GreenNav: Spatiotemporal Prediction of CO₂ Emissions in Paris Road Traffic Using a Hybrid CNN-LSTM Model

Abstract

In a global context where reducing the carbon footprint has become an urgent necessity, this article presents a hybrid CNN-LSTM prediction model to estimate CO₂ emission rates of Paris road traffic using spatio-temporal data. Our hybrid prediction model relies on a real-time road traffic database that we built by fusing several APIs and datasets. In particular, we trained two specialized models: a CNN to extract spatial patterns and an LSTM to capture temporal dynamics. By merging their outputs, we leverage both spatial and temporal dependencies, ensuring more accurate predictions. Thus, this article aims to compare various strategies and configurations, allowing us to identify the optimal architecture and parameters for our CNN-LSTM model. Moreover, to refine the predictive learning evolution of our hybrid model, we used optimization techniques like gradient descent to monitor the learning progress. The results show that our hybrid CNN-LSTM model achieved an R² value of 0.91 and an RMSE of 0.086, outperforming conventional models regarding CO₂ emission rate prediction accuracy. These results validate the efficiency and relevance of using hybrid CNN-LSTM models for the spatio-temporal modelling of CO₂ emissions in the context of road traffic.

1. Introduction

The increase in greenhouse gas (GHG) emissions in recent years has significantly contributed to the global warming of our planet. According to the International Energy Agency, transportation is the second most responsible sector for GHG emissions [1]. Greenhouse gases are a combination of several gases: carbon dioxide (CO₂), nitrous oxides (N₂O), methane (CH₄), and hydrofluorocarbons (HFCs). These gases are the primary GHGs emitted by automobiles in road traffic [2,3]. Transportation activities produce more than 75% of the total environmental CO₂ emissions [3,4].

Consequently, reducing transportation emissions has become a global goal to combat climate change [3,5]. Moreover, excessive traffic can have various undesirable impacts, including increased noise and gaseous pollution. Additionally, humans exposed to high amounts of CO₂ have a significantly increased likelihood of contracting various diseases and conditions, such as cancer, heart disease, respiratory problems, and premature births [3,6,7,8,9]. Therefore, estimating and predicting vehicle CO₂ emissions is essential to evaluating and mitigating these health and environmental impacts. Hence, several models can be used to anticipate CO₂ emissions from transportation [10,11,12,13]. Early traffic CO₂ models are based on traditional approaches that primarily rely on data sampling and technologies, such as GPS data. These models simulate the distribution of traffic CO₂ emissions in a region by combining thematic maps and vehicle emission equations [3,11,14].

Two main approaches are commonly used to estimate CO₂ emissions in road traffic: the top-down approach and the bottom-up approach [12,13]. The top-down approach uses aggregate data, such as fuel consumption or total traffic volume, to calculate emissions at a national or regional scale. While this method allows for a quick analysis of overall trends, it often lacks precision when capturing local details or smaller areas. On the other hand, the bottom-up approach relies on more detailed data, such as individual vehicle behavior or specific trip conditions. By using granular traffic and vehicle data, this method provides more accurate and localized estimates of CO₂ emissions, making it useful for targeted policies and urban-scale analyses. Moreover, recent approaches have introduced additional variables to refine emission estimates further. These variables include vehicle speed, vehicle-specific power (VSP), acceleration, real-time traffic congestion, weather conditions, driving behavior, road gradient, and historical CO₂ release rates. These newer methods enable a more comprehensive analysis of the factors influencing CO₂ emissions, resulting in more accurate modeling tailored to the dynamic conditions of an urban road network [11,12,15].

In our research [16,17,18], we have developed a solution capable of predicting CO₂ emissions by considering various road traffic parameters and spatio-temporal conditions to address this sustainable development challenge.

Regarding the spatial representation of our data, road traffic can be modeled on a spatial map, as road traffic data are linked to the space they occupy. For instance, congested roads have higher emissions due to abrupt speed variations and frequent stops/restarts. This gives us an indication of the spatial occupancy of cars on a map.

As for the temporal representation of our data, road traffic can also be modeled in a temporal context. Road traffic varies according to the times of the day and days of the week, which affects CO₂ emission levels. For example, during peak hours, there is often a significant increase in traffic congestion, leading to higher emissions. This temporal variation is crucial for understanding and predicting CO₂ emissions, reflecting changes in travel behaviors and traffic conditions.

By combining these two dimensions, spatial and temporal, we can obtain a comprehensive and accurate understanding of the factors influencing CO₂ emissions in road traffic. In our article, we address the following question: How can we predict CO₂ emission rates for each street in Paris in road traffic using a predictive AI model?

We proposed a deep learning, hybrid CNN-LSTM prediction model to capture CO₂ emission variations based on spatio-temporal data to achieve this goal. Our model uses a convolutional neural network (CNN) [19,20] to extract features from spatial data and a long short-term memory (LSTM) network [21,22] to model temporal sequences. Our research methodology is organized as follows:

First, we describe the architecture of our hybrid CNN-LSTM model. We detail the hyperparameters used, such as the number of layers, filter size, learning rate, and activation functions, as well as the data normalization process and the training method. Next, we evaluate our model’s performance using various analysis techniques to monitor its learning. One method used is gradient descent, which optimizes our model’s configuration parameters. Additionally, dashboards compare different model configurations and adjust hyperparameters accordingly. Finally, we deploy our model on our web application, GreenNav [18], which allows us to test our prediction model using visualization tools such as 2D and 3D maps and graphs. GreenNav aims to provide an interactive platform to explore and analyze CO₂ emission forecasts in urban road traffic, thus contributing to more sustainable and ecological traffic management solutions.

