Nothing Special   »   [go: up one dir, main page]
Next Article in Journal
The Impact of Environmental Regulation on the Growth of Small and Micro Enterprises: Insights from China
Previous Article in Journal
Spatiotemporal Patterns and Influencing Factors of Carbon Emissions in the Yangtze River Basin: A Shrinkage Perspective
You seem to have javascript disabled. Please note that many of the page functionalities won't work as expected without javascript enabled.
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Sustainability
Volume 17
Issue 5
10.3390/su17052111
sustainability-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Enhancing Travel Time Prediction for Intelligent Transportation Systems: A High-Resolution Origin–Destination-Based Approach with Multi-Dimensional Features

by
Chaoyang Shi
1,2,3,4,5,
Waner Zou
1,
Yafei Wang
6,*,
Zhewen Zhu
1,
Tengda Chen
1,
Yunfei Zhang
2 and
Ni Wang
3,4,5
1
School of Civil and Hydraulic Engineering, Huazhong University of Science and Technology, Wuhan 430074, China
2
Engineering Laboratory of Spatial Information Technology of Highway Geological Disaster Early Warning in Hunan Province, Changsha University of Science & Technology, Changsha 410114, China
3
Anhui Province Key Laboratory of Physical Geographic Environment, Chuzhou University, Chuzhou 239000, China
4
Anhui Engineering Laboratory of Geo-Information Smart Sensing and Services, Chuzhou 239000, China
5
Anhui Center for Collaborative Innovation in Geographical Information Integration and Application, Chuzhou 239000, China
6
School of Surveying and Geo-Informatics, North China University of Water Resources and Electric Power, Zhengzhou 450046, China
*
Author to whom correspondence should be addressed.
Sustainability 2025, 17(5), 2111; https://doi.org/10.3390/su17052111
Submission received: 3 February 2025 / Revised: 24 February 2025 / Accepted: 26 February 2025 / Published: 28 February 2025
Browse Figures
Figure 1
<p>Spatial information of the study area: (<b>a</b>) Location of Xi’an; (<b>b</b>) The central area of Xi’an; (<b>c</b>) The road network of the study area.</p> "> Figure 2
<p>Overview of the proposed methodology.</p> "> Figure 3
<p>Structure of the BPNN model.</p> "> Figure 4
<p>Performance of the proposed method: (<b>a</b>) CDF of the absolute percentage error; (<b>b</b>) Influence of travel distance on MAPE; (<b>c</b>) Influence of travel time on MAPE; (<b>d</b>) Influence of travel speed on MAPE.</p> "> Figure 5
<p>MAPE distribution with travel time and distance.</p> ">
Versions Notes

Abstract

:
Accurate travel time prediction is essential for improving urban mobility, traffic management, and ride-hailing services. Traditional link- and path-based models face limitations due to data sparsity, segmentation errors, and computational inefficiencies. This study introduces an origin–destination (OD)-based travel time prediction framework leveraging high-resolution ride-hailing trajectory data. Unlike previous works, our approach systematically integrates spatiotemporal, quantified weather metrics and driver behavior clustering to enhance predictive accuracy. The proposed model employs a Back Propagation Neural Network (BPNN), which dynamically adjusts hyperparameters to improve generalization and mitigate overfitting. Empirical validation using ride-hailing data from Xi’an, China, demonstrates superior predictive performance, particularly for medium-range trips, achieving an RMSE of 202.89 s and a MAPE of 16.52%. Comprehensive ablation studies highlight the incremental benefits of incorporating spatiotemporal, weather, and behavioral features, showcasing their contributions to reducing prediction errors. While the model excels in moderate-speed scenarios, it exhibits limitations in short trips and low-speed cases due to data imbalance. Future research will enhance model robustness through data augmentation, real-time traffic integration, and scenario-specific adaptations. This study provides a scalable and adaptable travel time prediction framework, offering valuable insights for urban traffic management, dynamic route optimization, and sustainable mobility solutions within ITS.

