Velocity and Energy Consumption Prediction of Medium-Duty Electric Trucks Considering Road Features and Traffic Conditions

Lately, electric vehicle (EV) technology has emerged as a promising approach to address growing environmental concerns, attracting significant interest within the trucking industry [1]. Last-mile delivery fleets particularly stand to benefit significantly from electrification due to their access to existing charging infrastructure, daily return-to-depot operations, and urban driving conditions [2]. Despite these advantages, the adoption of electric trucks (eTrucks) has been slow, primarily due to the issue of limited driving range compared to the high daily mileage expectations in such applications [3]. To ensure the necessary mileage is achieved, careful route planning enabled by accurate estimation and prediction of eTruck energy consumption is essential [4,5].

The energy usage of eTrucks is dependent on multiple features, notably the payload and vehicle velocity. The velocity is influenced by road features and traffic conditions [6]. Unlike passenger vehicles, trucks exhibit distinct behaviors due to their larger size and heavier weight, further compounded by the payload. In most operational scenarios, the anticipated payload variations during delivery are known in advance. Therefore, an accurate forecast of the velocity profile is essential to ensure a reliable estimate of the energy consumption dynamics of eTrucks, which differs considerably from passenger EVs.

A handful of literature explored the EV velocity and energy consumption prediction. Li et al. [7] presents a method for predicting the energy consumption of EVs over a planned route, focusing on statistical characteristics of vehicle speed. Using backpropagation and support vector machine algorithms to model the relationship between speed characteristics and energy consumption, the study achieves an improved prediction accuracy in several driving environments without relying on detailed input features. Hegde et al. [8] proposed a modified intelligent driver model for predicting future speed profiles and the corresponding energy consumption. The model was evaluated using simulationofurbanmobility (sumo) software, highlighting the distinct driver behaviors associated with various road features, and developed speed-limit-to-distance relationships for different scenarios. For nonpassenger vehicle scenarios, the work in Ref. [9] explored a two-step methodology for predicting electric bus energy consumption that combines stochastic speed profile generation with machine learning (ML) models. Random forest and k-nearest neighbor algorithms are leveraged to analyze the generated data and predict energy needs with high accuracy, reflecting the complex dynamics of urban bus transit systems.

While extensive research has been conducted on passenger vehicles, eTrucks, particularly medium-duty (MD) eTrucks for urban use, have received relatively less attention. These trucks exhibit behaviors distinct from passenger vehicles, thereby necessitating further research. In addition to the potentially unique deceleration and acceleration characteristics influenced by the truck weight and size, the significance of payload in energy consumption and operational efficiency becomes more pronounced. The operation behavior of eTruck fleets is characterized by predictability in routes, similar to bus fleets, with operations primarily confined to urban areas. Although the limitation imposed by the vehicles' maximum velocity may restrict the choice of routes, the consistent and foreseeable operation routes offer advantages in the predictions of their velocity and energy consumption.

To this end, this paper contributes a framework for predicting the velocity and energy consumption of MD-eTrucks by significantly modifying the methods for long-term predictions proposed in our prior work [10]. The factors crucial to trucking fleets that have been previously neglected in the context of passenger vehicles are reevaluated. To enhance prediction accuracy, it is essential to consider the distinctive characteristics of various road features, as well as the varying traffic conditions in both highway and city driving scenarios. The proposed framework prioritizes computational efficiency, with a potential future extension to real-time scenarios. Real-world driving data from an ego vehicle are collected and processed to extract relevant road features. The states of each route segment are classified into different levels of traffic congestion using fuzzy c-means (FCM) clustering. These traffic-labeled data, combined with payload information and road features, are utilized to train a gated recurrent unit (GRU). For each predefined delivery route, the GRU generates the most probable velocity profile using a sequence of traffic states predicted by a maximum entropy Markov model (MEMM). Subsequently, eTruck energy consumption prediction is conducted using a truck model created in autonomie.

The remainder of this paper is organized as follows: Sec. 2 provides a review of relevant work. Section 3 depicts the real-world driving data collection and processing, as well as data labeling with road features. Section 4 details the classification of traffic states using principal component analysis (PCA) and FCM clustering on the collected data. An in-depth explanation of the GRU neural network structures and the energy consumption model is described in Sec. 5. Section 6 presents the velocity prediction and energy consumption estimation results, along with a comparison with benchmark methods. A discussion of the proposed method and the limited feature scenarios is conducted in Sec. 7. Finally, Sec. 8 summarizes the key findings and challenges of the prediction model with future research directions.

