This paper presents COWINE (Community Twin Ecosystem), an ecosystem that harnesses Digital Twin (DT) to elevate and transform community resilience strategies. COWINE aims to enhance the disaster resilience of communities by fostering collaborative participation in the use of its DT among the decision-makers, the general public, and other involved stakeholders. COWINE leverages Cities:Skylines as its base simulation engine integrated with real-world data for community DT development. It is capable of capturing the dynamic, intricate, and interconnected structures of communities to provide actionable insights into disaster resilience planning. Through demonstrative, simulation-based case studies on Brevard County, Florida, the paper illustrates COWINE’s collaborative use with the involved parties in managing tornado scenarios. This study demonstrates how COWINE supports the identification of vulnerable areas, the execution of adaptive strategies, and the efficient allocation of resources before, during, and after a disaster. This paper further explores potential research directions using COWINE. The findings show COWINE’s potential to be utilized as a collaborative tool for community disaster resilience management.
Proactive safety systems aim to mitigate risks by anticipating future potential conflicts between vehicles and enabling early intervention to prevent work-zone-related crashes. This study presents an infrastructure-enabled proactive work zone safety warning system that leverages a digital twin (DT) environment, integrating real-time multisensor data, detailed high-definition (HD) maps, and a trajectory prediction model based on a historical-prediction attention mechanism (AM). Specifically, the contributions of this study are in twofolds: 1) developing an infrastructure-enabled vehicle trajectory prediction framework that leverages the historical prediction network (HPNet) model integrated with Lanelet2 HD maps and 2) developing a proactive safety warning generation application that incorporates probabilistic conflict modeling and vehicle bounding-box representations. The development and effectiveness demonstration of the presented trajectory prediction model and proactive warning system is conducted using a cosimulation environment integrating simulation of urban mobility (SUMO) and CAR learning to act (CARLA) simulators. To evaluate the accuracy of predicted trajectories (PTs), we use two standard metrics: joint average displacement error (ADE) and joint final displacement error (FDE). Our analyses revealed that the infrastructure-enabled HPNet model demonstrates superior performance on the work zone datasets generated from the cosimulation environment, achieving a minimum joint FDE of 0.3228 m and a minimum joint ADE of 0.1327 m, lower than the benchmarks on the Argoverse (minimumJoint FDE: 1.0986 m, minimumJoint ADE: 0.7612 m) and INTERACTION (minimumJoint FDE: 0.8231 m, minimumJoint ADE: 0.2548 m) datasets. In addition, our proactive safety warning generation application demonstrates its ability to issue alerts for potential vehicle conflicts.
Accurate traffic prediction during hurricane evacuations is essential for effective emergency management, yet conventional approaches may not fully capture the behavioral dynamics underlying evacuation traffic. This study develops a behavior-aware framework that combines an adversarial behavior-learning component inspired by Adversarial Inverse Reinforcement Learning (AIRL) with deep learning models. Using probe vehicle trajectories, the behavior-learning component identifies route-choice patterns and produces two model-internal behavioral features: route-choice entropy and maximum transition probability. These features are incorporated into LSTM models for short-term speed prediction and MLP models for long-term congestion classification, with strict temporal separation between predictors and targets. The framework is evaluated using probe vehicle data from Hurricane Sally in Escambia County, Florida, an evacuation event shaped by pandemic-era non-congregate sheltering guidance. AIRL-inspired features improve short-term speed prediction across all four forecast horizons, reducing MAE by 0.5–0.85 kph, and improve long-term congestion classification, increasing macro-F1 by approximately 0.05. Although baseline accuracy is high (96–98\%), macro-F1 provides a more informative measure under class imbalance. Improvements remain statistically significant after Holm–Bonferroni correction, with effect sizes considered alongside p-values given serial correlation in the observations. Transfer learning from normal traffic provides, at most, a marginal and non-significant benefit at 15 minutes and no consistent improvement at longer horizons. Overall, the results demonstrate the value of incorporating learned route-choice information into evacuation traffic prediction while motivating further causal and survey-based research to establish the behavioral meaning of these learned features.