Traffic incidents encompass a variety of events, including traffic accidents, vehicle malfunctions, vehicle fires, and arguments or fights in traffic. These incidents can lead to traffic congestion, which can have significant social, economic, and environmental impacts, such as increased travel time, excessive fuel consumption, air pollution, and stress [1,2]. Predicting the duration of traffic events is a challenging task due to the complexities arising from the stochastic nature of traffic events. Accurate duration prediction offers advantages to drivers in route selection and traffic operations managers in congestion management [3]. Efficient traffic incident management (TIM) is essential for mitigating adverse traffic effects. Continuous enhancement of TIM systems ensures effective incident handling and minimises traffic disruptions. Accurate estimation of event duration, reliant on environmental and event-specific analyses, is pivotal to TIM success. Precise duration estimation informs resource allocation and intervention planning. Diverting affected road users to alternative routes attenuates the incident impact [4].

Forecasting event duration is vital in traffic event management, representing the temporal gap from incident onset to clearance. It is essential for assessing the severity of the incident and determining the temporal and spatial distribution of traffic flow on the road network. Traffic incidents can be divided into sequential and distinct time intervals, as described by several studies [5–8]. The duration between the incident occurrence and the response of the traffic control center operators after receiving the call is known as the detection-reporting time. Preparation-dispatch time is the time between the receipt of the call by the operators and the dispatch of the response team members to the incident. The travel time is simply the duration between receiving the dispatch order and arriving at the scene for the incident response team members. Detection-notification time, preparation-dispatch time, and travel time are three important time intervals in incident response. Clean-up time is the time between the arrival of incident response team members at the scene and the completion of the cleanup of the incident. Clean-up time is especially used in planning and dispatching activities in traffic incident management. In this study, incident duration is defined as the time between the occurrence of the incident and the opening of the roadway.

During traffic incidents, TIM centers first attempt to collect incident coordinates and other relevant information. However, this estimation process is challenging due to the complexity of traffic events and the multitude of variables affecting duration. Uncertainty is inherent at the outset and escalates with incident size. The impact of different factors on incident duration during a traffic event or accident may vary depending on the circumstances, such as partial lane closure, complete road closure, or long-term road infrastructure works. TIMs may make a decision that no intervention is necessary in a low-level traffic incident and then make a similar assessment in a very similar incident that actually requires extensive intervention. Real-time data collection is essential to mitigate assessment errors. Machine learning algorithms are currently the most effective tools in prediction studies of traffic events [9].

Recent studies have focused on predicting traffic incident durations with the objective of enhancing resource allocation, emergency response, and traffic management [10–14]. These studies examine the sub-components of traffic incident durations, such as intervention, and scene cleaning. They utilize various methodologies, including machine learning models, hazard-based modeling, and ensemble learning approaches, in order to improve the accuracy and interpretability of incident duration predictions. Factors influencing incident duration include road type, casualties, weather conditions, and the number of vehicles. The necessity of considering time-varying traffic variables during incident episodes is emphasized, underscoring the importance of dynamic modeling to capture traffic flow dynamics. For real-time forecasting, variables for which real-time data can be collected should be examined and their success in forecasting investigated. The present study proposes an integrated methodology for dynamic estimating traffic incident duration, which addresses a significant gap in existing literature.

The majority of studies employ regression estimation for the purpose of predicting the duration of a traffic event. In contrast, classification is performed in a relatively limited number of studies [15]. In classification studies, the duration of the components of the traffic incident, such as traffic incident notification, response and traffic accident scene cleaning, was studied instead of the total time between the occurrence of the traffic incident and the cleaning of the scene [16]. Unlike previous studies focusing on regression, our novel approach employs classification methods for real-time prediction, reducing the initially examined variables to enhance predictive accuracy. Furthermore, the current study differs from previous research in that it considers the entire period from the occurrence of a traffic accident to the completion of the cleanup, rather than focusing on specific sub-components of traffic incident duration. The present study employs four machine learning methods to categorize traffic accidents and incidents into four duration classes. In order to facilitate dynamic prediction and more effective solutions, we initially reduced the number of variables, which had previously been extensive. Variables that were ineffective in forecasting were eliminated, and it was determined whether the variables that were effective in forecasting could collect real-time data. Subsequently, numerous databases containing the data of these variables were integrated, and a dynamic forecasting environment was provided. The categorization of incident durations and utilization of real-time data enables the swift intervention of incident management centers, fostering agile decision-making and resource allocation compared to conventional regression-based approaches. Feature selection enables the model to improve interpretability while also optimizing training and execution speed, thus mitigating the risk of overfitting. Ultimately, this study provides a pragmatic solution for dynamic traffic incident management, facilitating strategic interventions and resource optimization in urban transportation systems.