Given the significant impact of vehicle velocity on the energy consumption of various transportation systems and applications, a wide range of short-term and long-term prediction algorithms is explored in the literature. Short-term prediction techniques are designed to anticipate the speed profile in the near future. Zhou et al. [11] concentrates on predicting vehicle velocity at traffic lights by leveraging a vehicle trajectory generation algorithm and incorporating traffic light information. Rhode et al. [12] employed kernel adaptive filtering methodologies to achieve real-time power prediction for electric vehicles. Yavasoglu et al. [13] considered road type as a critical factor in their real-time remaining range estimation. However, their classification of roads was solely based on speed limits and failed to account for other road factors, such as stop signs and traffic lights.

On the other hand, long-term prediction methods offer the advantage of forecasting a projected speed trajectory for the entire planned driving route in advance. This valuable insight aids drivers in making informed decisions and enables the optimization of global energy management within the system. Li et al. [14] introduced a machine learning approach known as backpropagation long-short-term memory (LSTM) for predicting vehicle velocity in various driving environments, including city, suburban areas, and freeways. In Ref. [15], a long-range speed prediction framework that integrates historical data classification with vehicle-to-infrastructure communication is proposed.

Clustering analysis is employed to categorize traffic patterns without relying on prior knowledge. In Ref. [16], k-means clustering was used to test out different combinations of driving features and numbers of clusters. The results showed that idle time percentage and average acceleration with three clusters yielded the highest traffic condition recognition score for the given dataset. In Ref. [17], Lu et al. used FCM-based clustering to extract representative traffic state patterns from historical data. Rao et al. [18] used interval data-based k-means method to identify traffic states at urban intersections. Azimi and Zhang [19] applied three different clustering algorithms, k-means, FCM, and clustering large applications to identify freeway traffic conditions based on the flow characteristics. Other clustering methods explored include spectral clustering [20,21], adaptive k-means [22], and density-based spatial clustering of applications with noise [23].

In future traffic forecasting, clustering methods coupled with ML algorithms and parametric data-based approaches have been among the most popular strategies [24]. Park [25] used FCM followed by radial basis function neural network, with the experimental results showing more reliable outcome than the linear regression method. In Ref. [26], an improved k-mean algorithm in conjunction with kernel principal component analysis was implemented. Two convolutional neural network models using different architectures were used for both speed and traffic state prediction. Liu et al. [27] proposed a combination model consisting of an FCM-based traffic state partition model and a LSTM traffic state prediction mode. In Refs. [28] and [29], seasonal autoregressive integrated moving average time-series model was employed to capture the strong seasonal pattern in the traffic data. Guo et al. [30] added an adaptive Kalman filter to improve the seasonal autoregressive integrated moving average model by updating the process variances in real-time. Qi and Ishack [31] utilized a hidden Markov model (HMM) for short-term freeway traffic prediction, with the traffic condition in a set time interval as the hidden state. Using a sequence of traffic speed observations, HMMs estimated the most likely sequence of traffic states. This work was extended in Ref. [32], focusing on the contrast measure of traffic.

In summary, long-term velocity prediction methods are suitable approach for fleet applications for the capability to anticipate the energy demands before a trip and plan the routes accordingly. This study will utilize the GRU ML model for long-term velocity and energy prediction of electric trucks, supplemented with traffic conditions. For traffic classification and prediction, FCM and Markov-based MEMM have been selected, leveraging their demonstrated efficiency in short-term passenger vehicle cases as evidenced in the literature reviewed. The model will prioritize truck-specific features and fleet characteristics. Consequently, real-world data capable of providing significant operational insights are essential.

Real-world operational data were collected in Austin, TX from May 2022 and May 2023 using a class-4 eTruck. The participants in the study represented a diverse range of fleet services, from government and university fleets to those specializing in the delivery of goods, beverages, and mails. During this period, each fleet operated the truck for three-month intervals. The subject truck, an electrified Ford E-450, is equipped with a permanent magnet motor powered by a 120 kWh nickel manganese cobalt lithium-ion battery. The detailed specifications of the eTruck can be found in Table 1.