2. State of Art

In this section on the state of the art, we compare the different models for predicting CO₂ emissions in road traffic as presented in the following articles: [3,23]. These models all share the same objective: predicting the rate of CO₂ emissions in the context of road traffic. Thus, we analyze the types of data used as input, the model architectures, the estimation scales, the models’ accuracy, and the quantity and sources of the data. This comparative analysis allows us to better understand each approach’s advantages and disadvantages and position our CNN-LSTM model in this context.

2.1. Hybrid Models for Traffic Prediction

Before focusing on models specifically designed for CO₂ emissions prediction, it is essential to recognize recent studies utilizing hybrid models for traffic flow prediction. Accurate traffic modeling is directly correlated with CO₂ emissions, and hybrid models, combining statistical approaches with deep learning techniques, have shown great potential for capturing complex traffic dynamics.

For instance, Pan et al. (2024) proposed a hybrid model [24], the FD-Markov-LSTM, which integrates a fundamental diagram (FD) to identify traffic states, a Markovian model to capture probabilistic transitions between these states, and an LSTM network to capture residual temporal dependencies not accounted for by traditional approaches. This model was validated on traffic data from Beijing and Los Angeles, demonstrating high accuracy in different congestion scenarios. Similarly, Pan et al. (2022) developed a CNN-LSTM hybrid model that integrates traffic flow models with deep learning to predict real-time traffic on complex urban highways [25]. These models highlight the effectiveness of hybrid approaches for traffic prediction, which can be extended to CO₂ emissions prediction due to the strong relationship between traffic conditions and vehicle emissions. Hence, the strong performance demonstrated by these hybrid models for traffic flow prediction can be leveraged to refine CO₂ emissions estimation, given the close correlation between traffic dynamics and greenhouse gas production. However, while these models effectively capture complex spatiotemporal variations in traffic, their initial design does not systematically consider environmental parameters linked to CO₂ emissions (e.g., fuel consumption, vehicle types, and weather conditions). Therefore, in the following sections, we focus on models specifically developed for CO₂ emissions prediction, incorporating additional environmental parameters and precise traffic characteristics to improve the granularity and robustness of the estimates.

2.2. Comparative Analysis of CO₂ Emission Prediction Models

In ref. [3], Al-Nefaie and Aldhyani developed deep learning models to predict vehicle CO₂ emissions using detailed Kaggle [3] data, covering seven years and including 7385 vehicle records in Canada. LSTM and BiLSTM models managed temporal dependencies and accurately predicted CO₂ emissions. The model’s performance was evaluated using statistical metrics such as MSE, RMSE, R%, and R², with the BiLSTM model outperforming the LSTM in terms of accuracy and reliability.

Table 1. Review of Existing Artificial Intelligence Models for Predicting CO₂ Emissions: Approaches and Methodologies.

Models	Dataset Description	Model Architecture and Scaling Approach	Number of Data	Model Accuracy	Complexity and Spatio-Temporal Capacity
[3]	Data from Kaggle: vehicle class, engine size, etc.	LSTM and BiLSTM Macro/ Top-down	12 columns, 7385 rows over 7 years	LSTM: R2 = 77.81% BiLSTM: R2 = 93.78%	Medium/No
[23]	Data from World Bank, Turkish Statistical Institute, energy consumption data, GDP, etc.	MLP, XGBoost, SVM/ Macro/ Top-down	47 years of data (1970–2016)	MLP: R2 = 0.9689 XGBoost: R2 = 0.9886 SVM: R2 = 0.9886	Medium to High/No
[16,17,18]	Data from OpenDataSoft API, Google API, AirParif: geolocation, traffic flow, speed, etc.	CNN-LSTM Micro/Bottom-up	48 GB of data, 8 million rows over 4 months	CNN-LSTM: R2 = 0.91 RMSE = 0.29	High/Yes
[26]	Data from Data on urban governance elements in China: population, road network density, etc.	Random Forest/Macro/Top-down	Data from 1903 Chinese cities in 2010, 2012, 2015	R2 = 0.9431	Medium/No
[15]	PEMS and GPS data for LDDTs in China	LSTM/Micro/Bottom-up	Real-time data for 2 vehicles on different routes	R2: 0.986–0.990 RMSE: 0.165–0.167	Medium to High/Yes

summarizes the comparative analysis of existing artificial intelligence models for predicting CO₂ emissions, including their datasets, methodologies, and performance metrics.

In ref. [23], Çınarer et al. (2024) applied three different artificial intelligence algorithms (multi-layer perceptron (MLP), extreme gradient boosting (XGBoost), and support vector machine (SVM)) to estimate CO₂ emissions in the transportation sector in Turkey. The input parameters considered are energy consumption (ENERGY), vehicle kilometers traveled (VK), population (POP), year (Y), and gross domestic product per capita (GDP). The algorithms were tested under four scenarios based on the effect of correlation: scenario 1 (ENERGY/VK/POP/Y/GDP), scenario 2 (ENERGY/VK/POP/Y), scenario 3 (ENERGY/VK/POP), and scenario 4 (ENERGY/VK). The models’ performance was evaluated using statistical indicators (R², RMSE, MSE, and MAE), with the XGBoost algorithm offering the best performance in scenario 4 [23]. The AI algorithms effectively estimate CO₂ emissions, which has significant implications for policymakers and stakeholders.

In refs. [16,17,18], our GreenNav model uses a dataset from multiple APIs to create a comprehensive database for our analysis. The data sources include OpenDataSoft and Google API. The data considered include street names and geolocation, road traffic flow, occupancy rate, date and hours for OpenDataSoft, road distance, average speed, and speed limits for Google API. Based on a CNN-LSTM architecture, the model was applied at a micro-scale, i.e., by street. The study was conducted in the city of Paris, and four months of data covered all the roads in the town, representing a total of 48 gigabytes of data, i.e., 8 million rows. The performance of the CNN-LSTM model was evaluated with a coefficient of determination (R²) of 0.91.

