A Dual-Stage Attention-Based Vehicle Speed Prediction Model Considering Driver Heterogeneity with Fuel Consumption and Emissions Analysis

Abstract

With the development of intelligent transportation systems (ITSs), personalized driving systems are receiving more and more attention, and the development of advanced systems cannot be separated from the practical exploration of drivers’ heterogeneous driving behaviors. An important foundation for subsequent driver-targeted research is how to mine the key influencing factors that characterize drivers through real driving data and how to appropriately classify drivers as a whole. This study took heterogeneous drivers as the object, based on a dual-stage attention-based vehicle speed prediction model, and carried out research on the speed prediction of traffic flow and the impact of fuel consumption and emissions in the car-following state considering the heterogeneity of drivers. Specifically, first, Spearman’s correlation analysis and K-means clustering were used to classify different types of drivers. Then, speed predictions for different types of drivers were separated via the dual-stage attention-based encoder–decoder (DAED) model and the prediction results between models and drivers were compared. Finally, the heterogeneous drivers’ fuel consumption and emissions were further analyzed via the VT-micro model. The results show that the proposed speed prediction model can effectively discriminate the influences of heterogeneous drivers on the prediction model, and the aggressive type presents the best effect. In addition, from the experiments on traffic fuel consumption and emissions, it can be concluded that the timid driver is the friendliest to the environment. By researching individual drivers’ driving characteristics, this study may help sustainable development in traffic management.

1. Introduction

With the continuous improvement in social and economic levels, residents’ living standards and travel needs are also increasing day by day, and the number of vehicles continues to grow. However, with the rapid concentration of the population, urban traffic problems, such as congestion, environmental pollution, and traffic accidents, have significantly affected traffic efficiency and urban development. To solve these problems, intelligent transportation systems (ITSs), which provide efficient transportation services and management, have attracted widespread interest [1]. Unlike building new infrastructure, which is subject to space constraints and high costs [2], ITSs can collect and process various data through smart infrastructure as well as advanced algorithms to improve traffic efficiency [3]. Specifically, characterizing traffic states by predicting future traffic conditions to optimize the routes of self-driving vehicles and taking these measures to preventively mitigate the effects of congestion and its associated problems may help improve the overall traffic environment [4].

Vehicle speed is an important reference factor for road traffic condition evaluation. Therefore, vehicle speed prediction has also become one of the research hotspots in ITS development and has important research significance. Speed prediction aims to estimate the driving speed of traffic participants over a future period based on the current and historical traffic conditions. Accurate short-term vehicle speed predictions can react to traffic conditions [5] and support traffic management, such as the optimized timing of signals and traffic resource allocation [6]. In addition, traffic participants can use the traffic conditions to plan their trips [7].

Currently, deep learning models have been proven to be effective at mining high-dimensional features of road speed data [8]. In terms of model methods, many scholars have improved the model prediction accuracy by improving the model structure or combining the advantages of multiple models, and the results have been significant. Traffic flow speed prediction is a typical time series prediction task. In recurrent neural networks (RNNs), long short-term memory (LSTM) networks, and gated recurrent units (GRUs), for example, the self-circulation mechanism can be effectively used to mine the long-term changing characteristics of road section driving speeds, thereby reducing prediction errors [9,10]. Based on the classical models, scholars have developed all kinds of combination models to improve the prediction accuracy in different traffic scenarios [11,12,13,14]. The feature-enhanced LSTM neural network prediction model with an attention mechanism (Attention) is proposed by Vaswani et al. This improved model is applied to California highway data. The results show that the average absolute error of the model is lower than those of other LSTM-based models [15]. Subsequently, more attention mechanisms combined with LSTM models are used to predict highway traffic [16]. Currently, in order to grasp the key data characteristics and further improve the prediction effects, the encoder–decoder structure is incorporated into the traffic speed prediction model. This type of model can be used to solve the problem of the difference between the input sequence length and the output sequence length in the prediction model. Douglas applies BiLSTM and LSTM layers to the encoder and decoder parts, respectively, for the prediction of vehicle speeds on road sections [17]. There are various methods for predicting vehicle speeds using deep learning models. However, finding the most fitting one in research is not so easy.

At the same time, driving is a complex project, and its safety is affected by various factors, including driver behavior, vehicle design, and the road environment. To date, promising results have been achieved in improving road safety and vehicle performance. However, due to the subjectivity of human factors, there is uncertainty in driver behavior, which has a specific impact on safe driving, making the study of driver behavior more complex than that of other influencing factors [18,19]. Different drivers show different driving characteristics. If all drivers are predicted as a whole, then the prediction results may be affected. Driver heterogeneity is mainly reflected in the heterogeneity of the drivers themselves and the heterogeneity between drivers. The heterogeneity among drivers refers to the differences and diversity among a group of drivers. Driver heterogeneity can be reflected in two traffic scenarios: car following and lane changing. Among them, car following is common, simple to implement, and a hotspot in this research direction [20]. In order to explore the driving characteristics of different drivers, it is of great practical significance to classify drivers with different behaviors. Regarding the issue of driver classification, previous studies have classified drivers into four categories based on the three observed types of trajectory anomalies and speed characteristics, which have inspired subsequent research [21]. One study found five acceleration-related factors to accurately classify driving styles [22], but its research object is public transportation, and the influencing factors are only related to acceleration. Chen et al. classified drivers into three categories, timid, average, and aggressive, based on headway, maximum acceleration, and maximum deceleration [23]. Khakzar used K-means clustering to classify drivers and further predict their vehicle trajectories [24].