1. Introduction

Accurate travel time prediction (TTP) is a fundamental component of intelligent transportation systems (ITSs), supporting applications that benefit travelers, traffic managers, and service providers. Reliable TTPs enable travelers to make informed route choices, improving their commuting experience. Traffic managers can leverage accurate prediction to proactively mitigate congestion and optimize traffic flow, while service providers, such as ride-hailing platforms, can enhance operational efficiency and customer satisfaction [1,2,3,4]. These advantages underscore the importance of developing robust TTP models that enhance urban mobility, alleviate congestion, and support sustainable transportation planning [5].
Despite its significance, TTP remains a challenging problem due to the highly dynamic, nonlinear, and stochastic nature of urban traffic systems, which fluctuate across both spatial and temporal dimensions. The continuous advancement of traffic data acquisition technologies, including loop detectors, video surveillance, automated vehicle identification (AVI) systems, and floating car data, has facilitated progress in this field. However, existing TTP methodologies can be broadly classified into link-based and path-based approaches [6,7,8], with inherent limitations.
Link-based models estimate travel time by summing the travel times of individual road segments and accounting for intersection delays. While intuitive and widely adopted, these models face challenges such as error accumulation from segmentation, high computational complexity, and reliance on sparse spatial–temporal detector data [9,10,11,12,13,14,15]. The accuracy of link-based models is further compromised when applied to highly congested road networks with dynamic traffic conditions. In contrast, path-based models predict travel time for an entire route using historical trip data, effectively bypassing the need for segmentation. These models perform well in highway and expressway settings, where entry/exit points are limited, and traffic conditions are relatively stable [16,17,18,19]. However, path-based models struggle with scalability due to their dependency on predefined routes and historical sample availability, limiting their applicability in complex urban road networks. Recent works have addressed these limitations by incorporating additional contextual factors such as electronic toll collection data and segment-based data imputation, demonstrating improved performance in controlled environments such as expressways [20,21].
While both link-based and path-based methods have contributed to advancements in TTP, they exhibit key limitations, including limited data utilization due to reliance on predefined routes or road segments, computational inefficiency in large-scale urban networks, sparse spatial and temporal coverage reducing prediction generalizability, and an over-reliance on heuristic modeling that limits adaptability to real-time traffic variations. To overcome these constraints, origin–destination (OD)-based TTP has emerged as a promising alternative. Instead of relying on predefined routes or link segmentation, OD-based models predict travel time using only the OD pairs, leveraging historical trip data without requiring full trajectory information. This approach is particularly valuable in applications such as urban mobility analysis, travel demand forecasting, and large-scale traffic optimization, where precise route details are often unavailable or irrelevant.
In early research, OD-based TTP models primarily relied on AVI-based toll station data and automated license plate recognition (ALPR) systems. However, these approaches required continuous vehicle tracking across a fixed route, making them vulnerable to data sparsity issues, particularly when vehicles entered or exited midway through a journey. Additionally, prediction accuracy deteriorated in low-traffic periods due to insufficient historical samples. With the rise of ride-hailing platforms such as Uber, Lyft, and Didi Chuxing, unprecedented volumes of high-resolution trajectory data have become available. These data provide detailed spatiotemporal records of trip origins, destinations, and routes, offering a unique opportunity to develop more comprehensive OD-based models [7,22,23,24,25,26,27]. However, a major challenge remains: historical trip records are insufficient for many OD pairs for direct prediction.
Recent research has explored nearest-neighbor approaches for OD-based TTP, where historical trips with similar OD pairs and departure times are identified and used for prediction. However, these methods overlook the inherent randomness of urban traffic conditions. Simply averaging historical travel times for similar trips fails to account for external factors, leading to reduced accuracy in real-world scenarios. Additionally, most OD-based models lack integration of key external influences, including traffic infrastructure attributes (e.g., number of intersections, traffic signals, and congestion levels), weather conditions (e.g., temperature, precipitation, and air quality), and driver behavior patterns (e.g., route preferences and acceleration/deceleration tendencies). The absence of these critical factors introduces significant variability in travel time, underscoring a crucial research gap in improving OD-based models’ predictive accuracy and generalizability. Recent studies have attempted to address some of these limitations by incorporating deep learning models such as graph convolutional neural networks (GCNs) [28] and adaptive graph embedding [29], demonstrating improvements in predictive performance. Additionally, the integration of weather conditions into travel time estimation has been shown to enhance model accuracy, particularly for path-based TTP models [30]. Moreover, studies on multi-source data fusion have provided new insights into how different data types can be effectively leveraged for improved TTP in urban arterial networks [31].
Another emerging challenge in OD-based TTP is the impact of crowdsourced and shared mobility services on travel time variability. For example, the rapid expansion of crowdsourced door-to-door delivery services introduces additional unpredictability in urban travel patterns. Recent research [32] has explored the impact of such services on travel demand and congestion, highlighting the need for more adaptable TTP models that can incorporate these dynamic factors.
While previous studies have incorporated weather conditions and driver behavior into travel time prediction, many rely on oversimplified feature representations that fail to capture their dynamic impact. Some approaches use categorical weather labels or predefined driver classifications, ignoring the evolving nature of these factors. For instance, Lagos et al. [33] applied k-nearest-neighbor methods for OD-based travel time estimation but did not integrate multi-dimensional external influences. Fang et al. [34] employed hierarchical clustering to segment road networks without explicitly modeling environmental or behavioral factors.
To bridge these gaps, this study systematically integrates weather and driver behavior features within an OD-based prediction framework, optimizing feature utilization through a Bayesian-tuned Back Propagation Neural Network (BPNN). Specifically, a Random Forest-based approach is employed to dynamically quantify the impact of weather conditions, capturing variations in travel time across different weather scenarios. Additionally, a data-driven clustering technique is applied to infer driver behavioral tendencies based on speed and acceleration variations, ensuring a more accurate and adaptive representation of driving behavior.
Building on these advancements, this study proposes a comprehensive OD-based TTP model that fully incorporates high-resolution ride-hailing trajectory data and multi-dimensional external factors. The key contributions of this study are summarized as follows:
(1) This study introduces an enhanced OD-based TTP framework that eliminates the constraints of route segmentation while leveraging large-scale ride-hailing trajectory data. Unlike traditional OD-based models that rely solely on historical travel time averages, our approach integrates contextual factors to improve prediction accuracy and adaptability, addressing key limitations of traditional link- and path-based approaches.
(2) A refined multi-feature integration method systematically incorporates spatiotemporal, weather, and driver behavior features. Instead of simplistic categorical labels, weather effects are dynamically quantified using a Random Forest-based approach, capturing the nuanced impact of meteorological conditions on travel time. Similarly, driver behavior is characterized through data-driven clustering techniques based on speed and acceleration variations, offering a more accurate representation of driving tendencies. These refinements provide deeper insights into external and behavioral variability in urban travel dynamics.
(3) A Bayesian-optimized Back Propagation Neural Network (BPNN) enhances predictive performance and computational efficiency. While deep learning models such as GNNs and transformers have been explored in prior studies, our approach optimizes the BPNN through Bayesian hyperparameter tuning, ensuring a balance between model complexity and interpretability while maintaining high prediction accuracy. Furthermore, extensive ablation studies systematically evaluate the contributions of different feature sets, providing empirical validation of feature effectiveness.
(4) Empirical validation on large-scale ride-hailing trajectory data from Xi’an, China, demonstrates superior predictive performance, achieving an RMSE of 202.89 s and a MAPE of 16.52%. Ablation experiments quantify the incremental contributions of spatiotemporal, weather, and driver behavior features, showcasing the model’s robust applicability in dynamic traffic management, ride-hailing service optimization, and congestion mitigation within intelligent transportation systems (ITSs).
The remainder of this paper is structured as follows: Section 2 overviews the online ride-hailing dataset, detailing its characteristics and preprocessing steps. Section 3 introduces the proposed OD-based TTP framework, detailing its feature integration and model architecture. Section 4 presents a detailed performance analysis, focusing on the contributions of spatiotemporal, weather, and driver behavior features. Finally, Section 5 concludes the study and outlines future research directions.

2. Data Collection and Preprocessing

2.1. Data Description

The dataset used in this study is derived from the Didi Chuxing “Gaia” Open Dataset, encompassing historical trip records and high-resolution GPS trajectory data for ride-hailing services within Xi’an, China. Xi’an was selected due to its well-documented urban traffic patterns, availability of extensive ride-hailing trajectory data, and diverse weather conditions affecting travel. This dataset features extensive spatial and temporal coverage, high sampling precision, and a substantial sample size, making it a reliable foundation for OD-based TTP.
Figure 1a illustrates the geographical location of Xi’an within Shaanxi Province and China, providing context for the study area. Figure 1b focuses on the urban area of Xi’an, highlighting the five administrative districts included in this study: Lianhu, Beilin, Xincheng, Weiyang, and Yanta. Figure 1c provides a detailed visualization of the dataset’s coverage within Xi’an’s second ring road, where red points represent the trajectory of a single vehicle recorded over one day. These trajectories are composed of GPS sampling points collected at intervals of 2–4 s, offering high-resolution data for spatiotemporal analysis. The data from 1 October 2016, to 30 November 2016, were collected from approximately 18,000 ride-hailing vehicles.