In ref. [26], Zhang et al. (2023) developed a random forest model to predict CO₂ emissions in Chinese cities using data on population, road network density, and other urban governance elements. The data covered 1903 cities for 2010, 2012, and 2015. The model was compared to traditional approaches such as linear regressions. The random forest model performed better, with a coefficient of determination R² of 0.9431.

In ref. [15], Li et al. (2024) developed an LSTM model to predict CO₂ emissions from light-duty diesel trucks (LDDTs) using PEMS and GPS data. The model considers vehicle-specific power (VSP), speed, road gradient, and acceleration. It achieved R² values between 0.986 and 0.990, with an RMSE between 0.165 and 0.167.

2.3. Limitations and Weaknesses of the Analyzed Models

The analyzed models present several limitations that affect their ability to provide generalizable and accurate predictions of CO₂ emissions in diverse contexts. The DL-DTCEM model [15] focuses on light-duty diesel trucks, using specific data from PEMS and GPS systems, but its generalization is limited to predefined routes and specific vehicle configurations. It does not account for other vehicle types or more complex road conditions (urban, mixed highway routes, etc.), restricting its applicability. The MLP, XGBoost, and SVM models [23] rely on macroeconomic variables (energy consumption, GDP, population, etc.) and perform well only in scenarios where the input data are consistent and homogeneous. However, these models fail to capture local and temporal traffic dynamics, and their dependence on macroeconomic aggregates makes it challenging to use them for predictions at a micro level, such as individual road segments. The random forest model [26] effectively identifies emissions trends based on urban governance factors (city size, economic development, etc.). Still, it is unsuitable for fine-grained predictions at the road segment level due to the absence of specific traffic-related parameters such as traffic flow or speed. Finally, the BiLSTM model [3] effectively models static time series by capturing temporal dependencies. Still, it cannot incorporate spatial variations or interactions between road segments, limiting its accuracy in dense urban environments where the road network structure plays a crucial role.

3. Research Methodology

3.1. Data Structure

Before detailing the development of our hybrid prediction model, it is essential to present the dataset structure used for training our model. Proper structuring is a critical element of our quest to obtain reliable predictions. Our dataset includes vehicle specifications, traffic, and environmental conditions. Figure 1 below illustrates the composition of this dataset:

• The first source is OpendataParis [27], an API that provides real-time information about Paris’s roads. These data include road names, locations, traffic flow, occupancy rate, road traffic conditions, and recording dates, offering an overview of the traffic status;
• The second data source is Google API [28], which provides additional important information such as road distances, average speeds, and speed limits, adding an essential dimension to our dataset;
• The third component is GreenNav Algo, our probabilistic algorithm for calculating the CO₂ emission rate for each road in Paris. This algorithm allows us to generate precise estimates of CO₂ emissions based on real-time traffic conditions;
• Finally, we use data from the Airparif API [29] and the Getambee API [30], which provide information on air quality in Paris. Additionally, the statistical algorithm developed by ADEME integrated into the Airparif platform allows us to calculate the CO₂ emission rate between two specific positions [31]. By combining these different sources, we built a rich and diverse dataset essential for the training and performance of our hybrid CNN-LSTM model.

Figure 2 is an example of our integrated dataset after merging all sources. This sample corresponds to Boulevard de Belleville at 3:00 p.m. on 1 January 2024.

After collecting and structuring our data, the next step was to assign a reference CO₂ emission value to each road segment. Each segment is thus considered as an individual CO₂-emitting unit. To this end, we developed a probabilistic algorithm that estimates, on an hourly basis, the CO₂ emissions produced by every segment. This estimation is based on traffic data, vehicle characteristics, and various emission factors. The following equation represents the modeling:

$$E_{segment}=D\times \sum _{i=}^N{FE}_i$$

(1)

where

$E_{segment}$ is the total CO₂ emissions for the segment (in g/h).

$D$ is the segment distance (in km).

$N$ is the number of vehicles passing through the segment during the given period.

${FE}_i$ is the emission factor of the ith vehicle, encompassing parameters such as the vehicle’s environmental class, maintenance status, driving behavior (aggressive or not), air conditioning/heating use, and cold starts.

This calculation provides the hybrid CNN-LSTM model with a coherent reference dataset of CO₂ emissions per segment by incorporating these emission factors and traffic conditions. These values serve as a training base, allowing the model to establish correlations between traffic characteristics and emission intensity, thereby improving the accuracy of its predictions.

3.2. Description of the Hybrid CNN-LSTM Model

In this section, we detail the use of a hybrid CNN-LSTM model for predicting CO₂ emission rates. This model was specifically designed to address the unique challenges of emission forecasting in complex urban environments, where CO₂ emissions depend not only on the geographical characteristics of road segments but also on the temporal evolution of traffic.

CNNs capture spatial dependencies between road segments, such as network density, congestion, and interactions between adjacent streets. In our study, CNNs help model the local characteristics of roads based on their geographical position, providing a fine-grained representation of the spatial structure of the road network. However, CO₂ emissions also vary depending on traffic fluctuations over time, which requires accounting for temporal dynamics. LSTMs were therefore integrated to capture these temporal dependencies, modeling traffic variations, congestion periods, and changes in driving patterns across different periods (daily, weekly, etc.).

The CNN-LSTM hybrid addresses these challenges by simultaneously processing spatial and temporal information. This ability to capture spatio-temporal dynamics is crucial for predicting CO₂ emissions, as it allows us to model the complex variations from one road segment to another while considering changes over time. Unlike a sequential approach, where a CNN is used for spatial feature extraction followed by an LSTM for temporal modeling, our hybrid architecture jointly exploits spatial and temporal relationships at each learning stage. This integration captures cross-interactions between local characteristics (spatial dependencies) and sequential dynamics (temporal dependencies), thereby improving the granularity and accuracy of predictions.

