}\\ \dots & \dots & \dots & \dots & \dots \\ x_{i1}& x_{i2}& x_{i3}& x_{i4}& x_{i5}\\ \dots & \dots & \dots & \dots & \dots \\ x_{n1}& x_{n2}& x_{n3}& x_{n4}& x_{n5}\end{array}\right] \\ \text{Extreme min min min min min} \\ w \text{0.5 0.3 0.1 0.05 0.05} \end{gathered}$$

(6)

The next step in the SAW method assumes the normalization of criteria values. Because it is desirable to minimize all criteria, the normalized matrix is achieved in the following way (Equation (7)):

$$r_{ij}={\displaystyle \frac{\min \left(x_{ij}\right)}{x_{ij}}}$$

(7)

where r_ij are normalized values of x_ij.

Each i-th alternative is evaluated in the following way (Equation (8)):

$$A_i=sumproduct(w_j,r_{ij})$$

(8)

In that way, the values used to compare the alternatives are aggregated. The alternative (route) with the largest A_i value is declared the eco-route.

4. Results

Nine different O-D (origin–destination) routes on the University of Pittsburgh Campus were chosen to evaluate the effects of route choice on fuel consumption and emissions. As a result, 27 route choices for nine O-D pairs were identified based on Google Maps recommendations. These selected routes offer different traffic conditions, including different geometric layouts (some with high-slope streets).

Table 1. Simulation results of fuel consumption and CO₂ emission.

O-D	Route	# of veh	D (mi)	# of SI	0–15 min		15–30 min		30–45 min		45–60 min
					FC	CO2	FC	CO2	FC	CO2	FC	CO2
1	1	11	1.2	12	253.08	700.43	298.21	827.42	324.37	923.44	307.58	874.00
	2		1.2	13	265.32	738.57	279.62	772.58	306.90	860.31	310.11	880.64
2	1	11	1.3	13	269.01	754.26	333.12	937.86	347.08	981.37	340.95	965.90
	2		1.3	13	272.00	751.44	304.11	861.16	339.00	962.68	381.66	1089.01
	3		1.4	10	279.99	783.14	312.47	879.20	372.37	1061.77	371.79	1054.03
3	1	11	1.3	7	226.68	625.60	218.04	600.68	209.72	588.22	210.45	581.64
	2		1.3	7	223.38	612.66	212.39	593.76	218.76	614.75	246.21	689.77
	3		1.3	7	215.84	588.21	205.26	571.79	205.76	574.66	227.74	633.53
	4		1.5	7	258.65	732.55	241.53	696.18	241.18	691.70	263.39	740.26
4	1	11	1.3	6	232.39	639.76	210.92	587.79	207.56	575.83	247.66	695.44
	2		1.4	8	253.36	699.47	234.38	644.59	233.83	655.47	239.74	665.29
	3		1.5	6	259.33	724.15	249.18	700.71	252.05	707.40	280.49	789.76
5	1	14	1.3	16	216.19	597.43	226.17	628.36	229.63	640.63	241.36	674.36
	2		1.6	16	312.36	882.35	302.30	847.74	286.15	808.07	310.02	876.65
	3		1.3	15	205.52	579.04	206.96	586.22	228.72	643.92	215.07	601.89
6	1	15	1.3	11	230.73	651.14	252.30	710.39	242.83	677.56	232.71	654.16
	2		1.3	6	228.61	636.70	237.42	662.75	236.83	658.90	231.66	646.26
	3		1.4	8	219.25	608.54	222.51	614.97	228.14	631.07	226.70	634.19
7	1	12	1.0	11	232.24	643.60	218.26	607.08	209.20	570.37	236.36	656.02
	2		1.1	7	243.04	666.24	233.15	633.23	254.53	703.57	244.13	674.35
	3		1.0	8	218.51	601.32	221.53	608.39	221.67	619.41	241.45	672.15
8	1	14	1.2	15	205.08	573.39	205.64	579.99	222.48	623.45	217.74	610.89
	2		1.3	8	231.11	664.52	250.89	716.59	249.59	713.82	249.89	725.60
	3		1.2	14	203.23	579.90	205.05	584.57	218.89	620.63	206.03	586.74
9	1	13	1.3	14	205.17	577.42	232.95	655.83	235.67	657.89	215.33	610.99
	2		1.1	6	179.01	512.18	191.79	547.29	193.32	551.28	187.21	538.77
	3		1.3	15	211.86	595.81	208.81	588.14	237.00	659.17	225.56	629.25

* Values highlighted in grey are exceptions in which lowest FC does not have lowest CO₂.

shows the fuel consumption (g/veh) and CO₂ emissions (g/veh) results of each of the O-D pairs (and their associated routes). By SI we denote the number of signalized intersections along the route, while D is the distance traversed by a route.