An experimental study was conducted in Istanbul to test the model proposed in the study to predict the duration of traffic events. Istanbul is one of the most crowded and traffic-heavy cities in the world, connecting two continents. To create the model, we identified the variables that affect incident duration and collected data from various sources. The objective is to facilitate prompt intervention from the incident management center by estimating the duration of incidents within a given time period using current conditions and primary data with minimal variables. To achieve this, Istanbul was divided into 682 geohash areas. The models were used to forecast the duration of traffic events during specific time period. The model incorporates dynamic prediction through feature selection to streamline complex data structures into fewer variables, facilitating real-time forecasting. Unlike other forecasting approaches, it responds promptly to traffic condition fluctuations by collecting influential variable data in a real-time database. Temporal and spatial considerations enable precise real-time predictions, enhancing the reliability and efficacy of traffic management and emergency response strategies.

The paper is structured as follows: Section 2 reviews prior studies on traffic incident duration, discussing methodologies and models. Section 3 outlines our proposed model, including the dataset, performance metrics, and methods employed. Section 4 presents experimental findings derived from the model. Finally, the concluding section provides objective evaluations of the study and suggests future research directions.

The prediction of traffic event duration plays a crucial role in enhancing TIM systems. Prior research has employed diverse regression models and statistical forecasting techniques for forecasting the duration of traffic incidents. Literature suggests that incident delay and duration exhibit variability contingent upon factors encompassing environmental conditions and incident-specific characteristics. Models predicting the duration of traffic incidents are based on various factors, such as incident type, time of day, weather conditions, and traffic volume. These models include regression models [10,15,17], cyclic subspace regression [18], and probabilistic statistical models [19–21]. Hazard-based models were studied by Hojati et al. [1,11,22–24]. Lee et al. [25] analyzed structural equation models, while Zou et al. [26,27] examined finite mixture models. Zou et al. [27] and Laman et al. [16] investigated copula-based models. The literature identifies incident characteristics (e.g., type of incident, first responder, and number of responders), road features (e.g., average annual daily traffic, geometric features, and functional classification), traffic situations (e.g., month, day, and time), and weather conditions (e.g., season, precipitation, temperature, and wind) as the most important independent variables for the developed models.

After a traffic incident, certain information, such as the number of injured individuals, number of vehicles involved, and cause of the accident, may not be immediately available. To improve model accuracy, a dynamic prediction model should be developed gradually using primary data such as the time and location of the incident, incident type, notification time, and weather conditions obtained after the incident notification. Secondary data, encompassing casualty counts, response time, and extent of lane obstruction, may serve as inputs for the initial prediction and subsequent refinement in a two-stage process [28]. Obtaining clear data on the number of fatalities or injuries may take days after the accident has occurred, and the scene may have been cleared during this time, making dynamic forecasting difficult.

Time-series methods can be used to calculate the duration of traffic incidents [29]. Various approaches and techniques, such as genetic algorithms [30], fuzzy logic [31], and Bayesian networks [32,33], can be employed in these models. However, machine learning algorithms have become more preferred over traditional methods in recent years due to the many variables that affect time and the constantly changing traffic conditions. Machine learning models are utilized to predict the duration of traffic incidents by analyzing historical data [9]. These models may undergo training utilising datasets incorporating variables such as the type of event, temporal occurrence, traffic density, meteorological parameters, and event duration. Several machine learning and data mining techniques have been employed to predict the duration of traffic incidents. These models are frequently used in decision trees [34,35], artificial neural networks [9,28], support vector machines [36,37], and random forests [9,12]. Li et al. [38] utilized deep learning in their recent studies. Deep learning, particularly Graph Neural Networks (GNNs), plays a pivotal role in Intelligent Transportation Systems (ITS). GNNs are extensively utilized in ITS applications due to their capacity to analyze graph-structured data effectively. They have evolved for a multitude of ITS tasks, including traffic forecasting, demand prediction, autonomous vehicles, intersection management, and urban planning [39]. Moreover, the integration of deep reinforcement learning (DRL) within connected and automated transportation systems has demonstrated potential in tasks related to automated driving systems and connected-vehicle applications [40]. These advancements in deep learning, including GNNs and DRL, are enhancing the efficiency, safety, and coordination of transportation modes within modern ITS infrastructure, thereby illustrating the potential of data-driven solutions to address complex challenges in the transportation domain.