In the following Research Methodology Section, we demonstrate through experimental scenarios that this integrated architecture outperforms the performance obtained by sequentially using separate CNN and LSTM models. These scenarios show that using two separate models leads to information loss during the transition from spatial features to temporal modeling, impacting the quality of the final predictions. In contrast, the hybrid model optimizes information transfer across layers by processing the data simultaneously and cohesively, resulting in a more refined CO₂ emissions modeling and significantly reducing prediction errors.

To enhance this approach, we also integrate a generative AI technique [23], which uses initial CO₂ predictions to generate new input data. These new data, in the form of CO₂ maps, are then reinjected into the model to refine predictions further. This allows us to simulate different traffic scenarios and better capture the effects of traffic variations on CO₂ emissions at a fine scale. Figure 3 below illustrates how the initial predictions are used to generate these maps, which are then reintroduced into the model to improve estimation robustness.

Furthermore, we employed the root mean squared error (RMSE) as the loss function during model optimization. RMSE is derived by taking the square root of the average squared differences between predicted and actual values, thereby providing an error measure that is in the same units as the target variable. By minimizing RMSE, we ensure that large deviations receive a stronger penalty and make the model’s performance more interpretable. This approach effectively guides the learning process towards more stable and precise CO₂ emission predictions.

In conclusion, combining the CNN-LSTM hybrid model with a generative AI [32] approach addresses the inherent challenges of CO₂ emission prediction in urban contexts. This model captures the complex interactions between road segments and temporal traffic variations, offering fine-grained and localized emission predictions essential for more robust and informed environmental analyses.

3.3. Structure of the Research Methodology

Our article addresses the following problem: How can we predict the CO₂ emission rate for each street in Paris under road traffic conditions using an AI model? To achieve our objective, we organized our research pipeline as explained below and illustrated in Figure 4.

• Phase 1 Data Preprocessing: The first step in our methodology involves preparing and normalizing our dataset to ensure better learning by our models. We divided our dataset into two distinct sets: a temporal set and a spatial set:

  (a)

  Temporal Set for LSTM model: Contains temporal variables such as hour_sin, hour_cos, weekday_sin, weekday_cos, traffic flow (q), occupancy rate (k), distance, and traffic indicators (etat_trafic_*). These data allow modeling the temporal variations in CO₂ emissions;

  (b)

  Spatial Set for CNN model: Includes spatial variables such as location_start_x, location_start_y, location_end_x, location_end_y, traffic flow (q), occupancy rate (k), distance, and traffic indicators (etat_trafic_*). These data capture the spatial relationships between different road segments. By separating the data into temporal and spatial sets, our two models can specialize in their respective domains, thus optimizing the learning process.

• Phase 2 Model Training: The second phase of our methodology involves training our models:

  (a)

  CNN Model: Designed to extract complex spatial features from our dataset;

  (b)

  LSTM Model: Designed to handle temporal dependencies in the data series.

  We then divided our dataset into training and testing sets to feed our data into the training and testing phases of our hybrid model:

  (a)

  Training Set: X_train_temporal, X_train_spatial, y_train;

  (b)

  Testing Set: X_test_temporal, X_test_spatial, y_test.

  To leverage both models, we merged the outputs from each model to obtain a final prediction that combines spatial and temporal information. During the training phase, we relied on MSE as the loss function to evaluate and guide the optimization process of our model parameters. By systematically minimizing MSE, we ensure that our model continuously refines its predictions, converging towards a solution that better aligns with actual CO₂ emission values.
• Phase 3 Model Testing and Evaluation: After training our models, the final phase is testing their performance on a test dataset. These test data are randomly selected from the dataset to evaluate the model’s generalization. In this context, our task is a regression problem rather than a classification problem, so using a traditional accuracy metric (commonly employed in classification tasks) is inappropriate. Instead, we adopted the coefficient of determination (R²) as our main evaluation metric [33,34]. R² is a standard measure for regression models that quantifies how well the predicted values approximate the real data. It represents the proportion of variance in the dependent variable that is predictable from the independent variables. Additionally, we employed statistical visualization tools to analyze learning errors and adjust our model accordingly [35,36,37]. This iterative process allows us to continually loop back to previous steps to enhance our model’s performance.

3.4. Description of Our Hybrid CNN-LSTM Model Architecture

Our hybrid model architecture for predicting CO₂ emissions in Paris traffic leverages the strengths of both CNN and LSTM networks, enabling the effective capture of complex interactions between spatial and temporal features, resulting in more accurate predictions of CO₂ emissions in road traffic. Figure 5 below illustrates how temporal and spatial features are separated, processed independently, and merged to obtain a final prediction.

• Temporal features include time, day of the week, traffic flow (q), occupancy rate (k), distance, and other traffic indicators (etat_trafic_*). An LSTM model processes them. This model consists of an input layer that takes sequences of variable length, followed by three successive LSTM layers with 50 units each, where the first layer returns the entire sequence, and the third layer returns only the last output. Each LSTM layer is followed by a dropout layer to reduce overfitting;
• Spatial features include start and end coordinates (location_start_x, location_start_y, location_end_x, location_end_y) and other spatial indicators. They are processed by a dense model simulating a CNN for vectors. This model starts with an input layer taking a feature vector, followed by a dense layer with 128 units and ReLU activation, another dense layer with 64 units and ReLU activation, and a dropout layer to prevent overfitting.

The outputs from the LSTM and dense models are then merged using a concatenation layer, and the final prediction of CO₂ emissions is obtained through a final dense layer. This architecture effectively captures complex interactions between temporal and spatial features, thus offering a more accurate and robust prediction [3]. To ensure optimal training quality, we performed data normalization and missing value imputation prior to model training. These preprocessing steps reduce scale discrepancies between variables, properly capture temporal cycles, and ensure a complete and consistent dataset for the model:

• Data Normalization: We normalized key quantitative variables (e.g., traffic flow q, occupancy rate k, and distance) and spatial coordinates as well as transformed temporal features (hour, day) into sinusoidal representations. This approach prevents any single variable from disproportionately influencing the training process and enhances the model’s stability and convergence;
• Missing Value Imputation: The initial dataset contained missing values, particularly for traffic flow and occupancy rate. We implemented a multi-tiered imputation strategy that combines linear interpolation, multivariate methods (MICE), and road network-based inference. This approach reconstructs coherent time series while preserving spatial and temporal consistency, ensuring that data gaps do not bias the model.