The first method that the authors utilized for comparing the routes is based only on fuel consumption. The reason for this decision is that Google Maps only considers fuel consumption when making eco-routing recommendations. Considering that each 15 min interval was a separate task, 36 different eco-routing decisions were examined, along with the nine O-D pairs. We have highlighted the best fuel consumption values for each of the route choices in bold (Table 1).

Table 1 also shows that, in most cases, the route with the lowest fuel consumption also has the lowest CO₂ emissions. The exceptions to this rule are highlighted (in gray) and occur on O-D pair #2 (0–15 min), O-D pair #5 (30–45 min), and O-D pair #8 (15–30 min).

A graphical representation of the results from Table 1 can be also seen in Figure 7, where the differences between FC and CO₂ are visualized for all of the routes studied. We examined thirty-six separate tasks (four per O-D pair). The results of each task are shown in a separate column as colored circles indicating the corresponding values for FC and CO₂ for each of the routes (between two and four per task). Due to the lack of space, the intervals are shown with the final value; thus, for example, the interval 0–15 is shown as 15.

CMEM considers not only CO₂ emissions but also other pollutants that can significantly affect air quality in urban areas. Thus, in addition to fuel consumption and CO₂ emissions, this study also considered other emissions from CMEM, such as Nitrogen Oxides (NOx), Carbon Monoxide (CO), and Hydrocarbons (HC). The results from those experiments are shown in

Table 2. Simulation results of NOx, CO, and HC emissions.

O-D	Route	0–15 min			15–30 min			30–45 min			45–60 min
		NOx	CO	HC	NOx	CO	HC	NOx	CO	HC	NOx	CO	HC
1	1	1.67	60.52	2.17	2.00	70.54	2.29	1.99	62.85	2.01	1.91	60.09	2.16
	2	1.66	60.88	2.19	1.83	68.05	2.22	1.90	67.58	2.09	1.97	60.92	2.19
2	1	1.71	58.54	2.11	2.11	70.70	2.31	2.14	71.42	2.20	2.14	68.58	2.34
	2	1.77	65.81	2.36	1.99	61.46	2.07	2.07	67.16	2.12	2.30	72.30	2.40
	3	1.80	62.05	2.24	2.00	66.53	2.22	2.25	71.36	2.17	2.27	74.49	2.47
3	1	1.45	54.90	2.13	1.40	53.69	1.96	1.39	45.10	1.83	1.36	50.49	1.95
	2	1.42	56.39	2.17	1.38	46.90	1.86	1.40	46.50	1.82	1.55	53.71	2.04
	3	1.39	56.80	2.15	1.38	46.56	1.83	1.32	45.80	1.81	1.45	52.34	1.98
	4	1.65	52.05	1.83	1.50	41.26	1.53	1.55	43.28	1.59	1.63	56.63	1.86
4	1	1.52	57.30	2.19	1.36	47.74	1.86	1.33	48.58	1.85	1.55	53.08	2.01
	2	1.66	61.45	2.28	1.54	58.49	2.07	1.55	50.76	1.94	1.60	56.14	2.07
	3	1.75	58.55	1.92	1.66	53.25	1.80	1.67	54.74	1.82	1.76	59.52	1.92
5	1	1.46	52.06	1.95	1.53	52.54	1.94	1.52	51.75	1.92	1.63	53.89	1.96
	2	2.18	64.31	2.22	2.13	65.95	2.25	1.96	58.96	2.08	2.10	63.29	2.18
	3	1.48	42.66	1.75	1.48	41.05	1.71	1.56	48.00	1.84	1.50	47.18	1.85
6	1	1.61	47.63	1.76	1.81	53.17	1.90	1.65	54.91	1.92	1.56	49.56	1.83
	2	1.57	52.27	1.89	1.65	53.39	1.92	1.60	54.61	1.95	1.58	52.31	1.91
	3	1.56	51.45	1.81	1.56	53.86	1.86	1.58	54.87	1.90	1.56	50.13	1.83
7	1	1.48	55.12	1.92	1.40	50.37	1.82	1.28	55.21	1.92	1.51	55.47	1.96
	2	1.54	61.97	2.18	1.48	62.96	2.20	1.57	61.36	2.20	1.55	58.95	2.21
	3	1.46	54.02	2.06	1.42	55.59	2.07	1.41	49.13	1.95	1.57	55.07	2.15
8	1	1.34	45.22	1.81	1.35	42.33	1.72	1.44	48.41	1.85	1.46	46.88	1.82
	2	1.56	40.03	1.70	1.67	46.54	1.83	1.66	45.75	1.80	1.71	39.07	1.70
	3	1.40	37.69	1.64	1.42	38.35	1.67	1.51	43.15	1.75	1.41	38.92	1.68
9	1	1.34	42.89	1.79	1.48	48.79	1.92	1.47	52.87	1.97	1.38	42.10	1.76
	2	1.19	32.12	1.55	1.26	35.41	1.62	1.26	35.96	1.62	1.14	31.81	1.52
	3	1.35	44.69	1.79	1.30	43.44	1.78	1.49	54.63	2.01	1.45	50.73	1.94

* The route with the lowest value for fuel consumption does not have to be the one with the lowest emissions. The highlighted cells (in gray) in Table 2 show these inconsistencies.