Based on the analysis of traffic driving data, it is also of certain practical significance to explore the impacts of traffic on the environment further [25]. Among them, fuel consumption, carbon dioxide emissions, and so on are usually used to measure the degree of environmental impact of vehicles during driving. Based on the model input perspective and the input volume scale, traditional fuel consumption models can be divided into three categories: macroscopic, mesoscopic, and microscopic. The macro-level models mainly use the average speed of vehicles in a specific driving area to infer the total fuel emissions in the entire area. Among them, models such as the mobile source emission factor (MOBILE) [26] and the computer program for road traffic emission calculation (COPERT) [27] use average speed as a parameter of pollution characteristics. The meso-level models are between macro and micro models and usually involve driving patterns. These models are suitable for measuring fuel emissions over a local area, for example, a neighborhood or a road, over a day or an hour [28,29]. The micro-level models are suitable for estimating the instantaneous fuel emissions in seconds when a vehicle passes through a specific road segment or intersection. Barth et al. developed the comprehensive driving mode emission model (CMEM). Based on a microscopic perspective, they calculated the operating status parameters of the vehicle’s engine in different driving modes second by second, thereby inferring exhaust emissions [30]. The VT-Micro model, created by Rakha and Ahn in 2002, models the vehicle by dividing it into two situations: acceleration and deceleration. This model uses the logarithmic transformation method to simulate the logarithm of fuel consumption, the power function of speed, and the acceleration combination, which gives the model high prediction accuracy [31].

Based on the above analysis of the research backgrounds and status, this paper mainly researched the impact of driver heterogeneity on traffic flow speed prediction under car-following conditions. This study took heterogeneous drivers as the object, based on deep learning models, to predict the effective speed of heterogeneous drivers by carrying out a traffic model that considers driver heterogeneity in the car-following state and analyzed the impact of traffic fuel consumption and emissions. This article’s framework and the specific processes of experiments are shown in Figure 1.

There are three main points of innovation in this research:

• Driver heterogeneity and vehicle speed prediction are combined to analyze the different types of drivers instead of macroeconomic analysis. That means speed prediction not only focuses on the total traffic but also aims at individual drivers. Thus, relevant policies can be more target-oriented;
• Two attention mechanisms are used in an encoder–decoder framework to predict vehicle speed. Meanwhile, besides speed, other three-vehicle traveling data are also input to the prediction model, which can further improve the prediction accuracy;
• The accurate prediction results are further analyzed in terms of fuel consumption and emissions, which are combined with the actual traffic impacts that may arise to improve the practicality of the study.

This study holds several implications for sustainability goals across environmental, social, and economic aspects. As for ecological implications, accurate vehicle speed forecasting aids in optimizing traffic flow, reducing congestion, and minimizing idling time, thereby lowering vehicle emissions and improving air quality; as for social implications, forecasting vehicle speeds can enhance road safety by identifying high-risk areas, aiding in the implementation of measures to prevent accidents, promoting safer driving practices, and improve overall transportation efficiency, for example, on junction signaling prediction; as for economic implications, accurate vehicle speed predictions aid in designing more efficient road networks and transportation systems, optimizing infrastructure investments, and reducing long-term maintenance costs.

The structure of this paper is as follows. Section 2 introduces the methods used in this study, which involve driver classification, speed prediction, and fuel consumption and emissions. Section 3 uses the method proposed in Section 2 to display the results and analysis of each part based on data denoising processing. Finally, some concluding remarks are given in Section 4.

2. Methodologies

In this section, the methodologies used for this study are presented in three parts: driver classification, predictive modeling, and environmental evaluation modeling. The narrative is developed in the logical order of this study.

2.1. Driver Classification

Before classifying drivers, correlation data about car-following should be organized. The correlation coefficient is a statistical indicator that reflects the degree of correlation between variables. Its value range is [−1,1]; taking the value of 0 indicates no correlation, taking the value of [−1,0] indicates a negative correlation, and taking the value of [0,1] indicates a negative correlation. The two commonly used correlation coefficient methods are the Pearson correlation coefficient and the Spearman correlation coefficient. The Pearson correlation coefficient is mostly used to assess a linear relationship between two continuous variables. In comparison, the Spearman correlation coefficient is used to determine a monotonic relationship between two continuous variables. In a monotonic relationship, the variables tend to change together, but not necessarily at a constant rate. In this paper, a specific linear relationship between two variables is not necessary to research. Therefore, Spearman’s is more appropriate as a correlation analysis method. The Spearman correlation coefficient formula is expressed as follows:

$$\rho =\frac{{\displaystyle {\sum }_{i=1}^N(R(x_i)-\overline{R}(x))(S(y_i)-\overline{S}(y))}}{{\left[{\displaystyle {\sum }_{i=1}^N{(R(x_i)-\overline{R}(x))}^2}{\displaystyle {\sum }_{i=1}^N{(S(y_i)-\overline{S}(y))}^2}\right]}^{\frac{1}{2}}}$$

(1)

where $R_i$ and $S_i$ are the ranks of the values of the observations $i$, and $\overline{R}$ and $\overline{S}$ are the average ranks of the variables $x$ and $y$, respectively.

After obtaining several factors with weak correlations, the number of classifications must be determined before categorizing drivers. In this paper, the elbow method is used to determine the optimal number of categories. In selecting the number of categories, the elbow method plots a cost function for each number of categories. As the value increases, the number of samples included in each category decreases, so the samples are closer to their center of gravity, and the average distortion decreases. As the value continues to increase, the improvement in average distortion decreases. The elbow is the value corresponding to the location where the improvement in distortion decreases the most as the values are increased.

After determining the number of clusters, selecting an appropriate clustering method to categorize all drivers is necessary. According to the existing research, this paper summarized five commonly used clustering methods: Gaussian mixtures model, spectral clustering, K-means, Ward hierarchical clustering, and Balanced iterative reducing and clustering using hierarchies (BIRCH).