These preprocessing measures result in a more balanced, complete, and homogeneous dataset, thereby improving the accuracy and robustness of our CO₂ emission rate predictions. In conclusion, this hybrid architecture leverages the strengths of LSTM networks to capture temporal dependencies and CNN (dense) networks to extract spatial features, allowing for more precise modeling of CO₂ emissions in an urban traffic context.

3.5. Testing and Performance Analysis of Our Hybrid CNN-LSTM Model

3.5.1. Introduction of Optimization Methods in Our Cyclic Optimization Phase

This section presents the various steps that allowed us to adjust our hybrid CNN-LSTM model during its training. Through these progressive adjustments, we improved the accuracy of our model from 61% to 91%. We underwent several modification, comparison, and evaluation cycles before achieving this optimal result. To achieve this, we conducted a comparative study on the several following aspects of our architecture:

• Data Normalization: Data normalization is a crucial step in enhancing the learning of our CNN-LSTM model. We identified several important features to normalize:

  (a)

  Quantitative Features: Features such as traffic flow (q), occupancy rate (k), and distance were normalized to ensure they were on a comparable scale. This allowed us to reduce the magnitude of differences between these features, facilitating the learning process of our model;

  (b)

  Temporal Features: Temporal data, particularly hours and days of the week, were transformed into their sinusoidal representations (sine and cosine) to capture daily and weekly cycles. This transformation enabled the model to understand better and predict the temporal variations in CO₂ emissions;

  (c)

  Location Coordinates: The start and end coordinates (location_start and loca-tion_end) were normalized to ensure a uniform scale. This allowed the model to process location data more effectively, reducing the effects of different geographical scales;

  (d)

  Categorical Features (traffic state and day type): We also applied categorical encoding to certain variables for better interpretability. The traffic state (etat_trafic) was classified into four discrete levels representing different road traffic conditions: 1 for fluid traffic, 2 for pre-saturated traffic, 3 for saturated traffic, and 4 for blocked traffic. We also introduced a binary feature indicating whether the data point was recorded on a weekday or weekend. This categorization provided the model with structured information about traffic intensity and temporal context, enhancing its ability to learn complex patterns related to CO₂ emissions.

Normalizing these data had a significant impact. It improved the model’s stability and convergence speed during training, leading to better overall performance. Due to the different scales and feature variations, the model would have struggled to learn consistently without this normalization.

• Model Hyperparameters: We targeted the hyperparameters of our model to improve accuracy while avoiding overfitting. During this phase, we monitored the loss function value throughout the learning process and adjusted the following parameters:

  (a)

  Number of Epochs: Optimizing the number of epochs was crucial to avoid overfitting. Too many epochs can lead to overfitting the training data, while too few can result in underfitting;

  (b)

  Number of Layers: Adding additional layers allowed us to capture more complex features from the input data, enhancing the model’s representation capability;

  (c)

  Dropout Layers: We integrated dropout layers to reduce overfitting by randomly turning off specific neurons during training. This technique improved the model’s generalization capability;

  (d)

  Number of Units per Layer: Adjusting the number of units in each layer helped us balance representation capacity and generalization ability, thereby optimizing the model’s overall performance.

By implementing these optimization methods, we significantly enhanced the accuracy and robustness of our hybrid CNN-LSTM model, enabling more precise and reliable predictions of CO₂ emissions in Paris traffic.

3.5.2. Comparison of Results from Different Model Configurations

We herein present the training and testing results of our respective models. Through these comparisons, we show the hyperparameters and architecture of each version of our model, which allowed us to optimize performance and achieve an accuracy of 91%.

Configuration 1

In the first version of our model, we divided our dataset into two sets: temporal and spatial. These sets were, respectively, fed into distinct models for training. The model architecture is presented in the Figure 6 below:

In this configuration, we trained our model for 14 epochs. The model achieved an accuracy of 51%, which needs to be revised for a reliable prediction model. By monitoring the loss function during training (see Figure 7 below), we observed that the model had yet to finish learning by the end of the last epoch. The loss curve shows that the loss function was still not stabilized, suggesting that the model required more epochs or further hyperparameter adjustments to converge correctly.

The analysis of our model’s prediction results (see Figure 8 below) shows that the predictions did not align well with the actual values in our test set. This indicates that our model needed more training and had yet to find the correlation between the dataset and the target, the CO₂ emission rate for each street in Paris. Despite a promising initial architecture, this first configuration’s performance was limited. The low accuracy and poor alignment of predictions with actual values highlighted the need for continued model optimization.

The graphs (a) and (b) in Figure 8 show that our model’s predictions did not follow the actual values, suggesting that the model needed to learn more from the data. Subsequent adjustments to the architecture and hyperparameters were necessary to improve these initial results. In conclusion, Configuration 1 helped us understand the initial limitations and identify areas of the model needing improvement. The following steps involved exploring more complex configurations and refining parameters to achieve better performance.

Configuration 2

We proposed a new configuration in this second phase, retaining the model architecture from Configuration 1 but modifying the hyperparameters and input data. First, we normalized certain data to help the different layers of our model better understand the correlations between traffic data and CO₂ emission rates. For example, for temporal data, we used cosine and sine normalization functions to represent the hours and days of the week, providing more structured information to our model.

