Predicting Real-time Crash Risks during Hurricane Evacuation Using Connected Vehicle Data

The study addresses the challenge of increased crash risk during hurricane evacuations, where sudden surges in traffic volume produce oscillatory speeds and higher rear-end crash potential. Traditional crash mitigation strategies based solely on historical crash data are insufficient for proactive management during evacuations. The research question is whether real-time connected vehicle (CV) data, which provide micro-level speed and acceleration information, can be used to predict crash risk on highways during hurricane evacuations. The purpose is to develop and validate a framework for real-time crash risk prediction using CV data, enabling traffic agencies to identify high-risk segments and deploy timely countermeasures. The importance lies in overcoming limitations of infrastructure-based sensors (coverage, cost, maintenance) by leveraging decentralized CV data to enhance safety decision-making during short, intense evacuation periods.

Prior work links naturalistic driving behavior (speed, acceleration, hard braking) to crash risk, but faces limitations like privacy and low participation rates. Infrastructure-based detectors (e.g., MVDS, loop detectors) have been widely used for real-time crash prediction, identifying key features such as average and downstream speeds and variability in occupancy and volume. However, these sensors have limited spatial coverage and are susceptible to failure and maintenance issues. Machine learning methods (SVM, Random Forest, XGBoost, CNN/LSTM) have shown strong performance for real-time crash risk prediction, though deep models require large datasets. Evacuation traffic differs from normal conditions, and studies (e.g., during Hurricane Irma) found elevated crash risk during evacuations. Emerging CV data can capture micro-level behaviors (speed/acceleration distributions, hard-braking) useful for surrogate safety measures and potentially improve crash risk prediction. This study contributes by leveraging CV data alone, using shallow ML models suitable for limited evacuation datasets, and applying GPBoost (boosting plus Gaussian processes/mixed effects) to capture spatial dependence.

Study area and data: The study focuses on 291 miles of I-10 in Louisiana during the three-day evacuation period surrounding Hurricane Ida (August 27–29, 2021). The corridor was divided into 124 bidirectional segments (≈2.25 miles each; SD ≈ 0.91 miles). Three datasets were used: connected vehicle data (Otonomo) with location, speed, and acceleration sampled at <30-second frequency; crash data (RITIS) aggregated to segment-time bins; and hourly weather data from the nearest airport stations. Data processing: CV raw data were cleaned for unreasonable values. Speed outliers were clipped using interquartile-based upper bound (≈245.4 km/h) with lower bound at 0. Acceleration outliers were removed using ±13 m/s² thresholds (based on max acceleration of Tesla Model S Plaid). Data were spatially mapped to segments and temporally aggregated into 5-minute intervals, consistent with detector-based literature practices. For each segment-interval, core traffic features were computed (e.g., mean speed, max speed, standard deviation of speed, vehicle count, calculated mean acceleration/deceleration from speed trajectories, counts above thresholds). Thresholds: speed threshold 160.9 km/h (100 mph) to capture super-speeding (>30 mph over a 70 mph limit); acceleration/deceleration threshold ±3.4 m/s² (AASHTO stopping sight distance deceleration). Additional features included upstream/downstream segment features and lagged features for the previous 5, 10, and 15 minutes. Hourly weather features were merged by nearest station. Crash labeling flagged segment-intervals with crash occurrence as 1, otherwise 0. Feature engineering and selection: A correlation analysis guided removal of redundant features; coefficient of variation of speed was dropped. Using Random Forest permutation importance and SHAP from XGBoost, features negatively impacting performance (e.g., acc_cnt_threshold, dcc_cnt_threshold, temperature, CVS and their spatial/temporal counterparts) were removed. The final model used 54 features. Class imbalance handling: After processing, the dataset had 180 crash and 88,060 non-crash instances (highly imbalanced). SMOTE oversampling was used for balancing with ratios of 1:1, 1:2, and 1:4 (crash:non-crash). To retain crash cases across splits, crash and non-crash subsets were split separately (70/30 train/test), then recombined; SMOTE was applied to both train and test sets to achieve desired ratios. Models: Binary classification models evaluated included Logistic Regression (sklearn, regularized), linear SVM (sklearn LinearSVC), Random Forest (sklearn), GPBoost (Gaussian Process Boosting integrating boosting with Gaussian process/mixed effects to capture spatial random effects), and XGBoost (learning rate 0.05, max depth 11 from grid search). Loss functions and default parameters were used except for noted hyperparameters and max iterations. Models were run 10 times per sampling strategy, and average metrics reported (Accuracy, Precision, Recall/Sensitivity, F1-score, Specificity).