Vehicle operational data, including longitudinal velocity, GPS coordinates with elevation, motor speed and torque, battery voltage, current, and state of charge, as well as auxiliary power associated with air conditioning, were recorded by an onboard data logger connected to the control area network bus at a frequency of 1 Hz. Additionally, the payload information for each day was documented by the drivers. Due to sensor limitations, no specific traffic data or information about neighboring vehicles were collected. A sample delivery route with a photo of the eTruck is presented in Fig. 1. The delivery destinations were identified based on the vehicle's idling duration and parking gear engagement.

Erroneous data points are first eliminated from the collected dataset. This included the removal of incomplete or partial trips resulting from data transmission issues, conversion problems, or logger errors. Additionally, brief trips that occurred solely within the fleet's premises, such as moving the vehicle between the charging station and cargo loading docks, were excluded from the analysis. As the raw signal contained a considerable amount of noise, a Kalman filter was used to refine the velocity and GPS data. In addition, the data collected in the time domain were converted to the distance domain [33], as road features are typically defined by the distance between the starting point and their respective locations.

The vehicle's positional data from the GPS were utilized to label the specific road that the ego vehicle was traversing at each time instance, including the road type and the speed limit. Various road types in each route impact driving behavior differently, due to road-specific attributes such as stop signs on local roads, highway ramps, and the varying density of road signages. While vehicle velocity during highway driving is notably affected by traffic density and the lead vehicle's speed in relation to the speed limit, urban/local driving is also influenced by road characteristics. The geographical locations of the road features were sourced from the public Texas Department of Transportation database [34,35]. Using arcgis software, the positions of these traffic control devices were integrated into the vehicle route data.

The vehicle's turning maneuvers were labeled as these actions are inherently associated with the vehicle's acceleration and deceleration. Identification of turning direction was based on the vehicle's heading angle. Moreover, incorporating road signage to distinguish between signaled and unprotected turns enables a more precise evaluation of driving behaviors, such as stopping distances or the degree of deceleration when making turns. A sample velocity plot is shown in Fig. 2, annotated with the road features and driver maneuvers.

This section describes the traffic classification methods.

To incorporate the traffic states into the velocity forecast, the states need to be first identified for the collected data. In real-world driving, categorizing traffic levels becomes a complex task due to the traffic congestion level being a continuous spectrum rather than a discrete state. For the classification of traffic on a fixed road segment using data from traffic detectors [36], utilizing predefined traffic levels can be effective. For example, the highway capacity manual employs six categories (A, B, C, D, E, and F) [37], while Cheng et al. [38] employed five categories, representing states as fluent, basic fluent, slight congestion, moderate congestion, and severe congestion.

In this study, the ego vehicle data collected from various operating routes are utilized, with the destinations changing on a daily basis. The traffic is categorized into three states (light, medium, and heavy traffic [39]) to ensure a broader and more adaptable applicability than fixed-route classifications. The “light” state corresponds to unhindered driving, with the vehicle often surpassing the road speed limit, depending on driver characteristics and road features. Drivers in this state are unaffected by the presence of other vehicles. The “medium” state describes lower average velocities with varying degrees of travel disruption. The “heavy” traffic state signifies traffic density at or near its peak due to adverse conditions and extensive queuing. Vehicles experience significant delays, and driving freedom is greatly restricted.

As the ego vehicle will experience dynamic changes in traffic conditions in a single delivery route, a trip needs to be divided into shorter road segments [40]. Each unique road within the delivery route is treated as a distinct route segment. This implies that each segment possesses road characteristics, including road type, number of lanes, and speed limit. Segments exceeding a maximum distance of 4 km are further divided to capture more localized traffic details. Conversely, routes with lengths less than 200 m are considered too short for accurate traffic state identification and are not assigned a traffic state. Each segment is paired with a distinct and consistent traffic state.

Classification indicators are features used for the evaluation and categorization of traffic conditions. Commonly utilized indicators include traffic flow, speed, density, occupancy, and queue length [38]. From the primary indicators, additional parameters may be derived, such as the occupancy-to-flowrate ratio and occupancy-to-speed ratio. Typically, a combination of these indicators serves to describe the dynamic characteristics of traffic.