As shown in Figure 7, the previous model had yet to finish learning and had yet to reach the local minimum. Therefore, we increased the number of training epochs to allow more profound knowledge and better identification of correlations between the data. The results showed an improvement with an accuracy of 61% after making these modifications, as illustrated in Figure 9 and Figure 10. However, although the results were better, they needed to be more satisfactory. By analyzing Figure 10, which compares predicted values to actual values, we found that the model’s predictions could have more perfectly followed the ideal prediction line (red line). The dispersion of points around the prediction line indicated that the model had yet to correctly capture all relationships between the input variables and the target. In other words, the predicted values did not match the actual values well enough, suggesting that the model still required adjustments to improve accuracy and robustness.

The points in Figure 10 show that although the predictions generally followed the trend of the actual values, significant variation was still needed in the model. This variation indicates that the model had yet to thoroughly learn the complex relationships between the input data and CO₂ emission rates. Consequently, we decided to repeat the process of optimizing our model’s architecture in a new cycle to improve performance further. The following adjustments focused on finetuning hyperparameters and model structure to reduce prediction dispersion and align predicted values more closely with actual values. In conclusion, although Configuration 2 showed significant progress over Configuration 1, the results indicate that further adjustments were necessary to achieve optimal accuracy. The following optimization cycle aimed to refine the model’s architecture and hyperparameters to achieve even better performance.

Configuration 3

In this final configuration, our model fully leverages the complementarity between temporal and spatial dimensions. The temporal branch now consists of three sequential LSTM layers: two initial layers with 100 units each, both accompanied by dropout layers (rate 0.3) to mitigate overfitting, followed by a third LSTM layer with 50 units. This arrangement enables a more nuanced analysis of traffic sequences, taking into account daily and weekly variations, and thereby capturing longer-term trends.

On the spatial side, the architecture includes two dense layers: The first, with 128 units, employs a ReLU activation and L2 regularization (0.01), followed by a dropout layer (rate 0.5) to further prevent overfitting. The second dense layer, with 64 units, also uses ReLU activation and L2 regularization (0.01). Together, these layers provide a more robust and stable spatial representation of the road segments.

In parallel, the data were structured to support sequential learning. Samples were organized into coherent temporal segments (e.g., weekly intervals), ensuring that the model receives temporally ordered data. Temporal variables were formatted into sequences, allowing the LSTM layers to identify recurring patterns over time. Additionally, a thorough normalization of the input variables ensured that traffic-related measures (such as flow and occupancy) share comparable scales, facilitating the model’s convergence and enhancing the quality of feature extraction.

Finally, the outputs of the temporal and spatial branches were merged via a concatenation layer, and a single-unit dense layer produces the CO₂ emission prediction. By incorporating these improvements additional LSTM layers, enhanced regularization of dense layers, temporally coherent data structuring, and rigorous normalization, our model achieves more stable convergence and significantly improved predictive accuracy.

With the final model configuration (Figure 11), we significantly improved predictive performance. As shown in Figure 12 below, the learning curve demonstrates that the model’s training loss stabilized around the 30th epoch, indicating that the network effectively converged. This stable convergence reflects a better balance between complexity and generalization, allowing the model to capture the intricate spatio-temporal patterns of CO₂ emissions in urban traffic. The resulting R² score of approximately 91% and an RMSE of 0.086 highlights the robustness and accuracy of the model in approximating actual emission values.

A more detailed evaluation is presented in Figure 13. In Figure 13a, the strong correspondence between actual and predicted CO₂ emissions across various road segments demonstrates the model’s effectiveness in capturing complex, traffic-related emission dynamics. In Figure 13b, the close clustering of points around the reference line indicates that the predictions align closely with observed values, confirming that the architectural enhancements and data preprocessing steps significantly improved predictive accuracy.

In summary, the enhancements to the LSTM and CNN components along with more effective data preprocessing and hyperparameter tuning led to a model that not only converges more efficiently but also produces highly accurate and reliable predictions of CO₂ emissions. This final configuration represents a substantial step forward in our approach to forecasting urban traffic-related emissions.

Overall Comparison of Different Versions of Our Hybrid Model

To evaluate the evolution of our hybrid CNN-LSTM model’s evolution, we provide a visual and tabular comparison of the three configurations. Figure 14 illustrates the training loss curves, showing the progressive improvement in the model’s convergence and accuracy. Additionally,

Table 2. Comparison of the Three Configurations of the Hybrid CNN-LSTM Model.

Config	Input Data	Data Preprocessing	Layer Architecture	Epochs and R2 Score
Config 1	Temporal: [’hour’, ’day’, ’q’, ’k’, ’distance’, ’etat_trafic’] Spatial: [’location_start_x’, ’location_start_y’, ’location_end_x’, ’location_end_y’, ’q’, ’k’, ’distance’, ’etat_trafic’] Target: CO2 emissions	- Quantitative_Features (q, k, distance) - Temporal_Features (hour, day) - Localization (location_start, location_end)	LSTM => Dense (see Figure 6)	Epochs: 14 R2: 52%
Config 2	- Same as Configuration 1	- Same as Configuration 1 - Additional normalization using sine/cosine transformations for temporal features (hours, days)	LSTM => Dense	Epochs: 40 R2: 61%
Config 3	Temporal: [’hour’, ’day’, ’weekday’, ’q’, ’k’, ’distance’, ’etat_trafic’] Spatial: [’location_start_x’, ’location_start_y’, ’location_end_x’, ’location_end_y’, ’q’, ’k’, ’distance’, ’etat_trafic’] Target: CO2 emissions	- Same as Configuration 2 - Categorical features (etat_trafic, weekday) - Sorted data by date and road segment - Weekly batching for temporal sequences	LSTM => Dropout => LSTM => Dropout => Dense => Dropout => Dense (see Figure 11)	Epochs: 40 R2: 91%

summarizes the key modifications made to the input data, data preprocessing, layer architecture, and hyperparameters for each configuration, alongside their respective impacts on predictive performance.