The clustering effectiveness of various methods can be measured by the silhouette coefficient. The silhouette coefficient can be expressed as the mean silhouette value over the entire dataset. The silhouette value is used to characterize the similarity of a target concerning the clusters in which the target is located concerning other clusters. It ranges from −1 to +1. The larger value indicates that the target has a higher degree of matching relationship with its cluster and a lower degree of matching relationship with other clusters.

2.2. Prediction Model

In order to achieve accurate vehicle speed prediction in the car-following driving environment while considering the impact of intra-driver heterogeneity, the model needs to capture the correlation between other feature-influencing factors and vehicle speed changes while ensuring the prediction accuracy is not affected by the long training time. Given the excellent performance of the current encoder–decoder framework on sequence prediction tasks [17,32], the ability of the attention mechanisms to extract focused, and the relevant information [33,34], this paper proposed a dual-stage attention-based encoder–decoder (DAED) model, which is based on the characteristics of vehicle speed prediction considering driver heterogeneity in the car-following state.

As shown in Figure 2, in the first stage (encoder), a new input attention mechanism is introduced, which can adaptively extract the relevant driving sequences for each time step by referring to the previous encoder’s hidden state. The extracted driving feature inputs are then further fed into the predictive neural network to perform the encoding operation at historical speeds. In the second stage (decoder), a temporal attention mechanism is used to select the relevant encoder hidden states for all time steps while encoding the output speeds in conjunction with the outputs of the previous moment to finally obtain the predicted speeds of the target vehicle. These two attention models are well integrated into an LSTM-based recurrent neural network and can be jointly trained using standard backpropagation. In this way, the model can adaptively select the most relevant input features and appropriately capture the long-term temporal correlation of the time series.

Module 1: Speed coding. This module encodes the prediction of the historical vehicle speeds based on the input drive sequences via a neural network. For time series prediction, given an input sequence, the encoder can be applied to learn a mapping from input to hidden state at the time step. In this paper, we use LSTM cells as a nonlinear activation function to capture long-term dependencies. Each LSTM cell has a memory cell in the time step $t$ state $s_t$, and access to the memory cell is controlled by three “gates”, namely, the forgetting gate $f_t$, the input gate $i_t$, and the output gate $o_t$. The update of an LSTM cell can be summarized in the following expression:

$$f_t=\sigma \left(W_f\left[h_{t-1};x_t\right]+b_f\right)$$

(2)

$$i_t=\sigma \left(W_i\left[h_{t-1};x_t\right]+b_i\right)$$

(3)

$$o_t=\sigma \left(W_o\left[h_{t-1};x_t\right]+b_o\right)$$

(4)

$$s_t=f_t\odot s_{t-1}+i_t\odot \tanh \left(W_s\left[h_{t-1};x_t\right]+b_s\right)$$

(5)

$$h_t=o_t\odot \tanh \left(s_t\right)$$

(6)

where $[h_{t-1};x_t]\in R^{m+n}$ is a cascade of the last hidden state $h_{t-1}$ and the current input $x_t$. $W_f,W_i,W_o,W_s\in R^{m\times (m+n)}$ and $b_f,b_i,b_o,b_s\in R^m$ are parameters to be learned. $\sigma$ and $\odot$ are logic functions and element-wise multiplication, respectively. The LSTM cell structure is shown in Figure 3.

Module 2: Input feature extraction. This module is used to adaptively select the relevant driving sequences. Given the input driving sequence, an input attention mechanism is constructed according to a deterministic attention model, a multi-layer perceptron, by referring to the hidden state and the previous moment in the LSTM cell state in the encoder. The input attention mechanism is a feed-forward network that can be trained jointly with other derivatives of RNNs, and these attention weights can be extracted adaptively to drive the sequence. Then, the hidden state at the time step can be updated. By utilizing the proposed input attention mechanism, the encoder can selectively focus on some specific driver sequences instead of treating all input feature sequences equally. The flow structure of the input attention mechanism is shown in Figure 4.

Module 3: Time series capture. This module is used to adaptively select the relevant encoder hidden states at all time steps. Specifically, each attention weight of the hidden state in the encoder at the time step is computed based on the previous decoder’s hidden state and the LSTM neuron state. The flow structure of this module is shown in Figure 5.

In this paper, minibatch stochastic gradient descent (SGD) and adaptive moment estimation (Adam) optimizer are used to train the model. In order to allow the model to quickly converge to the optimal solution in the early stage of training while preventing model overfitting in the late stage of training, the model learning rate starts from 0.001 and decreases by 10% after every 10,000 iterations. In addition, this article uses the mean squared error (MSE) as the standard backpropagation of the objective function to learn the parameters:

$${\left(y_T,{\widehat{y}}_T\right)}_{\mathrm{MSE}}=\frac{1}{N}{{\displaystyle \sum _{i=1}^N({\widehat{y}}_T^i-y_T^i)}}^2$$

(7)

In order to further reduce the model training error caused by the sample data slicing, we trained the model in this paper by combining k-fold cross-validation with predictive modeling. K-fold cross-validation is a technique commonly used to evaluate the performance of machine learning models, which helps us more accurately estimate the performance of the model on unseen data [35]. Using this method, the dataset is divided into k subsets, one of which is kept for validating the model, while the remaining k−1 subsets are used for training the model. This process is repeated k times, each time choosing a different validation subset, and eventually, the average of the k performance evaluation results is obtained. This method allows for a more accurate assessment of the predictive ability and reduces the risk of overfitting while also maximizing data utilization. Considering the length of data taken in this paper, training requirements, and literature references, k = 5 is determined to be applied for predictive model training. K-fold cross-validation (k = 5) is schematically shown in Figure 6.