As velocity prediction will be conducted in the distance domain, traffic-related information retrieved from publicly available navigation databases, namely, googlemapsplatform, may not be directly applicable for traffic level classification as they primarily focus on temporal characteristics like idling time or delays. Thus, in addition to utilizing the navigation databases, parameters that can effectively encapsulate essential information for identifying driving conditions are obtained from the collected driving profiles using a feature engineering process. A total of nine parameters, shown in Table 2, that characterize the traffic conditions were selected following the methodology outlined in Refs. [41] and [42]. Major parameters include:

• Speed ratio (average and maximum and median): The ratio between the averaged/maximum/median velocity of the vehicle and the road speed limit.

• Coefficient of variation: The ratio between the mean and standard deviation of the velocity measurements.

• Average acceleration and deceleration with magnitude larger than 0.15 m/s².

• Duration ratio: The ratio of estimated travel duration with and without traffic from googlemapsplatform.

• Percentage of low-speed cruising: The percentage of travel distance with an instantaneous velocity less than 5 m/s.

The calculation of these characteristic parameter values excludes data points where the vehicle is in close proximity to traffic signals. This exclusion is based on the rationale that deceleration and idling at red lights or stop signs may not accurately represent the traffic level, as frequent stops due to traffic signals could erroneously indicate a higher level of congestion.

Principal component analysis is performed on the characteristic parameters before the clustering process. Analyzing these parameters from the collected driving cycle data, PCA aids in deriving a reduced set of uncorrelated principal components. These components effectively capture the most significant sources of variance in the data while minimizing redundancy, thus simplifying the dataset while preserving essential information. Such dimensionality reduction is especially beneficial when handling numerous parameters, streamlining the analysis, enhancing computational efficiency, and improving data manageability.

The characteristic parameters of a total of 270 driving segments were analyzed using PCA. The eigenvalues, the percentage of variance explained, and the cumulative variance of each principal component are shown in Table 3. As can be seen from the table, the cumulative variance explained by the first three principal components exceeds 80%, indicating their ability to adequately capture the driving segment's traffic characteristics. Additionally, the eigenvalues of the first three principal components are all greater than 1, signifying that the principal component captured more variance than the individual original features according to the Kaiser–Guttman criterion. Consequently, these first three principal components were chosen to represent the characteristic parameters.

For traffic state classification of each route segment with the principal components, fuzzy c-means clustering was used. While unsupervised clustering methods are primarily used for pattern discovery and data grouping without pre-existing labels, these methods can be used to reveal hidden data structures to serve as a basis for classification. By using cluster memberships as soft labels, the clustering result is matched to the traffic level defined in Sec. 4.1 as a semisupervised classification problem.

Fuzzy c-means is a robust clustering method, characterized by its ability to handle data points that belong to multiple clusters simultaneously. Unlike other clustering methods which assign data points to a single cluster with a crisp membership such as k-means, FCM introduces the concept of fuzziness, allowing data points to have partial membership in multiple clusters. This property is essential in traffic state classification, as it accommodates the inherent uncertainty and ambiguity often associated with real-world traffic data. FCM's capacity to offer a soft assignment of data points, where each point contributes to all clusters to varying degrees, can result in more nuanced and accurate traffic state classifications.

The algorithm iteratively updates the membership matrix and the cluster centers until the maximum difference between the updated and the previous matrix is less than a predefined threshold, or a maximum number of iterations is reached.

The 270 route segments, with features reduced to three principal components, were classified into three states using FCM. The scatter plot depicting the clustering results is displayed in Fig. 3, with each colored marker representing a distinct cluster. The average values of key characteristic parameters for each resulting cluster are shown in Table 4. FCM achieved reliable segment classification: Cluster 1, characterized by vehicle velocities exceeding the road speed limit and minimal low-speed cruising at 0.9%, excluding traffic wait times, indicating a state of light traffic. Cluster 2, with a mean speed ratio of 0.64, but still with minimal cruising proportion and a maximum velocity mirroring the speed limit, represented a state of medium traffic. During the delivery scenarios in cluster 3, the vehicle spent approximately 20% of the travel distance either idling or low-speed cruising with a mean speed equal to half of the speed limit, displaying the distinct characteristics of high traffic congestion. The traffic label also corresponds with the duration ratio estimates provided by googlemaps, with the exception of cluster 2. The minor deviation observed in cluster 2 could be attributed to the driving and operational characteristics of the truck, resulting in more frequent stops compared to the average traffic in urban environments. Refer to Fig. 4 for an illustration of driving data split into route segments, each with its corresponding identified traffic condition. The pins on the map mark the locations of traffic lights.