The training loss curves reveal a clear progression in the stability and optimization of the model. Each configuration builds upon the lessons learned from the previous iterations, refining the architecture and training approach to achieve better predictive accuracy.

The progression of our model configurations is summarized in Table 2, highlighting the input data, preprocessing strategies, model architectures, training durations, and corresponding R² scores. Each configuration reflects iterative enhancements aimed at improving the predictive performance of our hybrid CNN-LSTM model. Configuration 1 relied on basic normalization for quantitative, temporal, and spatial features but achieved only moderate accuracy. Configuration 2 introduced more advanced temporal transformations (sine/cosine) and extended the training epochs, resulting in modest improvements. Finally, Configuration 3 incorporated categorical features for traffic state and day type, organized data into weekly segments for more coherent temporal learning, and refined the layer architecture with additional LSTM and dense layers plus dropout and L2 regularization. These enhancements led to a substantially improved R² score of approximately 91%, demonstrating the effectiveness of a more comprehensive preprocessing regimen and a richer model architecture.

4. Experimental Scenario for CO₂ Emission Prediction Through the GreenNav Application

In order to demonstrate the practical utility and accuracy of our hybrid CNN-LSTM model within a real-world context, we integrated the prediction system into our GreenNav application. This integration allows users to visualize and compare CO₂ emissions for specific streets in Paris under various conditions. We present two distinct experimental scenarios to illustrate the model’s capabilities.

4.1. Scenario 1: Comparison of Predicted vs. Actual CO₂ Emissions

In the first scenario, Marc, a user of the GreenNav application, seeks to evaluate the accuracy of our CNN-LSTM prediction model by comparing predicted CO₂ emissions with actual values obtained from our baseline estimation algorithm. For this experiment, he selects Boulevard Saint-Germain in Paris, specifies 6 January 2023 as the date, and chooses 01:00 p.m. as a reference time for focused analysis (see Figure 15). To gain a more comprehensive overview, he also decides to retrieve emission data for all hours of the day, enabling a full 24-hour evaluation.

The GreenNav application queries the prediction model and retrieves both actual and predicted CO₂ emission values. These results are displayed through a comparative table (see

Table 3. Comparative table between predicted and actual values for the given scenario.

Hour	Flow (q)	Occupancy (k)	Actual CO2 (g/CO2)	Predicted CO2 (g/CO2)
00:00	517.0	3.9	15,932.5	15,895.2
01:00	323.0	2.8	10,143.7	10,120.5
02:00	212.0	1.4	6434.6	6400.1
03:00	142.0	0.8	4472.7	4450.3
04:00	104.0	1.0	4017.2	4000.9
05:00	183.0	1.5	5617.1	5600.6
06:00	431.0	3.7	13,350.9	13,310.8
07:00	849.0	7.7	26,319.9	26,300.9
08:00	1038.0	7.5	32,084.7	32,050.3
09:00	987.0	14.3	30,945.5	30,900.9
10:00	976.0	36.5	41,121.1	41,050.7
11:00	989.0	29.0	34,096.8	34,040.3
12:00	1129.0	24.8	38,812.5	38,750.6
13:00	1063.0	22.6	36,668.4	36,600.9
14:00	1093.0	28.5	38,598.0	38,500.1
15:00	1151.0	37.9	48,950.5	48,750.3
16:00	1147.0	35.2	46,788.4	46,700.8
17:00	1099.0	22.6	38,750.4	38,640.5
18:00	1149.0	28.4	40,071.1	40,000.1
19:00	1092.0	15.8	39,742.3	39,830.4
20:00	1185.0	14.8	36,328.2	36,250.9
21:00	870.0	5.6	27,290.5	27,250.6
22:00	771.0	6.6	23,569.2	23,500.5
23:00	888.0	6.5	27,229.4	27,200.1

) and an interactive graph (see Figure 16), facilitating a clear and comprehensive performance analysis. Key variables such as traffic flow (q), occupancy rate (k), and traffic state are also presented to provide additional context on emission-driving factors.

An analysis of the results reveals that the predicted values closely follow actual emissions throughout the day. For example, at 10:00 a.m., the actual CO₂ emission rate is 41,121.14 g/h, while the predicted value is 41,050.67 g/h, with a minimal error of 70.47 g/h, indicating a prediction accuracy of 99.83%. Similarly, at 03:00 a.m., the actual emission rate is 4472.72 g/h compared to a predicted rate of 4450.34 g/h, resulting in a difference of only 22.38 g/h, reflecting an accuracy of 99.50%.

The analysis further highlights a direct correlation between traffic flow and CO₂ emissions. For example, as traffic flow increases from 517 vehicles at midnight to 1151 vehicles at 03:00 p.m., CO₂ emissions rise accordingly. This strong relationship demonstrates the model’s ability to capture traffic dynamics and their impact on emissions accurately.

4.2. Scenario 2: Scenario for a Future Date

In this second scenario, Marc, a user of the GreenNav application, wants to estimate future CO₂ emissions for Boulevard Saint-Germain at a specific time. He selects the reference date and time on Wednesday, 10 July 2024, at 01:00 p.m. (see Figure 17). However, he is particularly interested in knowing the expected CO₂ emissions for 02:00 p.m. on the same day.