2.3. VT-Micro Model

This paper investigates the effect of driver heterogeneity on transportation fuel consumption and emissions, so the micro fuel consumption model, VT-micro, is chosen to better fit the research content and purpose of the article.

In existing studies, measure of effectiveness (MOE) is commonly used to measure fuel consumption and emissions per vehicle unit. In the VT-micro model, the MOE relation is obtained from the power product of the current vehicle’s instantaneous speed and acceleration [31]:

$$\ln (MO{E}_e)={\displaystyle {\sum }_{i=0}^3{\displaystyle {\sum }_{j=0}^3{k}_{i,j}^e{v}_n^i{a}_n^j}}$$

(8)

where $MO{E}_e$ denotes the fuel consumption or emission value of the $n$ vehicle at the time $t$, $v_n(t)$ denotes the speed of the vehicle $n$ at the time $t$, $a_n(t)$ denotes the acceleration of the vehicle $n$ at the time $t$, $i$ is the power index of the speed, $j$ is the power index of the acceleration, and $k_{i,j}^e$ is the regression coefficient when the power index of the speed is $i$ and the power index of the acceleration is $j$.

Fuel consumption and emissions per unit of time for a vehicle at a given time are expressed as follows:

$$E_n^e=Exp(MO{E}_e)\times 1000/v_n$$

(9)

where $Exp$ denotes an exponential function with base $e$, $E_n^e$ can represent fuel consumption $E_n^{Fuel}$, carbon dioxide (CO₂) emission $E_n^{C{O}_2}$, and nitrogen oxide (NO_x) emission $E_n^{N{O}_x}$, respectively, in the text.

Li et al. analyzed the energy consumption and emissions of different types of vehicles by testing. They found that the energy consumption and emission characteristics of vehicles under acceleration and deceleration are significantly different [36]. Based on the VT-micro model, they fitted three sets of regression coefficients $k_{i,j}^e$ for different types of energy consumption and pollutants for model calculation. The values of the MOE regression coefficients for fuel consumption are shown in

Table 1. Coefficients for the MOE of fuel consumption.

		$\mathit{i}=0$	$\mathit{i}=1$	$\mathit{i}=2$	$\mathit{i}=3$
$a\ge 0$	$j=0$	−7.735	0.02799	−2.228 × 10−4	1.09 × 10−6
	$j=1$	0.2295	0.0068	−4.402 × 10−5	4.80 × 10−8
	$j=2$	−5.61 × 10−3	−7.722 × 10−4	7.90 × 10−7	3.27 × 10−8
	$j=3$	9.773 × 10−5	8.38 × 10−6	8.17 × 10−5	−7.79 × 10−9
$a<0$	$j=0$	−7.735	0.02804	−2.199 × 10−4	1.08 × 10−6
	$j=1$	−0.01799	7.72 × 10−3	−5.219 × 10−5	2.47 × 10−7
	$j=2$	−4.27 × 10−3	8.375 × 10−4	−7.44 × 10−6	4.87 × 10−8
	$j=3$	1.8829 × 10−4	3.387 × 10−5	2.77 × 10−7	3.79 × 10−10

, the MOE regression coefficients for CO₂ emission are shown in

Table 2. Coefficients for the MOE of CO₂ emission.

		$\mathit{i}=0$	$\mathit{i}=1$	$\mathit{i}=2$	$\mathit{i}=3$
$a\ge 0$	$j=0$	6.916	0.02754	−2.070 × 10−4	9.80 × 10−7
	$j=1$	0.217	0.968 × 10−2	−1.0138 × 10−4	3.66 × 10−7
	$j=2$	2.354 × 10−4	−0.175 × 10−2	1.966 × 10−5	−1.08 × 10−7
	$j=3$	−3.639 × 10−4	8.35 × 10−5	−1.02 × 10−6	8.50 × 10−9
$a<0$	$j=0$	6.915	0.0284	−2.266 × 10−4	1.11 × 10−6
	$j=1$	−0.032	8.53 × 10−3	−6.594 × 10−5	3.20 × 10−7
	$j=2$	−9.17 × 10−3	1.15 × 10−3	−1.289 × 10−5	7.56 × 10−8
	$j=3$	−2.886 × 10−4	−3.06 × 10−6	−2.68 × 10−7	2.95 × 10−9

, and the MOE regression coefficients for NO_x emission are shown in

Table 3. Coefficients for the MOE of NO_x emission.

		$\mathit{i}=0$	$\mathit{i}=1$	$\mathit{i}=2$	$\mathit{i}=3$
$a\ge 0$	$j=0$	−1.080	1.791 × 10−2	2.412 × 10−4	−1.060 × 10−6
	$j=1$	0.2369	4.053 × 10−2	−4.078 × 10−4	9.418 × 10−7
	$j=2$	1.470 × 10−3	−3.750 × 10−3	−1.284 × 10−5	1.860 × 10−7
	$j=3$	−7.822 × 10−5	1.052 × 10−4	1.520 × 10−6	4.419 × 10−9
$a<0$	$j=0$	−1.080	2.111 × 10−2	1.630 × 10−4	−5.832 × 10−7
	$j=1$	0.2085	1.067 × 10−2	−3.230 × 10−5	1.830 × 10−7
	$j=2$	0.02193	6.550 × 10−3	−9.429 × 10−5	4.473 × 10−7
	$j=3$	8.816 × 10−4	6.265 × 10−4	−1.008 × 10−5	4.573 × 10−8

. In the above three tables, the upper halves of the tables are the fitted regression coefficients when acceleration is non-negative, and the lower halves of the tables are the fitted regression coefficients when acceleration is positive.

3. Experiments and Result Analysis

In this section, the analysis of the results of the experiments is given, corresponding to the methodological content in Section 2, respectively.