2.2. Data Preprocessing

Data preprocessing is essential for ensuring the accuracy and reliability of TTP models. This study’s preprocessing workflow includes map matching, UTC conversion, anomaly removal, and trajectory refinement to construct a high-quality dataset.
Map matching (MM) and path inference were conducted using the algorithm proposed by Chen et al. [35]. This algorithm converts latitude and longitude into geodetic coordinates to improve trajectory accuracy, ensuring precise distance calculations and path alignment. Additionally, UTC timestamps were transformed into daily time values (0–86,400 s) to facilitate temporal feature extraction.
Rigorous data cleaning was conducted to exclude abnormal trips and outliers based on predefined criteria, including unreasonable travel speeds, excessively short or long travel times, and erroneous GPS coordinates, as outlined in Ref. [36]. After preprocessing, the dataset was reduced from 6,584,397 trips to 6,203,848 valid trips, excluding approximately 6% of the original records. Detailed OD trip data were extracted, including origin and destination coordinates, departure and arrival times, travel time, and distance. Table 1 provides an overview of the dataset attributes, forming a robust foundation for subsequent analyses and model development.

2.3. Feature Extraction

Accurate TTP relies on capturing the multidimensional variability of influencing factors. This study extracts four main categories of features: spatial, temporal, weather, and driver behavior features. These features provide a robust representation of travel dynamics, ensuring the model can account for diverse influencing factors.
Spatial features define the geographical context of trips and serve as fundamental inputs for OD-based travel time modeling. The key spatial attributes include origin location ( o i ), destination location ( d i ), and trip length ( l i ). This compact representation simplifies the spatial aspect of trip characteristics while preserving essential geospatial information. For a trip i, the spatial feature vector is represented as F S = o i , d i , l i . This concise yet practical representation of spatial features forms the basis for analyzing and predicting travel time. It enables the model to incorporate essential geospatial information without overcomplicating the feature space.
Temporal features are pivotal in capturing traffic variations across different time scales. These features reflect fine-grained (intra-day variations) and coarse-grained (day-of-week or holiday-based) traffic dynamics. The departure time ( s i ) is recorded as the number of seconds elapsed since midnight (00:00:00), ranging from 0 to 86,400 s. The day of the week ( t d i ) is categorized into three weekday groups (Monday, Tuesday–Thursday, and Friday) and two weekend groups (Saturday and Sunday). A holiday indicator is also included to account for variations during public holidays.
The temporal feature vector for trip i is expressed as F T = s i , t d i . Additionally, a 4-bit binary encoding is used for day-of-week classification, where the first three bits indicate the weekday category, and the fourth bit denotes a holiday (0 for holidays and 1 for non-holidays), as shown in Table 2. For instance, a binary value of “0000” signifies Monday during a holiday, while “0011” indicates Tuesday/Wednesday/Thursday on a regular weekday.
Weather conditions significantly affect travel time variability. To systematically incorporate weather influences, this study adopts a hybrid approach, combining qualitative classification and quantitative impact estimation. The Kolmogorov–Smirnov (K-S) test groups weather conditions based on their impact on travel time. By analyzing travel time distributions across different weather scenarios, the original 12 weather types were consolidated into three categories: Category 1 includes fair, cloudy, partly cloudy, and mostly cloudy conditions; Category 2 comprises rain, light rain, snow, and light snow; and Category 3 consists of fog, haze, mist, and widespread dust. Average travel time (ATT) represents the mean travel time per kilometer under specific weather conditions, while standard deviation of travel time (SDTT) reflects the variability of travel times, highlighting congestion disparities across the network.
The weather feature vector is defined as F W T O D , D O W , W = { A T T , S D T T } , where ATT and SDTT are computed using a Random Forest model, incorporating departure time (TOD), day-of-week (DOW), and weather category ( W ). The results indicate that adverse weather conditions (e.g., rain or fog) lead to higher ATT and SDTT values, particularly during peak hours. These effects highlight the significant impact of weather variability on urban travel time.
Driver behavior significantly affects travel time variability. To systematically classify driving styles, this study applies K-means clustering based on speed and acceleration metrics extracted from GPS trajectory data. Three distinct driving styles were identified: aggressive, conservative, and neutral. Aggressive drivers exhibit higher speed variability and frequent acceleration changes, while conservative drivers show lower maximum speeds and more minor fluctuations in acceleration. Neutral drivers demonstrate moderate behavior between these two extremes.
For a driver j, the driver feature vector is expressed as F D = { p 1 j , p 2 j , p 3 j } , where p i j represents the probability of a driver belonging to one of the three clusters. These probabilities are derived based on the distances between the driver’s behavior feature vector ( x j ) and the cluster centroids ( C i ):
p i j = 1 / d i j ∑ i = 1 3 1 / d i j
where d i j = | | x j − C i | | 2 2 denotes the Euclidean distance between the driver’s feature vector and the centroid of cluster i. For drivers without historical trajectory data, equal probabilities (one-third) are assigned across all clusters, ensuring all drivers are considered even in the absence of prior data. This driver behavior modeling enhances predictive accuracy by incorporating personalized travel time variability, making the OD-based prediction framework more robust. Table 3 provides a concise summary of the extracted features.

2.4. Problem Statement

Accurate TTP is essential for improving traffic efficiency, congestion management, and intelligent transportation planning. However, existing OD-based models still face two key limitations: limited utilization of external factors and lack of adaptive feature representation. To address these challenges, this study formulates the OD-based TTP (OD-TTP) problem as a spatiotemporal regression task, incorporating high-resolution ride-hailing trajectory data and multi-dimensional external features.
Problem definition: Given a historical trip database T P containing m trips, where each trip   t p i consists of an origin o i , destination d i , starting time s i , travel time t i , and associated multi-dimensional feature set, the objective is to predict the travel time t ^ q for a query t p q = o q , d q , s q .
To enhance predictive accuracy, this study integrates spatiotemporal, weather, and driver behavior features into a Back Propagation Neural Network (BPNN) optimized via Bayesian algorithms, ensuring a robust and adaptive modeling framework. The next section details the model construction and implementation.