By entering the future date and time, GreenNav queries the hybrid CNN-LSTM prediction model to estimate CO₂ emissions based on projected traffic conditions. The model considers key traffic features such as traffic flow (q), occupancy rate (k), and traffic state, leveraging its spatio-temporal learning capabilities to produce an accurate prediction. The application returns the following result:

The model predicts that Boulevard Saint-Germain’s CO₂ emissions at 2:00 p.m. on 10 July 2024 will be 41,053.21 g CO₂e/h

5. Discussion: Model Advantages, Limitations, and Generalizability

• Model Accuracy: The CNN-LSTM hybrid model provides highly accurate predictions of CO₂ emissions, as evidenced by the slight difference between actual and predicted values. This performance is achieved by combining spatial-temporal features captured by the CNN with temporal sequences modeled by the LSTM, allowing the model to handle the complex dynamics of road traffic in Paris effectively;
• Impact of Traffic Parameters: The model effectively captures the effects of various traffic-related factors on CO₂ emissions, including traffic flow, occupancy rate, road location, congestion, and peak hours. The analysis shows a direct correlation between the increase in these factors and CO₂ emissions, accurately demonstrating the model’s ability to represent different traffic conditions, whether fluid or congested, accurately;
• Practical Applications: The model is ready for real-world applications, particularly in GreenNav, our ecological navigation application. It can provide route recommendations optimized based on real-time CO₂ emission forecasts, contributing to urban traffic management and emission reduction policies;
• Data Dependency: The model’s limitation is its reliance on a significant amount of precise spatiotemporal data to capture traffic variations. The quality and coverage of the training data directly impact the accuracy of the model’s predictions;
• Adaptability to Other Urban Environments: Although our model has been precisely calibrated and trained using data from the Paris road network, it is designed to be modular, allowing it to adapt to different urban contexts. We propose two complementary strategies to generalize its application to cities with diverse road network characteristics and traffic behaviors: transfer learning and federated learning:

  (a)

  Transfer learning [38] is beneficial when only a limited amount of data is available for the target city. In our context, the model’s convolutional layers (CNN), which capture the spatial features of the road segments, can be reused as-is since many cities share similar spatial patterns, such as road intersections and density. The LSTM layers, which capture temporal dependencies, can then be retrained to reflect the unique temporal dynamics of traffic in the new urban environment. For instance, if our model were to be applied to the city of Lyon, we could retain the CNN layers and only retrain the LSTM layers to adapt to Lyon’s specific temporal traffic patterns, such as peak hours and weekend behaviors. This approach would enable a quick adaptation to the local context without requiring extensive new data, thereby maintaining high performance on critical traffic segments;

  (b)

  On the other hand, federated learning [39] is more suitable for large metropolitan areas or cities that differ significantly from Paris regarding road network density, driving behavior, or vehicle types. In this scenario, the model is trained collaboratively using data from multiple cities while ensuring data privacy. Each city trains the model locally using its data, and only the model’s updated parameters are shared, allowing for creating a global model without transferring sensitive information. For example, if we wanted to adapt the model for New York City, factors such as road congestion, speed distribution on major highways, and seasonal variations could be incorporated through federated learning without exposing Paris or New York’s local data. This would allow the model to generalize its understanding across multiple urban environments, making it more resilient to varying geographical conditions;

• To enhance both the reproducibility and scalability of our approach, we recommend following a standardized protocol: replicate the preprocessing steps (normalization, imputation, and temporal transformations) for the new city, load the pre-trained CNN weights, and then retrain the LSTM layers using locally collected data. Ideally, these data should include detailed traffic measurements similar to those used in Paris, such as the number of vehicles per road segment and occupancy rates obtained from sensors (e.g., electromagnetic) and open data platforms. These traffic parameters are crucial for enabling the model to learn the relationship between traffic intensity and temporal variations (e.g., hours, days of the week, etc.), ensuring an effective adaptation to the specific urban context at hand. Publicly available scripts and configuration files can be provided to streamline this process. Additionally, employing dedicated frameworks (e.g., TensorFlow Federated) will facilitate the implementation of federated learning, allowing other researchers to easily reproduce and adapt our methodology to diverse urban environments while maintaining data quality and consistency;
• Computational Requirements and Comparison with Simpler Models: The CNN-LSTM model has higher computational requirements due to its hybrid structure, which combines convolutional layers to capture spatial relationships between road segments and LSTM layers to model temporal dependencies. In contrast, simpler models like MLPs or standard LSTMs handle only static correlations (MLP) or temporal sequences (LSTM), making the CNN-LSTM significantly more resource-intensive. This is because it processes a larger volume of data and multi-dimensional feature matrices, resulting in much higher parameters and more significant memory and computational capacity needs. This difference in computational costs makes the CNN-LSTM less suitable for environments with limited resources. However, the additional complexity is justified by a significant improvement in prediction accuracy, especially in urban scenarios where capturing spatio-temporal dynamics is crucial for accurately modeling CO₂ emissions. Thus, choosing a CNN-LSTM and simpler architectures will depend on the trade-off between available hardware resources and the required prediction accuracy;
• Future Improvements: Transfer learning mechanisms could enhance the model’s generalizability, allowing it to be adapted to different urban environments with minimal retraining. Techniques like federated learning also enable model sharing without transferring data between regions, increasing the model’s scalability across cities.

6. Conclusions

To address the challenge of predicting CO₂ emissions for each street in Paris, we developed a hybrid deep learning model combining convolutional neural networks (CNNs) [1] and long short-term memory (LSTM) networks [40,41]. This approach effectively captures the spatiotemporal variations in CO₂ emissions using data from multiple sources such as OpenDataSoft [27], Google API [28], and AirParif [27]. Our structured methodology, which includes preprocessing, training, and evaluation phases, allowed us to refine the model parameters and achieve an accuracy of 91%. The results show that our model outperforms traditional approaches, validating the effectiveness of the CNN-LSTM combination for CO₂ emission modeling in an urban environment. Integrating the model into the GreenNav application [18] enables users to predict CO₂ emissions accurately and in real time, contributing to more ecological decisions and better environmental management. To further improve our model’s accuracy and relevance in natural conditions, we plan to inject real-time CO₂ emission rates captured by our sensors. This approach will provide our model with updated data collected directly from the field. Indeed, it is crucial to differentiate CO₂ emission rates captured by sensors from those produced directly on the specific street. This distinction is necessary because factors like wind can shift CO₂ emissions from one street to another. By integrating real-time data, we can better understand and model these variations, enhancing the accuracy and reliability of our predictions.