Traffic incident duration can be divided into sequential time intervals, typically two, three, four, or five. Traffic incident durations are classified as either other or severe/major incidents. Lin et al. [41] defined durations as below or above 60 min, while Zhang et al. [42] defined them as below or above 120 min. The triple classification divides tasks into minor/short, medium, and major/long. According to Smith et al. [43], short tasks take less than 15 min, medium tasks take between 15 and 30 min, and long tasks take over 30 min. The US Department of Transportation [44] and Islam [45] define short tasks as taking less than 30 min, medium tasks as taking between 30–120 min, and long tasks as taking over 120 min. In this study traffic event durations were stratified into four distinct categories, delineated according to the frequency of minor events and the implementation of traffic event management protocols. Simple incidents are those that last less than 10 min and do not require intervention. Minor incidents are incidents that require intervention but last between 10 and 30 min, while mid-level incidents last between 31 and 60 min. The study considered incidents that required 61 min or more and a large amount of resources as major incidents.

Recent studies have shown an increase in research on estimating traffic incident duration. Unlike various statistical methods for designing data-driven models, machine learning techniques are frequently utilized and have demonstrated effectiveness [8,38]. These studies have examined the factors that affect incident duration, focusing on individual components of post-event duration, such as notification, response, and cleanup [16]. Our study proposes a model that takes a different approach to estimating traffic event durations. The model employs four machine learning methods and was tested in an experimental study conducted in Istanbul, a large metropolis with complex traffic. The study classified traffic accidents and incidents into four duration classes and estimated their duration based on which duration class the incidents were in. The model simplified the problem by reducing the number of complex features. As a result, instead of extracting data from numerous databases, a few databases were integrated, enabling real-time data for dynamic forecasting.

This paper presents a model for forecasting the duration of traffic incidents and an experimental study of the model. In this context, the prediction of the duration of traffic incidents is approached as a classification problem rather than a regression problem. The accurate forecasting of the duration of a traffic accident is very difficult, and the specific estimation of the duration is not very useful for traffic management. Instead, knowledge of the accident duration class is much more useful for traffic management. This is because interventions and resources allocated for similar accidents with similar durations of traffic incidents do not differ. The purpose of this study is to manage traffic events and provide resource management by enabling decision-makers to act strategically. This includes determining whether urgent measures need to be taken and directing traffic police based on the duration of the incident in an agile manner. Additionally, reducing the number of features can simplify the model, resulting in faster training and execution, and decreasing the risk of overfitting. This can also improve the model’s ability to generalize and make it more interpretable. It can also prevent resource waste and facilitate effective management by eliminating unnecessary features.

Four machine learning methods were selected for the experimental study: Decision Trees (DT), K-Nearest Neighbors (KNN), Random Forest (RF), and Support Vector Machine (SVM). DT is suitable for modeling simple, non-linear relationships, while KNN excels in classification with easy integration of new data points. RF is chosen for its high performance and robustness, while SVM demonstrates excellent generalization ability and adapts well to high-dimensional data. While other classification methods exist, this study focuses on these four due to their strong performance and general availability, considering factors such as problem requirements and data structure. Furthermore, the combination of databases of variables is employed for real-time prediction. The proposed approach, as illustrated in Fig. 1, allows for the immediate estimation of the duration of traffic incidents.