While the selected features and principal components are considered reliable indicators for distinguishing traffic states, it is essential to assess the precision and effectiveness of clustering outcomes. To provide a systematic evaluation of the results, two assessment criteria were selected: the silhouette score and the Davies–Bouldin score.

where a(i) and b(i) are the cohesion and separation of sample i, respectively. The overall silhouette score for a clustering solution is defined as the average silhouette score for all data points.

In addition, the silhouette score is used to identify the optimal number of clusters. If the number of clusters is too small, the difference between samples in the same cluster is large, and if the number of clusters is too large, the difference between different clusters in the clustering results will not be obvious. Experimentation with three to six clusters on the processed dataset revealed the highest score of 0.44, achieved when utilizing three clusters.

where N is the number of clusters, i and j are indices for individual clusters. R_i is the average distance of data points in cluster i to the centroid of cluster i. R_j is the average distance of data points in cluster j to the centroid of cluster j. d_ij is the distance between the centroids of clusters i and j.

The clustering result produced a DB score of 0.90, slightly higher than anticipated. This value is likely attributed to the traffic states' spectral and less distinctly separable nature. Nonetheless, the observed characteristic parameter values from cluster centers aligned sensibly, signifying a meaningful clustering outcome.

This section outlines the framework for velocity prediction and the corresponding energy consumption estimation for the eTruck.

The velocity profile for every road segment in a delivery route is sequentially predicted using a GRU neural network model. The model is trained using the collected data labeled with road features and traffic states identified in Sec. 4, along with the payload information, to effectively capture the vehicle's behavior under various driving conditions. For future segments, the trained model predicts the probable vehicle velocity, utilizing the known road features and payload, information from the previous segment, and traffic conditions estimated from MEMM.

A GRU is a type of recurrent neural network (RNN) architecture used for sequence modeling, designed to address some limitations of traditional RNNs. The gating mechanism enables the GRU to capture long-range dependencies in sequences while mitigating the vanishing gradient problem that exists in vanilla RNNs. Although sequential models like LSTM and the one-to-many transformer architecture utilized by coauthors in prior studies [10,43] could also be viable for sequence prediction, GRU was chosen for its computational efficiency.

The many-to-many GRU model is designed to predict a sequence of vehicle velocities for each road segment by learning from a sequential input of features. These features are structured in the distance domain, with sequence length proportional to the specific route's length. As the lengths of routes vary, padding, start, and stop tokens are added to assist the model in processing sequences of varying lengths. The three input features—road speed limit, traffic state, and payload—are constant value sequences. Meanwhile, road signage, including stop signs, traffic lights, and crossings, are categorical variables designated at their respective positions in the input sequence. This enables the model to discern whether changes in vehicle velocity are linked to traffic or passing a road feature.

With each segment denoting a specific road along the delivery route, the transition between segments commonly involves a turning maneuver to enter a new road. Unlike the approach in Ref. [43], which trains acceleration, deceleration, and steady-state driving separately, the GRU model is trained using complete sequences. In longer sequences, the weight of turning maneuvers compared to the overall profile may be diminished, which can lead to decreased estimation performance. Hence, the expected turning velocity is determined prior to the prediction and added as an input feature. For this “guide” velocity input, a compact single-layer feedforward neural network with 32 neurons is implemented. The network takes initial velocity, road class, and payload weight as inputs, with a focus on the substantial influence of payload weight on maximum vehicle acceleration and deceleration.

The architecture is structured as a sequential model consisting of two layers of GRU. Each GRU layer has a dimension of 64 units and employs the hyperbolic tangent (tanh) as its activation function. A dropout rate of 0.1 is applied within the model to mitigate overfitting. The optimizer of choice is adaptive moment estimation (Adam) with a learning rate set to 0.001, which is used to adjust the model's parameters during training. The mean absolute error (MAE) serves as the loss criterion, with the padded values masked in the calculation. The network was trained for 200 epochs with an input batch size of 16.