3.1. Driver Heterogeneity

In this subsection, the way of data processing and driver classification results are introduced. Meanwhile, driver heterogeneity is shown by analyzing the driving characteristics of different types of drivers.

3.1.1. Data Processing

The data selected for this study are extracted from video images of northbound traffic on I-80 in Emeryville, California, published in FHWA’s Next Generation Simulation (NGSIM) program, and the layout of the study site is shown in Figure 7 below [37]. This dataset records vehicle location information every 0.1 s from 4:00 p.m. to 4:15 p.m. on 13 April 2005. Among them, Lane 1 on I-80 is a high occupancy vehicle lane (HOV lane), and Lane 6 is an acceleration lane and connects with the entrance and exit gates, considering that the driving behaviors of these two special lanes will be different from the other normal driving lanes. Therefore, this paper only analyzed the vehicle driving data of the remaining four lanes. And the whole experiment distance is 503 m.

In the NGSIM dataset, the position information of the vehicle itself has a large noise due to the environment, acquisition technology, and other reasons, coupled with the velocity and acceleration information of the vehicle that is obtained by numerical differentiation [38], which will inevitably exacerbate the error of the vehicle data and affect the accuracy of the subsequent researches. Therefore, combining the existing research results and the characteristics of the data required for this article, this paper adopted a three-step method for noise reduction of the original vehicle position information, specifically including determining the location of the outliers, interpolation, and noise reduction in three steps, involving the methods of Gaussian kernel locally weighted regression, Lagrangian interpolation, wavelet noise reduction, and Butterworth noise reduction. Applying the above methods to the dataset for noise reduction, the results are shown in Figure 8, taking 1882 numbered vehicles on Lane 2 of the I-80 region in NGSIM as an example. The figure shows that after the noise reduction processing by the three-step method, the vehicle position curve will be smoother over time, especially at the position mutation, for example, the 120–135, 350–380, and 710–730 time steps. In this way, the data errors caused by the technical problems of the collection equipment, the influence of the external environment, and other conditions can be reduced. Thus, the impact on the subsequent experiments due to data problems can be reduced. At the same time, the smooth position change is also more in line with the behavioral changes of the vehicle in the actual car-following driving state.

In microscopic traffic flow theory, driver heterogeneity is mainly reflected in two traffic flow states, lane-changing and car-following [39,40], and this paper focuses on the study of driver heterogeneity under car-following conditions. Therefore, to distinguish the characteristics of different types of drivers from the vehicle information, this paper first extracted the following fleets from the denoised vehicle driving data so as to ensure the continuity of the car-following behavior of the vehicle in the lane as well as the reasonableness of the data of headway, acceleration, and so on. It is worth noting that this paper only considers the car-following situation of the current vehicle and the two vehicles in front of it, which means the following of multi-vehicle convoys with more than two vehicles is not considered in the experiments. Among them, the conditions for determining the vehicle group to keep following are that the following time is 30 s or more, and the following lengths of the current vehicle and the vehicle in front of it are strictly equal. Finally, 499 sets of vehicle driving information were extracted.

3.1.2. Driver Classification

For car-following behavior, Bergasa [41] summarized the most commonly used classification parameters, such as average speed, acceleration, headway, braking deceleration, and so on. Martinez and Cao [42] compared some driving style classification algorithms, for example, supervised, unsupervised, and reinforcement learning. Based on the existing studies [43], this paper summarized eight factors influencing car-following driving characteristics, which are maximum acceleration (Max acc), minimum acceleration (Min acc), maximum deceleration (Max dec), minimum deceleration (Min dec), average acceleration (Avg acc), average deceleration (Avg dec), average vehicle speed (Avg v), and average head time distance (Avg TH). These eight influencing factors were analyzed for correlation so as to identify a number of factors with the least correlation and to provide a clustering basis for subsequent driver classification.

Spearman’s correlation analysis was used to obtain the results of correlation analysis, as shown in Figure 9. The correlation significant rating is 0.05, which means if the absolute value of the correlation coefficient between the factors is lower than 0.05, then these two factors are not significantly correlated. In order to distinguish the driver types more significantly, it is necessary to select the influential factors with small correlations as the basis for clustering. By analyzing the pairs of influencing factors with no significant correlation marked (×) in the figure and combining them with the existing studies [23] and the actual driving situation, this paper selected Avg TH, Max acc, and Max dec as the discriminative basis for the subsequent driver categorization.

In this paper, the elbow method was used to determine the optimal number of classifications. The cost values of the function corresponding to values of the number of classifications from 1 to 10 are listed in Figure 10. The core idea of the elbow method is that the larger the number of classifications k is, the finer the sample division will be, and the degree of aggregation of each class will gradually increase; then, the error sum of squares will naturally become gradually smaller. However, if the k value keeps rising, the intraclass error variation is already small enough that it can no longer distinguish the samples better than the current k value, meaning the image’s inflection point should be selected. It can be noticed that the degree of distortion is greatly improved when the degree of change in the cost values is gradually insignificant as the values continue to increase. Therefore, for this dataset, k = 3 is the optimal number for clustering. Combining the research object and existing research results, this paper defined three types of drivers: aggressive, neutral, and timid.

Combined with the optimal number of clusters determined above, the clustering results for each clustering method, including the silhouette coefficient and the percentage of each category, are given in

Table 4. Comparison of clustering methods.

Method	Silhouette Coefficient	Percentage of Each Category
		Aggressive	Neutral	Timid
K-means	0.5271	235(36.5%)	284(44.2%)	124(19.3%)
Spectral Clustering	0.6209	5(0.8%)	2(0.3%)	636(98.9%)
Gaussian Mixtures	−0.0248	382(59.4%)	71(11.0%)	190(29.6%)
Ward Hierarchical	0.5368	241(37.5%)	98(15.2%)	304(47.3%)
BIRCH	0.5553	10(1.5%)	340(52.9%)	293(45.6%)

.