3. Model Construction and Implementation

The proposed OD-TTP model leverages high-resolution ride-hailing trajectory data and multi-dimensional external factors (weather, spatiotemporal conditions, and driver behavior) to enhance prediction accuracy. A Back Propagation Neural Network (BPNN) is the core prediction model optimized using Bayesian hyperparameter tuning. This section details the model framework, training procedure, and optimization strategy, ensuring robust and scalable predictions for urban transportation systems.
Figure 2 provides an overview of the proposed methodology, structured into three main components. The first component involves data preparation, including processing historical OD and GPS trajectory data to extract relevant features such as spatiotemporal characteristics, weather factors, and driver behavior. The second component focuses on training the BPNN model using the prepared dataset, ensuring that the network effectively captures the complex relationships between trip attributes and travel time. The third component applies the trained BPNN model to predict travel times for future trips, enabling practical and accurate OD-TTP solutions.

3.1. Back Propagation Neural Networks (BPNNs)

While BPNN is a classical machine learning algorithm, its application in OD-TTP has been relatively underexplored compared to more computationally intensive approaches such as Graph Neural Networks (GNNs) [37] and transformer-based models [38]. While these advanced models offer strong predictive capabilities, they require extensive computational resources and complex training pipelines. In contrast, this study optimizes BPNN using Bayesian algorithms to enable efficient hyperparameter tuning and adaptive learning rates, enhancing predictive robustness while maintaining computational efficiency. The BPNN is a widely used feedforward neural network that captures complex nonlinear dependencies between input features and travel time. The network consists of three primary components: an input layer, one or more hidden layers, and an output layer. The input layer encodes key OD-related trip attributes and external factors. The hidden layers capture hierarchical representations of traffic patterns and influencing factors. Finally, the output layer generates the predicted travel time. The network’s structure, as shown in Figure 3, ensures efficient data processing and error correction, making it suitable for high-dimensional and nonlinear data applications.
The BPNN training process follows a standard backpropagation optimization, which iteratively updates network weights to minimize the prediction error. The objective function is defined as the mean squared error (MSE):
L ( θ ) = 1 n   ∑ i = 1 n ( t i − t ^ i ) 2
where t i and t ^ i represent the actual and predicted travel times, respectively. Weight updates are performed using the gradient descent algorithm:
w i + 1 = w i − η ∂ L ( θ ) ∂ w
where w i and w i + 1 are the weights at the ith and i + 1th iteration, η is the learning rate controlling the step size of each update, and ∂ L ( θ ) ∂ w is the gradient of the loss function with respect to the weight.

3.2. Model Structure and Implementation

To achieve optimal prediction accuracy and computational efficiency, the BPNN model is designed with an input layer that encodes spatial (origin, destination, and distance), temporal (departure time and weekday category), weather (ATT and SDTT), and driver behavior features. Two hidden layers with ReLU activation, containing 366 and 787 neurons, are employed to capture complex relationships while balancing model complexity and generalization. The output layer uses the Identity activation function to produce direct TTPs. Model performance is evaluated using RMSE as the loss function, while the Adam optimizer ensures fast convergence through adaptive learning rates.
The BPNN model is implemented using the TensorFlow framework in Python 3.8, with PyCharm as the integrated development environment (IDE) for coding and execution. The computational setup includes an Intel Core i9-13900HX CPU (5.40 GHz), an NVIDIA RTX 4070 Ti GPU, and 64 GB of RAM, ensuring efficient handling of large-scale data and computationally intensive training processes.

3.3. Bayesian Hyperparameter Optimization

To improve model robustness, Bayesian optimization is employed to tune hyperparameters efficiently. In terms of convergence speed and accuracy, it outperforms grid search and random search. The algorithm models the relationship between hyperparameters and performance using a Gaussian process (GP), iteratively refining the search space.
Table 4 presents the final selected hyperparameters, balancing computational efficiency and accuracy. The number of neurons in the first and second hidden layers was selected from a range of 32 to 1024, with a step size of 1, resulting in optimal values of 366 and 787, respectively. The initial learning rate ( α 0 ) was tuned within a uniform distribution between 0.001 and 0.01, yielding an optimal value of 0.001683. The training batch size was optimized from 256 to 512, with a step size of 32, leading to a final value of 347. Lastly, the maximum number of iterations was chosen from 50 to 100, with a step size of 1, resulting in 72 iterations for convergence. Bayesian optimization dynamically refines hyperparameter selection, ensuring the model achieves high predictive performance while minimizing computational overhead.

4. Results and Discussion

4.1. Evaluation Metrics

To comprehensively evaluate the predictive performance of the proposed model, we employ four commonly used metrics: root mean squared error (RMSE), mean absolute error (MAE), coefficient of determination (R2), and mean absolute percentage error (MAPE). Their mathematical formulations are as follows:
R M S E = 1 n ∑ i = 1 n ( t i − t ^ i ) 2        
R 2 = 1 − ∑ i = 1 n t i − t ^ i 2 ∑ i = 1 n t i − t ¯ i 2
M A E = 1 n ∑ i = 1 n t i − t ^ i
M A P E = 1 n ∑ i = 1 n t i − t ^ i t i × 100 %      

4.2. Ablation Study and Performance Analysis

Furthermore, comprehensive ablation studies are conducted to systematically assess the contributions of different feature sets—including spatiotemporal, weather, and driver behavior—providing valuable insights into feature importance and improving model interpretability. Table 5 summarizes the different model configurations.
The results from the ablation studies demonstrate the effectiveness of incorporating external factors. The baseline model, which only considers spatial–temporal attributes, achieves an RMSE of 327.57 s and an R2 of 0.578, indicating limited predictive capability. As features are added, the model performance improves significantly, with the full model (Study 5) achieving an RMSE of 202.89 s and an R2 of 0.838. These improvements validate the necessity of incorporating weather and driver behavior features to enhance prediction accuracy.
Table 6 presents a comparative analysis of all ablation experiments. Integrating additional spatial–temporal attributes in Study 1 reduces the RMSE to 261.11 s and improves R2 to 0.732. Including quantified weather features in Study 2 further improves performance, reducing the RMSE to 244.32 s. Compared to raw weather attributes (Study 3), quantified weather metrics demonstrate superior predictive capability, confirming the advantage of transforming raw weather data into structured features. Study 4 highlights the importance of driver behavior, achieving an RMSE of 224.13 s, demonstrating that individual driving tendencies contribute significantly to travel time variability.
The improvement from the baseline to the full model underscores the importance of integrating multiple external features. The largest performance gain is observed when quantified weather metrics and driver behavior features are introduced, demonstrating their significant contribution to reducing prediction errors.
The spatial–temporal features serve as the foundation of the model, capturing essential route characteristics such as distance and departure time. While these features provide valuable insights, they cannot fully explain travel time variability. Incorporating weather factors further refines the model, as adverse weather conditions directly impact road traffic and driver responses. The quantified metrics for average travel time and its standard deviation (ATT and SDTT) outperform raw weather attributes, suggesting that structured representations of weather data enable the model to better capture external influences.
Driver behavior is crucial in travel time estimation, as individual driving styles significantly impact speed variations and acceleration patterns. The clustering-based driver classification method effectively represents these behavioral tendencies, contributing to improved model generalization. Ablation results indicate that driver behavior features reduce RMSE by approximately 20 s compared to models without them, reinforcing their importance in real-world TTPs.