The future traffic state prediction holds a significant role in forecasting vehicle speed. Traffic conditions serve as a constraint on velocity, wherein heavier traffic is anticipated to reduce the maximum cruising velocity and prompt more frequent deceleration. Given that the features utilized in traffic classification are not suitable for future traffic prediction, models based on the Markov property were examined. Markov property refers to the memoryless property of a stochastic process, stating that the future evolution is independent of its history. When applied to traffic states, these models operate on the assumption that the traffic state of the next road segment is solely influenced by the current traffic state. With the amount of available data classified from a single ego vehicle, Markov-based models can be more fitting than ML-based models that require extensive labeled datasets. A similar concept was applied in Ref. [44], where separate Markov chains, for rush hours and nonrush hours, were constructed to model the state transitions. As the Markov chain approach overlooks specific road attributes and complex temporal relationships, a MEMM has been chosen for the generation of future traffic sequences.

where f_j are real-valued or categorical feature-functions that capture the relationships between the observations, states, and transitions, λ_j are the corresponding weights, and $Z(xi,yi−1)$ is a normalization term ensuring that the distribution sums to one.

The MEMM uses the observed states to make the most probable sequence of predictions about the labeled states, by assuming these states are connected in a Markov chain rather than being conditionally independent of each other. The dependency between the labeled and observed states makes MEMM more suitable than other commonly used sequential models, such as HMMs, where the observed states are dependent on the hidden states. In the context of traffic level prediction, the observed road type, time of the day, and the corresponding relative traffic load are used in generating hidden traffic level states. During sequence generation, the traffic state is reset when the duration spent at each delivery destination exceeds a certain threshold. This reset becomes necessary as the traffic conditions from the last segment might no longer offer reliable input for predicting the upcoming route segments' traffic. In addition, shortening the sequence minimizes label bias, making the MEMM notably more efficient when compared to other discriminative models like conditional random fields.

The relative traffic load is calculated by taking the traffic rate at a specific time of day and dividing it by the peak traffic rate observed in Dallas, TX during weekdays. The traffic data collected in 240 locations throughout the city between 2017 and 2020 by the North Central Texas Council of Governments database [45] were normalized with peak traffic count, as shown in Fig. 5. This ratio effectively captures the relative traffic patterns at different times of the day, providing a measure of how heavy the traffic is compared to the peak traffic. Supplementing the time input with this variable can offer a more comprehensive view of the traffic situation, as the collected data are limited to specific times of the day, contingent upon the type of delivery fleet. While this approach assumes consistent traffic patterns on weekdays and generalizes the observed pattern from the Dallas metropolitan area to represent Austin's traffic, the combination of features can help to improve the model's predictive performance by providing a richer representation of the traffic data.

Due to the nonlinear relationship, the estimation of vehicle energy consumption from the predicted velocity profile can be challenging. To enhance accuracy in estimation, autonomie, developed by Argonne National Laboratory, was employed in this study. autonomie is a comprehensive vehicle system simulation tool designed to evaluate the energy consumption, performance, and cost of a wide range of vehicle technologies. It provides a unique set of full vehicle models and control algorithms, enabling precise energy consumption estimation for specific driving cycles. The eTruck model was constructed using component specifications provided by the manufacturer. Proprietary details were approximated from the collected driving data. The vehicle's specifications and model parameters are outlined in Table 1. The longitudinal vehicle model reliably delivers highly accurate estimation results for energy consumption utilizing the velocity profile, payload, and road grade as inputs.

Two routes were selected from observed data to assess prediction performance. These routes vary in length and exhibit diverse distributions of road features. The selection was aimed at highlighting the effectiveness of two key features—traffic states and guide velocity (v_guide) from feedforward neural network. It is presumed that a priori knowledge of delivery destinations, their associated routes, road features, and payload details is available.

In the context of vehicle velocity prediction, individual velocity profiles for each segment are independently forecasted and then combined to form a complete route prediction. Alongside the proposed GRU model that leverages full input feature sequences, two supplementary scenarios are tested using reduced features for comparison: one scenario trains without traffic states, and the other trains without guide velocity. These models undergo evaluation to analyze the influence of key features on velocity prediction performance. The model architecture remains consistent across all comparisons.

where L represents the road segment length, $v̂i$ and $vi$ are predicted and true velocity, respectively. The prediction results are illustrated in Fig. 6. The prediction results are provided in Table 5.

The energy consumption estimation is evaluated with RMSE and MAE between the actual and the predicted energy consumptions. The true energy consumption of the eTruck from the collected data serves as the reference, while the energy consumption estimation generated by autonomie using the true velocity profile serves as the baseline. The prediction results of the three input scenarios, in comparison to both references, are depicted in Fig. 7. The corresponding error results calculated in the distance domain are presented in Table 6.