The driver classification graph obtained using the K-means clustering method is shown in Figure 11. It can be found that different types of drivers show different driving characteristics. Among them, aggressive drivers (red dots) have the shortest average headway (hourly distance), and aggressive drivers are generally confident; they pursue higher speeds to pass quickly and keep a shorter distance from the vehicle in front of them. In terms of acceleration and deceleration, since this paper conducts a study of driver heterogeneity under the premise of following, in order to ensure the reasonableness of the driving data in the car, the acceleration range of the vehicle is limited to $(-3,3){\mathrm{m}/\mathrm{s}}^2$ when screening the data. As a result, the maximum acceleration and deceleration of all drivers presented in Figure 11 are closely clustered in $3{\mathrm{m}/\mathrm{s}}^2$ and $-3{\mathrm{m}/\mathrm{s}}^2$. Although the numerical range restriction makes the categorization results not differ much in terms of maximum acceleration and deceleration, it is still evident that maximum acceleration and deceleration are overall greater for the aggressive drivers.

3.2. Vehicle Speed Prediction

Combined with the characteristics of the data used in this study, the parameters were iteratively tuned according to the results of the model in the test set, and the parameters of the final model are shown in

Table 5. Parameters for each module of the prediction model.

Parameter	Values
The number of neurons	100
The number of neuron network layers	1
Time step	10
Learning rate	0.001
Encoder hidden state size	64
Decoder hidden state size	64
Batch size	128
Epochs	100

.

In addition to the MSE, the following three indicators were selected for the evaluation of the prediction model:

(1)

Mean Absolute Error (MAE)

$${\left(y_T,{\widehat{y}}_T\right)}_{\mathrm{MAE}}=\frac{1}{N}{\displaystyle \sum _{i=1}^N\left|{\widehat{y}}_T^i-y_T^i\right|}$$

(10)

(2)

Mean Absolute Percentage Error (MAPE)

$${\left(y_T,{\widehat{y}}_T\right)}_{\mathrm{MAPE}}=\frac{1}{N}{\displaystyle \sum _{i=1}^N\left|\frac{{\widehat{y}}_T^i-y_T^i}{y_T^i}\right|}\times 100\%$$

(11)

(3)

R-square (R²)

$${\left(y_T,{\widehat{y}}_T\right)}_{{\mathrm{R}}^2}=1-\frac{{{\displaystyle \sum _{i=1}^n({\widehat{y}}_T^i-y_T^i)}}^2}{{{\displaystyle \sum _{i=1}^n(y_T^i-\overline{y_T^i})}}^2}$$

(12)

3.2.1. Results Comparison in Drivers

For different types of drivers, the model prediction results are shown in Figure 12a, Figure 13a, and Figure 14a, corresponding to aggressive, timid, and neutral drivers, respectively. In order to facilitate the evaluation of the prediction effect for each type of driver, the evaluation indexes of the prediction results for different driver test sets are given in

Table 6. Evaluation indicators of prediction model.

Driver Types	RMSE	MAE	MAPE	R2
Aggressive	0.53	0.30	12.4	0.98
Timid	0.26	0.20	7.06	0.95
Neutral	0.61	0.33	12.94	0.89

. In addition, in order to further analyze the possible reasons for the corresponding prediction effects, Figure 12b, Figure 13b, and Figure 14b partly show the distribution of speeds during the driving process of the three types of drivers and the fitted curves of the corresponding distributions.

Overall, the model proposed in this paper has a good prediction effect for all three types of drivers. Among them, the aggressive driver has the best fitting degree, with a fitting coefficient of 0.98, followed by the timid driver with 0.95, and the neutral driver with a fitting coefficient of 0.89. The reason for the relatively low fitting degree of the neutral driver is mainly reflected in the fact that there is no obvious regularity in the driving process. The distribution of the speed range is relatively large, which increases the difficulty of the model training, and the error value is also relatively the largest. In addition, although the fitting coefficient of the aggressive driver is relatively high, its RMSE, MAE, and MAPE values are significantly larger than those of the timid driver. The reason for this is that the upper limit of the speed range distribution of the aggressive driver is larger, and the proportion of larger speed values is significantly higher than that of timid drivers. The proportion of aggressive drivers with speeds of 6–7 represents the largest proportion at 19%, and the timid driver accounts for the highest proportion of speeds of 2.5–3 and only represents 13%. Therefore, the error value of aggressive drivers also increases with their speed value.

3.2.2. Results Comparison in Models

In addition to analyzing the prediction results of different types of drivers, this paper also selected other prediction models as controls to verify the superiority of the prediction effect of the model proposed in this article. The selection of the control models mainly considers the degree of correlation between the model of this article and the classical model with more references for speed prediction. The results of the model comparison validation correspond to three types of drivers and are shown in Figure 15, Figure 16 and Figure 17, and the error evaluation metrics are shown in

Table 7. Evaluation indicator comparisons for prediction models.

Evaluation	RMSE			MAE			MAPE			R2
Models	A	T	N	A	T	N	A	T	N	A	T	N
DAED	0.53	0.26	0.61	0.30	0.20	0.33	12.40	7.06	12.94	0.98	0.95	0.89
ED	0.61	0.24	0.49	0.45	0.18	0.31	22.20	8.08	17.70	0.96	0.91	0.89
LSTMA	0.79	0.36	0.64	0.57	0.29	0.40	20.50	11.65	17.63	0.95	0.91	0.88
LSTM	0.86	0.39	0.72	0.59	0.31	0.48	21.73	11.85	21.10	0.94	0.89	0.85
CNN	0.94	0.29	0.66	0.62	0.21	0.42	24.20	7.51	16.98	0.93	0.94	0.87
SVM	1.02	0.29	0.68	0.59	0.16	0.31	11.83	5.77	9.77	0.89	0.92	0.80

. For a concise presentation, the models in the table are in abbreviated form. “A” refers to the aggressive type, “T” refers to the timid type, and “N” refers to the neutral type in this table.