4.3. Robustness and Generalization Analysis

The robustness and generalization capabilities of the proposed BPNN model were evaluated by analyzing prediction accuracy under varying travel conditions. This section examines the model’s performance across different travel distances, times, and speeds. Additionally, a cumulative distribution analysis of absolute percentage error (APE) and a comprehensive robustness analysis provide further insights into the model’s effectiveness and areas for improvement.

4.3.1. Cumulative Distribution of Absolute Percentage Error (APE)

Figure 4a presents the cumulative distribution function (CDF) of the absolute percentage error (APE) for the proposed model. Key percentile markers show that 10% of predictions have an APE below 2.39%, 50% (median) are below 13.12%, and 90% are below 35.29%. This indicates that most predictions are highly accurate, with minimal deviation from actual values.
The steep initial rise in the CDF reflects the model’s robustness for most predictions. However, the curve’s tail reveals challenges in achieving consistent accuracy for outlier scenarios, likely caused by irregular travel patterns or extreme weather and driver behaviors. These findings highlight the model’s overall effectiveness while identifying opportunities for improvement in edge cases.

4.3.2. Influence of Travel Distance on Prediction Accuracy

Figure 4b shows the relationship between MAPE and travel distance. The MAPE stabilizes around the global average of 16.89% for trips between 2.5 km and 12.5 km, where the model achieves its highest accuracy. Short trips (<2.5 km) exhibit higher errors due to limited spatial variability, which makes predictions more sensitive to minor factors like intersection delays. The MAPE increases for longer trips (>15 km) due to more significant variability in traffic, weather, and driver behavior over extended durations.
These results underscore the model’s optimized performance for medium-range trips. However, the increased errors for extreme distances suggest additional contextual factors, such as dynamic traffic data, to improve predictions in these scenarios.

4.3.3. Influence of Travel Time on Prediction Accuracy

Figure 4c illustrates the effect of travel time on prediction accuracy. Short-duration trips (<5 min) show higher MAPE values due to the disproportionate impact of initial delays, such as congestion near the origin. The error stabilizes for trips between 5 and 25 min, where the model achieves its lowest MAPE, demonstrating robust performance in this range.
For longer trips (>25 min), the error increases steadily, driven by the cumulative effect of unpredictable factors like weather changes, driver behavior variations, and traffic congestion. This trend aligns with the findings for travel distance, emphasizing the need for more advanced features to address variability in extended travel times.

4.3.4. Influence of Travel Speed on Prediction Accuracy

The MAPE increases for high-speed trips (>55 km/h) due to more significant variability in travel time over longer distances and reduced frequency of historical data samples. This result emphasizes improving the model’s performance in predicting high-speed scenarios, potentially through advanced data augmentation techniques or additional features.
Figure 4d analyzes the relationship between MAPE and travel speed. The MAPE is significantly higher for low-speed trips (<15 km/h), reflecting the challenges of stop-and-go conditions and frequent delays. Trips with average speeds between 15 and 55 km/h exhibit stable and low MAPE values, near the global average, demonstrating the model’s ability to handle consistent travel speeds effectively.
The MAPE increases for high-speed trips (>55 km/h), likely due to reduced historical data samples and variability over long distances. These results highlight the importance of targeted data augmentation and scenario-specific modeling to improve accuracy in low-speed urban and high-speed scenarios.

4.3.5. Comprehensive Analysis of Model Robustness

The robustness of the proposed BPNN model is evaluated through a combined analysis of travel distance, travel time, and travel speed, as shown in Figure 5 and Table 7. Figure 5 illustrates the MAPE distribution across travel distance and travel time, highlighting the model’s optimal accuracy for medium-range trips and moderate durations. However, the prediction error increases for short distances, long travel times, and extreme speeds, revealing the need for further enhancements in these conditions.
Table 7 compares model performance across different travel speeds. For low-speed trips (≤12 km/h), which constitute 11.8% of the dataset, the model achieves an RMSE of 488.16 s, an R2 of 0.56, and a MAPE of 31.19%, reflecting reduced accuracy due to stop-and-go conditions and high variability. In contrast, for trips with speeds >12 km/h (88.2% of the dataset), the model performs significantly better, with an RMSE of 151.21 s, an R2 of 0.88, and a MAPE of 14.57%, demonstrating its capability in more consistent scenarios.
These findings confirm that the training data distribution strongly influences the model’s predictive accuracy. Medium-range trips, moderate durations, and higher speeds, which are well represented in the dataset, yield the best predictions. Underrepresented scenarios, such as short trips, low speeds, or long durations, present higher errors due to limited training examples and increased variability.

5. Conclusions

This study proposes an OD-based TTP framework that integrates spatiotemporal, weather, and driver behavior features within a Back Propagation Neural Network (BPNN). By leveraging high-resolution ride-hailing trajectory data, the model effectively captures the complexities of urban traffic dynamics and significantly enhances predictive accuracy. The results demonstrate that incorporating multi-dimensional external factors improves the model’s robustness, achieving a root mean squared error (RMSE) of 202.89 s and a mean absolute percentage error (MAPE) of 16.52%. Notably, introducing quantified weather metrics and driver behavior clustering provides deeper insights into the external variability influencing travel time, reinforcing the necessity of integrating diverse contextual factors.
Despite its strong performance, the model exhibits limitations in extreme travel conditions, such as short-distance trips, low-speed scenarios, or prolonged travel durations. These shortcomings primarily stem from data imbalances, where underrepresented travel patterns introduce variability that the model struggles to capture effectively. To address these challenges, we propose future enhancements, including real-time congestion metrics, more granular road network characteristics, and data augmentation techniques to mitigate data sparsity and improve prediction accuracy across all trip distances.
Future research will enhance the model’s generalizability and scalability by incorporating diverse data sources and refining feature representations. Expanding the dataset to include travel scenarios from multiple cities, particularly for underrepresented conditions, will enhance adaptability. Additionally, integrating real-time traffic flow, dynamic road conditions, and finer-grained driver behavior metrics will refine predictive accuracy. To facilitate cross-city adaptation, transfer learning techniques and domain adaptation strategies will be explored, enabling the model to adjust to varying traffic dynamics. Furthermore, leveraging advanced deep learning architectures and developing scenario-specific models tailored to extreme conditions will enhance predictive robustness. Implementing real-time prediction capabilities will further strengthen its applicability in intelligent transportation systems, contributing to improved urban mobility, congestion mitigation, and dynamic traffic management.