.

Similarly, in Table 2, bold text is used to highlight the lowest values for NOx, CO, and HC pollutants that correspond to the best routes in terms of fuel consumption. Nevertheless, the route with the lowest value for fuel consumption does not have to be the one with the lowest emissions. The highlighted cells (in gray) in Table 2 show those inconsistencies. For example, in the case of O-D # 1 (0–15 min), Route 1 has lower fuel consumption than Route 2 (Table 1). At the same time, Route 1 is worse than Route 2 in terms of NOx (Table 2). This is the reason why the corresponding cell (1.66) is grayed out.

Inconsistencies between the lowest fuel consumption and the lowest NOx, CO, and HC pollutants occurred more frequently than those between fuel consumption and CO₂. However, when the final eco-routes from the SAW analysis were compared with the lowest fuel consumption routes, no difference was found there. This means that the SAW analysis only reconfirmed that fuel consumption is the ultimate criterion for establishing eco-routes. Therefore, it is significant to emphasize the distance of emission parameters from fuel consumption. We found inconsistencies between the routes with the lowest fuel consumption and the lowest values of NOx in seven out of thirty-six tasks (19.44%). Similarly, such inconsistencies for CO and HC occurred in proportions of 10/36 (27.78%) and 11/36 (30.56%), respectively (Table 2).

A graphical representation of the results from Table 2 can be also seen in Figure 8, where the differences between FC and NOx, CO, and HC are visualized for all the studied routes.

The most important part of the analysis part is to compare the eco-routes resulting from the applied methodology and the eco-routes recommended by Google Maps. While such results are discussed later in the paper, in

Table 3. Eco-routes proposed by our methodology (with Ai values) and by Google Maps.

O-D # 1						O-D # 6
Period	15	30	45	60		Period	15	30	45	60
R1	1.000	0.937	0.948	1.000		R1	0.952	0.885	0.945	0.978
R2	0.961	1.000	0.995	0.989		R2	0.958	0.939	0.967	0.978
O-D # 2						R3	0.995	0.999	1.000	0.999
Period	15	30	45	60		O-D # 7
R1	0.999	0.914	0.956	1.000		Period	15	30	45	60
R2	0.981	1.000	1.000	0.902		R1	0.948	1.000	0.994	1.000
R3	0.958	0.973	0.937	0.921		R2	0.902	0.932	0.819	0.964
O-D # 3						R3	0.997	0.980	0.938	0.973
Period	15	30	45	60		O-D # 8
R1	0.944	0.939	0.970	0.998		Period	15	30	45	60
R2	0.957	0.964	0.936	0.861		R1	0.983	0.992	0.983	0.945
R3	0.988	0.994	0.991	0.927		R2	0.881	0.820	0.882	0.837
R4	0.842	0.852	0.861	0.813		R3	0.993	0.993	0.995	1.000
O-D # 4						O-D # 9
Period	15	30	45	60		Period	15	30	45	60
R1	0.994	0.998	0.999	0.969		R1	0.872	0.826	0.822	0.863
R2	0.914	0.897	0.888	0.991		R2	1.000	1.000	1.000	1.000
R3	0.899	0.852	0.830	0.863		R3	0.847	0.922	0.816	0.821
O-D # 5						Legend: Eco-route proposed by our methodology— Google Maps eco-route—x.xxx
Period	15	30	45	60
R1	0.952	0.921	0.992	0.896
R2	0.665	0.686	0.800	0.704
R3	0.999	0.999	0.996	1.000

, we present the eco-routes identified by the SAW analysis and those from Google Maps. Table 3 also provides the final A_i values from SAW, which gives a reader the ability to understand how SAW values various routing options.

As already mentioned, SAW eco-routes are, at the same time, routes with the lowest fuel consumption. When those were compared with Google Maps’ eco-routes, there was only a 34.48% match—in 10 out of 29 cases. In eight cases, Google Maps did not recommend any eco-routes, whereas in the remaining 11 cases, the two methods proposed different routes as the eco-routes.