The proposed approach for predicting traffic events comprises four distinct stages. The initial stage involves data preprocessing, which is undertaken to refine the dataset for subsequent analysis. The creation of the data set was realized in four steps. In the first step, data was collected with information from nine different open sources or institutions. In the second step, merging was done according to time and location information. In this merging, coordinate information about the location was converted into Geohash codes with ArcGIS. This is because when the coordinates are pointwise, the merging process will be very difficult. In the study, other variables existing in the existing area were integrated thanks to the coordinates converted to a 6-cell geohash area (0.74 km² area with a cell width of 1.22 km and a cell height of 0.61 km). In this context, it is thought that data merging is rarely done. Missing data were then identified. In cases where missing data were not eliminated, data cleaning was performed, and a data set was prepared to estimate traffic event duration in certain areas. Following this, in the second stage, a non-real-time prediction of the times of the traffic events is carried out using machine learning models. These models leverage the abundance of variables inherent in the initial problem domain. Since this is a critical stage in training four different machine learning models, the hyperparameters must be done carefully. In this context, GridSearchCV was integrated into four different machine learning models and hyperparameters were adjusted. At the same time, the performance of four different machine learning methods was analysed. Subsequently, the complexity of the problem is mitigated through the application of feature reduction and hyperparameter optimization techniques, which facilitate the amalgamation of databases for real-time prediction while refining algorithmic performance. This process culminates in the real-time estimation of traffic event times within a simplified framework, enabling swift and accurate predictions by leveraging optimized algorithms. In this context, feature selection was made with the GridSearchCV algorithm based on the model that gave the best result in the previous stage. In the last stage, with the decreasing number of variables and the real-time data set, the overall prediction of the traffic event duration was dynamically calculated with four different machine learning models. In this context, the performance of the model was compared with the situation in the second stage, and the results obtained from the effective use of resources were evaluated.

This study presents a machine learning-based model for forecasting traffic event duration, integrating feature selection and dynamic modeling. Through comprehensive testing, the model reduced 75 variables to 14 significant ones, enabling predictions prior to incidents. Real-time database structuring facilitates dynamic forecasting. The performance evaluations conducted with machine learning algorithms, including DT, RF, KNN, and SVM, revealed that the RF model achieved the highest accuracy and balanced accuracy rates. The model demonstrates effectiveness in traffic incident duration prediction, particularly in complex urban environments, suggesting potential for sustainable prediction systems with reduced variables. Experimental findings underscore its significance for traffic management and planning. We utilized the SHAP technique to identify 14 features that contribute to predictions. The time period holds particular importance for emergency response. Variables such as GeoHash and temperature also significantly influence prediction accuracy across time classes, highlighting the nuanced dynamics of incident duration prediction in urban environments like Istanbul. Each method has its own strengths and weaknesses, highlighting the importance of selecting the appropriate method for a specific problem. RF’s non-parametric nature allows for versatile application across diverse datasets, avoiding rigid assumptions and identifying intricate data patterns. RF minimizes individual tree variations by aggregating predictions from multiple decision trees trained on distinct data subsets, resulting in more accurate predictions through classification voting. Both RF and KNN exhibit high precision, recall, and F1-scores. DT’s performance slightly lags due to potential overfitting with a single tree. Despite its longer runtime, RF is more robust against overlearning from unbalanced data distributions and less sensitive to hyperparameter tuning, resulting in consistently superior predictive performance compared to other models.The RF model’s ability to calculate with 99% accuracy demonstrates its usefulness in dynamic models. It can be executed quickly and efficiently on high-processing computers, making it a valuable tool for managing traffic incidents in cities with high accident rates.

Deep learning and neural networks play a crucial role in predicting the duration of traffic incidents. Various studies have proposed innovative models integrating deep learning techniques such as LSTM, Bi-LSTM, and ANN autoencoders to enhance prediction accuracy. These models leverage features such as traffic flow, incident descriptions, and sensor data [38,40]. The fusion of machine learning with traffic data has shown significant improvements over traditional regression models, achieving up to a 60% accuracy enhancement [40]. Moreover, the interpretability of models such as TabNet has enabled the identification of key factors influencing incident duration, including road type, casualties, weather conditions, and vehicle numbers [61]. These advancements in deep learning models offer valuable insights for efficient resource allocation, emergency response, and traffic management strategies. It is therefore anticipated that even more significant outcomes may be achieved with this proposed approach in future studies as the field of deep learning continues to evolve.

The generalizability of the study can be improved by extending the estimation of traffic incident duration to larger geographical areas. Conducting similar analyses in various regions and countries beyond Istanbul will enable us to gain a broader perspective on the impacts of urban features and traffic infrastructure. Examining various geographic scaling methods can offer a more detailed analysis to determine the most suitable scaling strategy for predicting the duration of traffic events. In this regard, it is crucial to evaluate the impact of geohash scaling and alternative scaling methods. Integrating dynamic factors is essential for the prediction model to better adapt to real-world conditions. With the advent of IoT solutions and smart city applications, traffic event data can now be obtained much faster and independently of human input. By incorporating hard-to-obtain data, the prediction success of the model can be increased, allowing for more effective traffic management strategies and quicker responses to potential issues.