Author Contributions

C.S. and Y.W. conceived the research idea. W.Z. and Z.Z. derived the model. Z.Z. and Y.Z. designed the experiments. Y.W., W.Z. and T.C. analyzed the results and drafted the manuscript. C.S., Y.Z. and N.W. edited the manuscript. All authors have read and agreed to the published version of the manuscript.

Funding

This research was supported by the National Natural Science Foundation of China (Grant Nos. 42101437 and 41903192), the Open Fund of the Engineering Laboratory of Spatial Information Technology for Highway Geological Disaster Early Warning in Hunan Province (Changsha University of Science & Technology, kfj230701), and the Open Fund of the Anhui Province Key Laboratory of Physical Geographic Environment at Chuzhou University (Grant No. 2023PGE004).

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

Publicly available datasets were analyzed in this study. These data can be found here: https://gaia.didichuxing.com (accessed on 1 January 2025).

Acknowledgments

The authors are grateful to Didi Chuxing for providing the data used in this paper. They also would like to thank the anonymous reviewers and editors for giving valuable comments and suggestions, which helped significantly improve the manuscripts. Additionally, the authors sincerely appreciate Lu Ao and Wang Shixin for their valuable contributions to the formation of this paper and the design of the algorithms.

Conflicts of Interest

The authors declare no conflicts of interest.

Abbreviations

The following abbreviations are used in this manuscript:
ODOrigin–destination
ITSsIntelligent transportation systems
BPNNBack Propagation Neural Network
TTPTravel time prediction
ATTAverage travel time
SDTTStandard deviation of travel time
TODTime of day
DOWDay of week
RMSERoot mean squared error
MAEMean absolute error
MAPEMean absolute percentage error