Before discussing potential reasons for such results, the authors wanted to further investigate some of the best re-calibrated eco-routes (from Table 3) and compare them with Google Maps. The reasoning behind this additional step was to check whether the match between the two methods for identifying eco-routes increases if the travel times (with all other things being equal) from Google Maps and microsimulation have an excellent match. The criteria for the best re-calibrations were applied to each of the 29 eco-routes resulting from our methodology in Table 3, starting with the following, where B_r is the set of the best re-calibrated eco-routes (Equation (9)):

$$IF\left|{GTT}_r-{STT}_r\right|<7\mathrm{\%} AND \left|{GTT}_s-{STT}_s\right|<20s,foreach s\in r THEN r\in B_r$$

(9)

Seven eco-routes were found after applying the filtering process from above. These eco-routes, along with their associated segment travel times, are shown in Figure 9.

Table 4. Matches between GTTs and STTs for the seven best re-calibrated O-D pairs.

O-D	Route	Period	Google Travel Times (s)	Simulation Travel Times (s)	Differences	Match
1	R1	15–30	433	415	4.2%	YES
2	R2	0–15	402	376	6.7%	NO
5	R1	15–30	356	369	3.6%	NO
5	R1	30–45	369	391	5.8%	NO
5	R1	45–60	423	430	1.6%	NO
6	R2	45–60	369	348	5.9%	NO
8	R1	15–30	375	384	2.4%	NO

presents the disparities between Google Travel Times (GTTs) and Simulation Travel Times (STTs). This examination substantiates our use of the best calibrated routes for comparative analysis, ensuring a meticulous match assessment between eco-routes obtained from Google and those derived from the simulation.

The results from Figure 9 and Table 4 reveal that only one of the seven best re-calibrated eco-routes matches Google Maps’ eco-routing recommendations, which is an even lower percentage (or 14.28%) than in the entire population. It is also worth mentioning that statistical t-tests revealed no significant differences between Google and simulation segments’ travel times.

5. Discussion

Based on the analyzed results, one can conclude that our research found significant inconsistencies between eco-routes identified by Google Maps and our microsimulation approach. While we did expect to see some inconsistencies, the results made us think that the magnitude of the difference is much larger than we expected. There are many possible reasons for such discrepancies.

On the Google Maps side, while we do not have access to the exact details of the applied Google methodology, we want to point to two major possible issues. First, Google Maps’ approach is macroscopic in its nature, and as such, in spite of all good intentions, it may never be so precise as to give a resolution that is good enough to capture the difference between some of the routes, especially if such routes cover short distances and contain many urban traffic impediments. Second, even if macroscopic, the Google Maps’ approach perhaps does not properly account for traffic signals [35] or the impact of road works, local on-street parking, and loading/unloading operations, which are all part of this campus network. Thus, before we make any conclusions let us discuss some limitations of our study:

• A very specific campus network with traffic activities that cannot be generalized as very common in some other networks.
• A microsimulation approach with all of the caveats that such an approach brings, starting with somewhat unrealistic driving behaviors, simplifications in underlying models, etc.
• A relatively old fleet of vehicles that was used as a base to develop fuel consumption and emission models in CMEM [11]. While this should not affect the relative ranking of multiple routes, we can certainly say that the impacts of hybrid and electric vehicles were not considered. In future research, more recent emissions models should be considered.
• Inability to perfectly match (ever) the field conditions in the simulation environment.
• In our study, the localized, short-distance routes within the campus network may not provide sufficient differentiation in emissions to make Google’s eco-routing method as effective as it might be over longer distances. Indeed, as our tables suggest, the variations in emissions are minimal at this scale, which could make fuel consumption the primary criterion influencing route selection.

Thus, the best conclusion that we can reach is that our study shows some worrying inconsistencies between the two approaches, which should be further investigated, hopefully with a more sophisticated approach, in the future.

6. Conclusions

This paper demonstrates how traffic microsimulation can be applied to identify eco-routes in urban street networks. A set of origin–destination paths, with eco-routes recommended by Google Maps, were generated within a campus network to compare with those mimicked by a Vissim simulation model. The main criterion for identifying and comparing the eco-routes was fuel consumption, but emissions such as CO2, NOx, CO, and HC were also considered.

The experimental results suggest significant inconsistencies between the eco-routes recommended by Google Maps and those resulting from the microsimulation approach. Proper matches were found in only about one-third of all cases. Further experiments to test whether better realignment of field and simulation travel times might improve the matching of the eco-routes did not yield more meaningful results.

These findings can open new avenues of investigation, such as the possibility of replicating (and incorporating) Google’s approach (if publicly available) into similar experiments in the future. Additionally, more sophisticated comparisons could be made with field data from probe cars, although such an approach would also be more expensive. In any case, future research is almost certainly needed to further validate the Google Maps methodology for eco-routes, especially for traffic conditions in heavily urbanized areas.