Chuanyun Fu

Accurate short-term traffic forecasting and efficient dynamic route guidance are pivotal for mitigating urban congestion. However, accurately capturing complex temporal dependencies and nonlinear variations in traffic flow remains challenging, while conventional path-planning algorithms may also suffer from local-optimum problems. To address these challenges, this study proposes a hybrid framework integrating an optimized prediction model with an enhanced ant colony optimization algorithm. First, a GA-BiGRU-ATT model is developed for traffic state prediction. By combining a bidirectional gated recurrent unit (BiGRU), a temporal attention mechanism, and genetic algorithm-based hyperparameter optimization, the model captures contextual temporal dependencies within the observed historical input window and emphasizes critical time-step features. Under the evaluated model configurations, the proposed approach achieved traffic-state classification accuracies of 90.28% and 87.50% on Segments 1 and 2, respectively. Second, based on the predicted traffic states, an Improved Ant Colony Algorithm (IACA) incorporating traffic-state feedback is proposed to mitigate slow convergence and local-optimum entrapment in conventional ACO. Within the ant-colony-based comparison, the IACA reduced the average computational time by 49.57% relative to the conventional ACA. Furthermore, under the evaluated SUMO evening-peak scenario, periodic dynamic guidance reduced the average travel time of the selected guided vehicles by 12.06% and increased their average speed by 27.07%. These results demonstrate the potential of prediction-guided dynamic rerouting under the evaluated simulation conditions.

Eye movement features of pilots are critical for aircraft landing, especially in low-visibility and windy conditions. This study conducts simulated flight experiments concerning aircraft approach and landing under three low-visibility and windy conditions, including no-wind, crosswind, and tailwind. This research collects 30 participants’ eye movement data after descending from the instrument approach to the visual approach and measures the landing position deviation. Then, a random forest method is used to rank eye movement features and sequentially construct feature sets by feature importance. Two machine learning models (SVR and RF) and four deep learning models (GRU, LSTM, CNN-GRU, and CNN-LSTM) are trained with these feature sets to predict the landing position deviation. The results show that the cumulative fixation duration on the heading indicator, altimeter, air-speed indicator, and external scenery is vital for landing position deviation under no-wind conditions. The attention allocation required by approaches under crosswind and tailwind conditions is more complex. According to the MAE metric, CNN-LSTM has the best prediction performance and stability under no-wind conditions, while CNN-GRU is better for crosswind and tailwind cases. RF also performs well as per the RMSE metric, as it is suitable for predicting landing position errors of outliers.

Per the World Health Organization, traffic accidents cause significant financial losses and fatalities annually