• Data summary: 31,186 unique vehicles; >5 million CV data points; 376 crashes recorded over I-10 during the evacuation; after aggregation and filtering, 180 crash and 88,060 non-crash segment-intervals used for modeling.
• Feature importance: Upstream vehicle count (proxy for volume/flow) was the most influential feature. Other important predictors included downstream mean speed, segment max speed, count of super-speeding observations (speed_cnt_threshold), upstream mean deceleration, downstream standard deviation of speed, time of day, and upstream time-difference between segments.
• Model performance (average over 10 runs):
  • XGBoost achieved the highest and most stable recall across sampling ratios: Recall = 0.87 (1:4), 0.88 (1:2), 0.91 (1:1); F1 = 0.93, 0.94, 0.95; Precision ≈ 0.99 for all ratios; Specificity ≈ 0.99.
  • GPBoost performance improved with higher crash sampling ratio: Recall = 0.58 (1:4), 0.77 (1:2), 0.91 (1:1); F1 = 0.73, 0.84, 0.89; Precision = 0.99, 0.94, 0.87; Specificity = 1.00, 0.98, (not reported for 1:1 but high).
  • Random Forest: Recall = 0.54, 0.73, 0.86; F1 = 0.69, 0.82, 0.90; Precision = 0.95, 0.93, 0.94; Specificity = 0.99, 0.97, (not reported for 1:1).
  • Logistic Regression and SVM had lower recall: 0.21/0.18 (1:4), 0.44/0.44 (1:2), 0.66/0.66 (1:1).
• Best-performing models: XGBoost and GPBoost both reached Recall = 0.91 at SMOTE 1:1; XGBoost also delivered high F1 (0.95) and maintained strong recall across all sampling ratios.
• Comparative context: The achieved recall (0.91) exceeds most prior real-time crash prediction studies summarized (e.g., CNN best recall ≈0.888 in Cai et al., 2020), despite using only three days of evacuation data and shallow ML models.

The findings demonstrate that connected vehicle data can effectively predict real-time crash risk during hurricane evacuations, addressing the research objective of enabling proactive, segment-level safety assessment without reliance on fixed infrastructure sensors. By leveraging micro-level speed and acceleration behaviors, the models capture dynamic traffic instability (e.g., speed variance, hard deceleration, super-speeding) known to precede crashes. High recall (up to 0.91) indicates strong capability to identify imminent crash-prone conditions, supporting targeted deployment of countermeasures (e.g., dynamic messaging, patrols, ramp metering). The importance of upstream vehicle count and downstream speed/variability aligns with known mechanisms of rear-end crash risk under congested, oscillatory flows typical of evacuations. GPBoost’s strong performance at higher sampling ratios suggests benefits of modeling spatial random effects across contiguous segments, while XGBoost’s robustness across sampling strategies underscores its practicality for real-time implementation. Overall, the results validate CV data as a scalable alternative to infrastructure sensors for wide-area safety monitoring, particularly valuable during short, high-impact evacuation windows.

This study presents a framework to process connected vehicle data for real-time crash risk prediction during hurricane evacuations. Using three days of CV data on I-10 in Louisiana during Hurricane Ida, along with weather and crash records, multiple machine learning models were trained on 5-minute, segment-level features. XGBoost and GPBoost achieved the best performance, with recall up to 0.91 (SMOTE 1:1); XGBoost attained F1-score of 0.95. Feature importance analyses identified upstream vehicle count, downstream mean speed, speed variability, and super-speeding prevalence as key predictors. The results indicate that CV data can provide broad coverage and capture critical micro-level driving behaviors, enabling proactive, real-time safety management during evacuations. Future work should expand data duration and spatial coverage, incorporate more precise weather observations, develop methods less reliant on synthetic oversampling, and further exploit spatial-temporal dependencies (e.g., GPBoost and graph-based models) to enhance generalizability and real-time deployment.

• Data sparsity and rarity of crashes: Only three days of evacuation data were available; final dataset included 180 crash cases versus 88,060 non-crash cases, necessitating oversampling.
• Oversampling dependence: Use of SMOTE, including on test sets, may inflate performance metrics (e.g., high precision/specificity) and limit generalizability to truly imbalanced real-world data.
• Weather data granularity: Hourly weather observations may not align precisely with 5-minute traffic intervals, introducing potential inaccuracies.
• Feature reliability: Some acceleration readings can be noisy; calculated acceleration/deceleration from speed trajectories was used to mitigate sensor artifacts, but residual errors may remain.
• Spatial and temporal coverage: Results are specific to a single interstate corridor during one evacuation event; transferability to other regions/events needs validation.