The GRU model, trained with proposed input features including speed limit, traffic state, payload, road characteristics, and guide velocity, successfully generated precise speed profiles across all driving stages for the tested routes. Notably, the acceleration and deceleration stages, primarily linked to deceleration at traffic signs or turning maneuvers at the start and end of road segments, exhibited higher predictive accuracy compared to the cruising stage, where the subject vehicle is following the speed limit or the preceding vehicle. This outcome was as expected, considering the inherently more stochastic nature of cruising stage velocities. These speeds are significantly influenced by various factors, some of which are not captured in the available data, such as the driving behaviors of neighboring vehicles. However, the inclusion of road signage along the route allowed the model to predict some speed variations in the cruising stage. This is evident in Fig. 6, route 1 at a distance of 8000 m, where the predicted velocity dropped as the ego vehicle passed a traffic light.

The results of the proposed model demonstrate superior performance when compared to the velocity predictions of models trained with two limited inputs: without utilizing the guide velocity (w/o v_guide) and without incorporating the traffic data (w/o traffic). In Fig. 6, the route 1 plot reveals that the prediction result that excludes the utilization of guide velocity, which captures driver turning maneuvers, exhibits less accurate acceleration and deceleration behavior. While the amount of overshoot relative to the speed limit remains similar, given its dependence on traffic conditions, the magnitude of acceleration and deceleration is slightly higher than optimal. Moreover, the vehicle exhibited hard deceleration behavior resembling a turning maneuver when passing through a traffic light (8000 m), even when the light was not red, with the absence of guide velocity.

In route 2, the last three segments share very similar driving conditions, including identical speed limits and the relative location of traffic lights. Despite this uniformity, the collected driving data display varying cruising speeds, likely due to fluctuating traffic conditions. While the proposed model adeptly captures most of this variation, the model trained without traffic state input predicted the same cruising velocity to all three segments. These results indicate that the model effectively captures the important characteristics of all input features, suggesting the significance and uniqueness of each feature in the overall predictive process.

The estimation results from the true velocity input validate the effectiveness of the autonomie model in estimating energy consumption with minimal error when given accurate velocity, payload, and road grade inputs. Some minor discrepancies observed are likely due to auxiliary power consumption, regenerative braking efficiency, and the Kalman filter used in data processing. Additionally, the conversion of data between distance and time domains for the simulations may also contribute to the observed error. When comparing the estimation results of the three input scenarios, the predictions generated by the proposed method align more closely with the actual energy usage profile than those based on limited inputs. Despite consistent payload and road grade inputs, variations in the velocity profile significantly impact the final energy consumption estimation, further emphasizing the importance of accurate velocity prediction. In all comparison cases, the main source of estimation error originates from the cruising sections. Nonetheless, the estimation performance of the proposed method proves sufficient for delivery fleets, enabling them to make more accurate operational decisions and enhancing their confidence in the eTruck driving range estimation.

This paper presents a specialized framework for predicting the velocity and energy consumption of medium-duty electric trucks. For accurate prediction performance, all available information, including various road features, payload, and traffic conditions, needs to be carefully considered. From real-world driving data, relevant features are extracted through feature engineering and PCA, and then classified into multiple traffic state levels using FCM. The resulting traffic-labeled data, combined with payload information and road features, contribute to training a GRU model. For predefined delivery routes, the GRU generates probable velocity profiles using a sequence of traffic states predicted by MEMM. Subsequently, energy consumption prediction is conducted using autonomie with forecasted velocity and payload. Experimental results showed that the proposed method accurately predicted eTruck velocity and energy consumption. Comparative analysis with limited input scenarios demonstrates the model's effectiveness in utilizing all input features for accurate predictions.

While the proposed method demonstrated notable performance, potential improvements can be made in future research. First, a more localized influence of traffic, including the impact of the lead vehicle, could enhance accuracy. In addition, integration of commercial real-time traffic update services may improve the traffic forecasting algorithms. Second, severe weather conditions affecting acceleration or deceleration behavior and extreme temperatures impacting auxiliary energy consumption could serve as valuable additional input features. Finally, nonstationary traffic signage such as traffic lights could be treated as stochastic elements. Probabilistic models that account for the frequency and durations of light changes, along with the corresponding response of the drivers, can be considered.

• Department of Energy under DOE (Award No. DE-EE0009223).

The datasets generated and supporting the findings of this article are obtainable from the corresponding author upon reasonable request.