It can be found that the proposed model DEAD has the best prediction effects from a general point of view. Specifically speaking, DEAD has the highest fitting degree at 0.89 R² and the lowest prediction error at 0.53 MAPE, 0.30 MAE, and 12.40 MAPE in aggressive-type drivers. As for timid- and neutral-type drivers, the encoder–decoder model may have better performance in some time, but the DEAD model is still at a more stable prediction level.

3.2.3. Model Examinations

In order to investigate the effectiveness of the input attention mechanism in DAED, this paper tested it by using noisy drive sequences as inputs. Specifically, four additional noisy drive sequences were generated by randomizing the original four drive sequences in the predicted input dataset. Then, these four noisy drive sequences are combined with the four original drive sequences as inputs, and the effectiveness of DAED is tested. For a time step of T = 5 and the size of the hidden state of m = p = 256, the model prediction results are shown in

Table 8. Validation of the input attention mechanism.

Driver Types	RMSE	MAE	MAPE	R2
Aggressive	0.64	0.41	13.3	0.97
Timid	0.31	0.28	7.87	0.95
Neutral	0.72	0.48	14.12	0.87

. The results are comparable to the performance in Table 8, which indicates that DAED is robust to noisy inputs.

In order to investigate the effectiveness of the temporal attention mechanism in DAED, DAED is compared with IAL (input attention LSTM) when the time step T varies from 3 to 5, 10, 15, and 25. The results of the RMSE for the three types of drivers at different time steps are shown in Figure 18. It can be found that DAED predicts significantly better than IAL when T is relatively large, which suggests that the temporal attention mechanism can capture long-term dependencies by selecting relevant encoder hidden states at all-time steps.

In addition, this paper investigates the sensitivity of DAED to its parameters, for example, the effect of the hidden layer size of the encoder m (decoder p) and the time step T on the prediction results. By setting m = p = 256, the RMSE is plotted against different time steps T in Figure 19. It is easy to observe that the performance of DAED and IAL will be worse when the length of the time step is too short or too long, while DAED will have relatively less error than IAL. By setting T = 5, the RMSE of the encoder and decoder (m = p ∈ {32, 64, 128, 256, 512}) in Figure 18 is plotted versus different sizes of hidden layers. It can be found that DAED usually achieves the best performance when m = p = 128 or 256. Furthermore, it can be concluded that DAED is more robust to parameters than IAL.

3.3. Environmental Impacts

In this subsection, the three types of drivers mentioned in the previous section will be investigated in terms of fuel consumption. The results are shown in Figure 20 after calculation using the VT-micro model in Section 2.3. The color representation of fuel consumption size in the figure becomes lighter as the value increases, and it can be found that timid drivers have the largest proportion of light colors, followed by aggressive drivers. The distribution of fuel consumption for neutral-type drivers is more balanced and remains lower overall. The maximum fuel consumption values of aggressive drivers are not as high as those of timid types. Still, their fuel consumption distribution is wider, from dark to light colors occupy a small proportion. At the same time, both timid and neutral drivers have a relatively significant distribution of dark and light colors.

The descriptive statistics of fuel consumption of different drivers are given in

Table 9. Descriptive statistics of fuel consumption for heterogeneous drivers.

Driver Types	Best Value (L/km)	Average Value (L/km)	Variance (×10−3)
Aggressive	0.230/0.510	0.367	3.872
Timid	0.280/0.552	0.431	1.307
Neutral	0.217/0.448	0.288	4.766

. It can be seen that the mean value of fuel consumption of timid drivers is relatively high. The minimum and maximum values of fuel consumption of this type of driver are the largest among the three types of values. At the same time, its variance is the smallest among the three types, which indicates that the fuel consumption of this type of driver stays at a high level stably. This may be related to the fact that this type of driver is more likely to be affected by changes in external traffic conditions and is prone to frequent acceleration and deceleration during driving. The neutral drivers have the lowest fuel consumption, which may be due to the fact that this type of driver does not have any special driving preferences and will mainly drive according to their driving rhythm with fewer sudden changes. On the other hand, aggressive drivers generally maintain higher driving speeds and smaller headway, so their changes in speed and acceleration are easily affected by the driving conditions of the vehicle in front of them, and their changes in fuel consumption are also more uncertain.

Figure 21 shows the comparison of CO₂ emissions of different types of drivers, and the descriptive statistics of CO₂ emissions of each type of driver are given accordingly in

Table 10. Descriptive statistics of CO₂ emissions for heterogeneous drivers.

Driver Types	Best Value (×105 kg/km)	Average Value (×105 kg/km)	Variance (×1010)
Aggressive	5.674/13.620	8.434	3.661
Timid	5.404/13.983	9.876	3.325
Neutral	4.706/9.959	6.634	0.966

.

Firstly, it can be noticed that the CO₂ emissions during vehicle driving are huge compared to NO_x. Secondly, observing the distribution of emissions of different types of drivers, a similar distribution to that of fuel consumption appears. The mean value of CO₂ emissions of timid drivers is still the highest at 987,600 kg/km, followed by the aggressive type at 843,400 kg/km and the average type at 663,400 kg/km. The neutral type of driver has the smallest emission maximum, mean, and variance of the three types of drivers, indicating that the CO₂ emissions of this type of driver remain steadily low. The results indicate that the CO₂ emissions of this type of driver are kept steadily at a low level. It can be inferred that the reason for this phenomenon may be the same as the analysis of fuel consumption distribution above, which means the distribution of fuel consumption and carbon dioxide emission is the same for the same types of drivers, while the distribution of different types of drivers is the same for different types of research species.