References

  1. Wang, S.F.; Li, L.; Ma, W.J.; Chen, X.Q. Trajectory analysis for on-demand services: A survey focusing on spatial-temporal demand and supply patterns. Transp. Res. Part C Emerg. Technol. 2019, 108, 74–99. [Google Scholar] [CrossRef]
  2. Luo, S.J.; Zou, F.M.; Zhang, C.; Tian, J.S.; Guo, F.; Liao, L.C. Multi-View Travel Time Prediction Based on Electronic Toll Collection Data. Entropy 2022, 24, 1050. [Google Scholar] [CrossRef] [PubMed]
  3. Ta, X.; Liu, Z.; Hu, X.; Yu, L.; Sun, L.; Du, B. Adaptive spatio-temporal graph neural network for traffic forecasting. Knowl.-Based Syst. 2022, 242, 108199. [Google Scholar] [CrossRef]
  4. Zhong, R.X.; Chen, C.; Huang, Y.P.; Sumalee, A.; Lam, W.H.K.; Xu, D.B. Robust perimeter control for two urban regions with macroscopic fundamental diagrams: A control-lyapunov function approach. Transp. Res. Part B Methodol. 2018, 117, 687–707. [Google Scholar] [CrossRef]
  5. Zhang, S.; Tang, J.J.; Wang, H.X.; Wang, Y.H.; An, S. Revealing intra-urban travel patterns and service ranges from taxi trajectories. J. Transp. Geogr. 2017, 61, 72–86. [Google Scholar] [CrossRef]
  6. Chughtai, J.; Haq, I.; Islam, S.; Gani, A. A Heterogeneous Ensemble Approach for Travel Time Prediction Using Hybridized Feature Spaces and Support Vector Regression. Sensors 2022, 22, 9735. [Google Scholar] [CrossRef] [PubMed]
  7. Li, Y.G.; Fu, K.; Wang, Z.; Shahabi, C.; Ye, J.P.; Liu, Y. Multi-task Representation Learning for Travel Time Estimation. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, London, UK, 19–23 August 2018. [Google Scholar]
  8. Yuan, H.; Li, G. A survey of traffic prediction: From spatio-temporal data to intelligent transportation. Data Sci. Eng. 2021, 6, 63–85. [Google Scholar] [CrossRef]
  9. Afrin, T.; Yodo, N. A long short-term memory-based correlated traffic data prediction framework. Knowl.-Based Syst. 2022, 237, 107755. [Google Scholar] [CrossRef]
  10. An, J.Y.; Guo, L.; Liu, W.; Fu, Z.Q.; Ren, P.; Liu, X.Z.; Li, T. IGAGCN: Information geometry and attention-based spatiotemporal graph convolutional networks for traffic flow prediction. Neural Netw. 2021, 143, 355–367. [Google Scholar] [CrossRef]
  11. Huang, Y.; Xu, L.; Qiang, L.; Kuang, X. Urban expressway travel time prediction method based on fuzzy adaptive kalman filter. Appl. Math. Inform. Sci. 2013, 7, 625–630. [Google Scholar] [CrossRef]
  12. Jin, G.; Yan, H.; Li, F.; Huang, J.; Li, Y. Spatio-temporal dual graph neural networks for travel time estimation. ACM Trans. on Spat. Algor. 2024, 10, 1–22. [Google Scholar] [CrossRef]
  13. Kolidakis, S.; Botzoris, G.; Profillidis, V.; Lemonakis, P. Road traffic forecasting—A hybrid approach combining artificial neural network with singular spectrum analysis. Econ. Anal. Policy 2019, 64, 159–171. [Google Scholar] [CrossRef]
  14. Lu, Z.L.; Lv, W.F.; Xie, Z.P.; Du, B.W.; Xiong, G.X.; Sun, L.L.; Wang, H.Q. Graph sequence neural network with an attention mechanism for traffic speed prediction. ACM Trans. Intell. Syst. Tec. 2022, 13, 20. [Google Scholar] [CrossRef]
  15. Shi, C.; Chen, B.Y.; Li, Q.Q. Estimation of travel time distributions in urban road networks using low-frequency floating car data. ISPRS Int. Geo-Inf. 2017, 6, 253. [Google Scholar] [CrossRef]
  16. Dion, F.; Rakha, H. Estimating dynamic roadway travel times using automatic vehicle identification data for low sampling rates. Transp. Res. Part B Methodol. 2006, 40, 745–766. [Google Scholar] [CrossRef]
  17. Fei, X.; Zhang, Y.; Liu, K.; Guo, M. Bayesian dynamic linear model with switching for real-time short-term freeway travel time prediction with license plate recognition data. J. Transp. Eng.-ASCE 2013, 139, 1058–1067. [Google Scholar] [CrossRef]
  18. Tam, M.L.; Lam, W.H.K. Application of automatic vehicle identification technology for real-time journey time estimation. Inform. Fusion 2011, 12, 11–19. [Google Scholar] [CrossRef]
  19. Shi, C.; Chen, B.Y.; Lam, W.H.K.; Li, Q.Q. Heterogeneous data fusion method to estimate travel time distributions in congested road networks. Sensors 2017, 17, 2822. [Google Scholar] [CrossRef] [PubMed]
  20. Jedwanna, K.; Athan, C.; Boonsiripant, S. Estimating toll road travel times using segment-based data imputation. Sustainability 2023, 15, 13042. [Google Scholar] [CrossRef]
  21. Lai, S.; Xu, H.; Luo, Y.; Zou, F.; Hu, Z.; Zhong, H. Expressway Vehicle Arrival Time Estimation Algorithm Based on Electronic Toll Collection Data. Sustainability 2024, 16, 5581. [Google Scholar] [CrossRef]
  22. Khaled, A.; Elsir, A.M.T.; Shen, Y. GSTA: Gated spatial–temporal attention approach for travel time prediction. Neural Comput. Appl. 2022, 34, 2307–2322. [Google Scholar] [CrossRef]
  23. Sun, Y.; Wang, Y.; Fu, K.; Wang, Z.; Yan, Z.; Zhang, C.; Ye, J. FMA-ETA: Estimating travel time entirely based on FFN with attention. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Toronto, ON, Canada, 6–11 June 2021. [Google Scholar]
  24. Wang, Z.; Fu, K.; Ye, J.P. Learning to Estimate the Travel Time. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, London, UK, 19–23 August 2018. [Google Scholar]
  25. Derrow-Pinion, A.; She, J.; Wong, D.; Lange, O.; Hester, T.; Perez, L.; Nunkesser, M.; Lee, S.; Guo, X.; Wiltshire, B.; et al. ETA prediction with graph neural networks in google maps. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, Gold Coast, Australia, 1–5 November 2021. [Google Scholar]
  26. Poongodi, M.; Malviya, M.; Kumar, C.; Hamdi, M.; Vijayakumar, V.; Nebhen, J.; Alyamani, H. New York City taxi trip duration prediction using MLP and XGBoost. Int. J. Syst. Assur. Eng. 2022, 13, 16–27. [Google Scholar] [CrossRef]
  27. Radhika, K.R.; Sheela, S.V.; Jha, T.; Sushant, V.S.; Rajibul, S. Dynamic Graph Convolutional Network-based Long Short-Term Memory for Taxi Trip Time Prediction. In Proceedings of the 2024 First International Conference on Software, Systems and Information Technology, Tumkur, India, 17–18 October 2025. [Google Scholar]
  28. Li, X.; Wang, H.; Sun, P.; Zu, H. Spatiotemporal features—Extracted travel time prediction leveraging deep-learning-enabled graph convolutional neural network model. Sustainability 2021, 13, 1253. [Google Scholar] [CrossRef]
  29. Wang, R.; Guo, Q.; Dai, S.; Deng, L.; Xiao, Y.; Jia, C. An OD time prediction model based on adaptive graph embedding. Phys. A 2025, 657, 130217. [Google Scholar] [CrossRef]
  30. Li, A.; Lam, W.H.; Ma, W.; Tam, M.L. A novel modelling framework for real-time prediction of path travel times using both traffic and weather data. Transp. A 2024, 1–34. [Google Scholar] [CrossRef]
  31. Zheng, L.; Ma, H.; Wang, Z. Travel Time Estimation for Urban Arterials Based on the Multi-Source Data. Sustainability 2024, 16, 7845. [Google Scholar] [CrossRef]
  32. Lu, F.; Du, Z.; Wang, Z.; Wang, L.; Wang, S. Towards enhancing the crowdsourcing door-to-door delivery: An effective model in Beijing. J. Ind. Manag. Optim. 2025, 21, 2371–2395. [Google Scholar] [CrossRef]
  33. Lagos, F.; Moreno, S.; Yushimito, W.F.; Brstilo, T. Urban Origin–Destination Travel Time Estimation Using K-Nearest-Neighbor-Based Methods. Mathematics 2024, 12, 1255. [Google Scholar] [CrossRef]
  34. Fang, H.; Liu, Y.; Chen, C.H.; Hwang, F.J. Travel Time Prediction Method Based on Spatial-Feature-based Hierarchical Clustering and Deep Multi-input Gated Recurrent Unit. ACM Trans. Sens. Netw. 2022, 19, 1–21. [Google Scholar] [CrossRef]