Figure 22 shows the results of the comparison of NO_x emissions for heterogeneous drivers, and

Table 11. Descriptive statistics of NO_x emissions for heterogeneous drivers.

Driver Types	Best Value (×102 g/km)	Average Value (×102 g/km)	Variance (×103)
Aggressive	1.740/4.535	2.746	3.674
Timid	1.776/4.480	3.221	4.087
Neutral	1.508/3.273	2.153	1.072

shows the descriptive statistics of NO_x emissions for different types of drivers.

First, unlike CO₂, it can be seen from the tables that NO_x emissions during vehicle driving are much smaller than CO₂ emissions. Secondly, it can still be found that the distribution of emissions for each type of driver is similar to that of CO₂ emissions and fuel consumption above. The mean value of NO_x emissions of timid drivers is the largest at 322.1 g/km, followed by aggressive drivers at 274.6 g/km and neutral drivers at 215.3 g/km. The emission maximum, mean, and variance of neutral drivers are still the smallest among the three types of drivers, indicating that the NO_x emissions of this type of driver are steadily kept at a low level. The best value of the aggressive driver is not much different from that of the timid driver. However, the mean value of the emissions can still be significantly smaller than that of the timid driver due to its smaller variance. Up to this point, it can be found that the three types of drivers have the same distribution order in terms of fuel consumption, CO₂ emissions, and NO_x emissions.

Based on the VT-Micro model and the categorized data of driver heterogeneity, the values of fuel consumption, CO₂, and NO_x emissions were calculated. From the results, it can be seen that the stratification of different types of drivers is more obvious and that the three types of drivers have the same order of hierarchical distribution in the three studies; in numerical order from the highest to the lowest, it is timid drivers, aggressive drivers, and neutral drivers.

In the study of fuel consumption, carbon dioxide emissions, and nitrogen oxide emissions, the distribution of the different elements is similar. To analyze this, fuel consumption and pollutant emissions are the result of fuel consumption in the process of vehicle driving, and except for the carbon dioxide and nitrogen oxides studied in this paper, the other emissions also come from fossil fuels. Therefore, the distribution of carbon dioxide and nitrogen oxides should be similar to that of fuel consumption, and it is hypothesized that the distribution of other pollutant emissions should also be similar.

4. Conclusions

With the development of ITS, personalized driving systems are receiving more and more attention, and the investigation of drivers’ heterogeneous driving behaviors is an extremely important foundation before the development of advanced systems. In this study, we carried out a study on vehicle speed prediction considering driver heterogeneity and the impact of fuel consumption and emission in traffic under the car-following state to realize the effective speed prediction and the impact of fuel consumption and emission of traffic with heterogeneous drivers. Based on the data processing, the drivers under the car-following state were classified. Then, the speed prediction results of different types of drivers were obtained by constructing a deep learning model for predicting the speed of heterogeneous drivers. Finally, the predicted speed information was combined with the microscopic fuel consumption and emission model to validate the impact of driver heterogeneity on traffic flow and the environment. The specific conclusions are as follows:

• Three types of drivers are obtained through the Spearman correlation analysis and K-means clustering. The results of driver characteristics show that in the car-following state, different types of drivers will have obvious driving behavior differentiation. Aggressive drivers will keep smaller headway and larger acceleration (deceleration), timid drivers have larger acceleration (deceleration) while there is a larger headway, while the neutral type of driver keeps a moderate level and has no particularly significant driving characteristics or rule changes;
• In order to accurately predict the speed of heterogeneous drivers, this paper constructs a speed prediction model with an encoder–decoder framework and two attention mechanisms named DEAD. The results show that the prediction effects of different types of drivers under the same prediction conditions will be different, and the timid type has the highest prediction accuracy, followed by the aggressive type and neutral type. The prediction model proposed in this paper has better prediction performance than other baseline models, and the parameter sensitivity analysis and robustness analysis also verify that the model has better prediction ability;
• To further study the impact of heterogeneous drivers on traffic, the distribution of fuel consumption, distribution of carbon dioxide emission, and nitrogen oxide emission of different types of drivers are obtained. The results show that timid drivers are generally higher than aggressive and average drivers in fuel consumption and pollutant emissions, while the volatility of neutral drivers is much smaller than that of the other two types of drivers.

In summary, micro vehicle speed forecasting, considering driver heterogeneity, directly contributes to environmental sustainability by reducing emissions and improving energy efficiency. Socially, it enhances safety and mobility, benefiting communities. Economically, it leads to cost savings and optimized infrastructure, aligning with sustainable development goals across these key dimensions.

Limited by the data and knowledge base, this paper still has some things that could be improved. For the heterogeneity of drivers, only the car-following premise is considered, while lane-changing is also a relatively common scenario in real traffic. In the future, we can continue to study the impact of driver heterogeneity in the lane-changing state. At the same time, more driver characteristic influencing factors can be integrated into the category determination. In this paper, we only verified that heterogeneous drivers affect the prediction effect of the model. And we mainly considered the overall prediction ability but did not propose an improvement program to improve the prediction accuracy for a certain category of drivers with poor prediction effect. Due to the limitation of available data, the microscopic fuel consumption and emission model adopted in this paper only considered vehicle speed and acceleration factors, and more relevant information such as weather, season, road friction, and vehicle engine conditions should be considered to improve the accuracy of the model calculation.