  35. Chen, B.Y.; Yuan, H.; Li, Q.Q.; Lam, W.H.K.; Shaw, S.L. Map-matching algorithm for large-scale low-frequency floating car data. Int. J. Geogr. Inf. Sci. 2014, 28, 22–38. [Google Scholar] [CrossRef]
  36. Shi, C.; Li, Q.; Lu, S.; Yang, X. Exploring Temporal Intra-Urban Travel Patterns: An Online Car-Hailing Trajectory Data Perspective. Remote Sens. 2021, 13, 1825. [Google Scholar] [CrossRef]
  37. Wang, D.; Zhu, J.; Yin, Y.; Ignatius, J.; Wei, X.; Kumar, A. Dynamic travel time prediction with spatiotemporal features: Using a gnn-based deep learning method. Ann. Oper Res. 2024, 340, 571–591. [Google Scholar] [CrossRef]
  38. Liu, S.; Zhang, Y.; Wang, Z.; Liu, X.; Yang, H. Personalized origin–destination travel time estimation with active adversarial inverse reinforcement learning and Transformer. Transp. Res. Part E Logist. Transp. Rev. 2025, 193, 103839. [Google Scholar] [CrossRef]
Sustainability 17 02111 g001
Figure 1. Spatial information of the study area: (a) Location of Xi’an; (b) The central area of Xi’an; (c) The road network of the study area.
Figure 1. Spatial information of the study area: (a) Location of Xi’an; (b) The central area of Xi’an; (c) The road network of the study area.
Sustainability 17 02111 g001
Sustainability 17 02111 g002
Figure 2. Overview of the proposed methodology.
Figure 2. Overview of the proposed methodology.
Sustainability 17 02111 g002
Sustainability 17 02111 g003
Figure 3. Structure of the BPNN model.
Figure 3. Structure of the BPNN model.
Sustainability 17 02111 g003
Sustainability 17 02111 g004
Figure 4. Performance of the proposed method: (a) CDF of the absolute percentage error; (b) Influence of travel distance on MAPE; (c) Influence of travel time on MAPE; (d) Influence of travel speed on MAPE.
Figure 4. Performance of the proposed method: (a) CDF of the absolute percentage error; (b) Influence of travel distance on MAPE; (c) Influence of travel time on MAPE; (d) Influence of travel speed on MAPE.
Sustainability 17 02111 g004
Sustainability 17 02111 g005
Figure 5. MAPE distribution with travel time and distance.
Figure 5. MAPE distribution with travel time and distance.
Sustainability 17 02111 g005
Table 1. Overview of OD order data fields.
Table 1. Overview of OD order data fields.
No.Field NameDescriptionExample 1Example 2Example 3Example 4
1VehicleVehicle ID1,396,5691,009,0121,001,7421,017,095
2OrderIDOrder ID22,652,14220,060,25920,011,4212,011,400
3TimeOriginDeparture time (s)49,77362,67130,49764,809
4TimeDestinationArrival time (s)49,93573,78030,56064,867
5RoadIDOriginThe link ID of the origin29,35647,66020,00758,449
6mmXTransformed coordinate of the origin497,452.30498,142.71491,888.56499,857.70
7mmY3,794,644.893794,621.433,790,399.543,787,044.29
8RoadIDDestinationThe link ID of the destination29,91063,37242,33862,597
9mmXTransformed coordinate of the destination498,141.79499,612.0649,248.83499,309.39
10mmY3,794,600.353,794,602.393,790,394.113,787,475.35
11PathTravelTimeOD travel time (s)16211,109 (>7200 s)6358 (<60 s)
12PathLengthTravel distance (m)991.2412,162.971476.81719.35
13TravelSpeedSpeed (km/h)22.033.94 (<5)84.34 (>80)44.65
Note: Bold values indicate invalid orders due to specific anomalies, including excessively low or high speeds, excessively long or short travel times. These cases were excluded during the data preprocessing stage to ensure data reliability and model accuracy.
Table 2. Quantified coarse-grained temporal features.
Table 2. Quantified coarse-grained temporal features.
BitNumberDate
Bits 1–3000Monday
001Tuesday/Wednesday/Thursday
010Friday
011Saturday
100Sunday
Bit 40Holiday
1Non-holiday
Table 3. Summary of the extracted features.
Table 3. Summary of the extracted features.
Feature CategoryFeaturesDescription
Spatial F S = o i , d i , l i Origin, destination, and trip length
Temporal F T = s i , t d i Departure time and day-of-week classification
Weather F W T O D , D O W , W = { A T T , S D T T } Average travel time and standard deviation
under different weather conditions
Driver Behavior F D = { p 1 j , p 2 j , p 3 j } Probability distribution of driving styles
(aggressive, conservative, and neutral)
Table 4. Parameter space and optimal parameters.
Table 4. Parameter space and optimal parameters.
ParameterParameter SpaceParameter DistributionOptimized Parameter
Number of neurons in hidden layer 1 [ 32 ,   1024 ] Step size: 1366
Number of neurons in hidden layer 2 [ 32 ,   1024 ] Step size: 1787
Initial learning rate [ 0.001 ,   0.01 ] Uniform distribution0.001683
Training batch size [ 256 ,   512 ] Step size: 32347
Maximum iterations [ 50 ,   100 ] Step size: 172
Table 5. The configurations for each ablation study.
Table 5. The configurations for each ablation study.
ModelSpatial FeatureTemporal FeatureWeather FeatureDriver Feature
OD
Location		Travel
Distance		Starting
Time		Travel
Date		Raw Weather
Feature		Quantitative
Evaluation with RF
Baseline model✓ ✓
Study 1✓✓✓✓
Study 2✓✓✓✓ ✓
Study 3✓✓✓✓✓
Study 4✓✓✓✓ ✓
Study 5✓✓✓✓ ✓✓
Note: The checkmark (✓) indicates that the model includes the corresponding feature in the prediction process.
Table 6. Performances of different ablation studies.
Table 6. Performances of different ablation studies.
Ablation StudiesEvaluation Metrics
RMSE (s) R 2 MAE (s)MAPE (%)
Baseline 327.570.578191.8924.09
Study 1: Baseline+Extended Spatial–Temporal Features261.110.732151.7418.37
Study 2: Study 1+Quantified Weather Features244.320.766142.9717.69
Study 3: Study 1+Raw Weather Features256.590.742148.8418.19
Study 4: Study 1+Driver Features224.130.803133.2716.89
Study 5 (BPNN): Study 1+Combined Features202.890.838130.4816.52
Table 7. Performance indexes of prediction models at different travel speeds.
Table 7. Performance indexes of prediction models at different travel speeds.
SpeedNumber of Trips R M S E   ( s ) R 2 M A E (s)MAPE (%)
≤ 12   k m / h 142,541 (11.8%)488.160.56359.1231.19
> 12   k m / h 1,070,308 (88.2%)151.210.88100.0414.57
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Shi, C.; Zou, W.; Wang, Y.; Zhu, Z.; Chen, T.; Zhang, Y.; Wang, N. Enhancing Travel Time Prediction for Intelligent Transportation Systems: A High-Resolution Origin–Destination-Based Approach with Multi-Dimensional Features. Sustainability 2025, 17, 2111. https://doi.org/10.3390/su17052111

AMA Style

Shi C, Zou W, Wang Y, Zhu Z, Chen T, Zhang Y, Wang N. Enhancing Travel Time Prediction for Intelligent Transportation Systems: A High-Resolution Origin–Destination-Based Approach with Multi-Dimensional Features. Sustainability. 2025; 17(5):2111. https://doi.org/10.3390/su17052111

Chicago/Turabian Style

Shi, Chaoyang, Waner Zou, Yafei Wang, Zhewen Zhu, Tengda Chen, Yunfei Zhang, and Ni Wang. 2025. "Enhancing Travel Time Prediction for Intelligent Transportation Systems: A High-Resolution Origin–Destination-Based Approach with Multi-Dimensional Features" Sustainability 17, no. 5: 2111. https://doi.org/10.3390/su17052111

APA Style

Shi, C., Zou, W., Wang, Y., Zhu, Z., Chen, T., Zhang, Y., & Wang, N. (2025). Enhancing Travel Time Prediction for Intelligent Transportation Systems: A High-Resolution Origin–Destination-Based Approach with Multi-Dimensional Features. Sustainability, 17(5), 2111. https://doi.org/10.3390/su17052111

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop