The passenger demand of urban rail transit (URT) lines often present asymmetric tidal time-varying characteristics. To match the demand fluctuation, the train operational plan (TOP) generally has asymmetric bi-directional frequency/headway setting and imbalanced circulation, leading to high operation cost. This paper incorporates circulation balance into TOP optimization to balance the bi-directional arrival, departure, circulation, and resource utilization, and reduce the overall operation cost. Based on time-varying section demand and predetermined service level, bi-directional stepped maximum headway functions are collaboratively constructed, and then the circulation process is described by the trip flow circulation network that is formulated as a cost-oriented integer linear programming model. Using the optimized frequency setting, the final TOP is obtained by a two-stage approach to successively solve the schedule and rolling stock circulations at terminals. The case study based on an URT line in Shenzhen indicates that the proposed approach can not only ensure the required service level for travel demand, but also improve the efficiency of circulation and utilization, and effectively reduce the overall operation cost. The proposed approach provides an effective technique to keep balanced, stable and sustainable operation for URT lines.

Recommendation systems play a pivotal role in improving product competitiveness. Traditional recommendation models predominantly use centralized feature processing to operate, leading to issues such as excessive resource consumption and low real-time recommendation concurrency. This paper introduces a recommendation model founded on deep learning, incorporating edge computing and knowledge distillation to address these challenges. Recognizing the intricate relationship between the accuracy of deep learning algorithms and their complexity, our model employs knowledge distillation to compress deep learning. Teacher–student models were initially chosen and constructed in the cloud, focusing on developing structurally complex teacher models that incorporate passenger and production characteristics. The knowledge acquired from these models was then transferred to a student model, characterized by weaker learning capabilities and a simpler structure, facilitating the compression and acceleration of an intelligent ranking model. Following this, the student model underwent segmentation, and certain computational tasks were shifted to end devices, aligning with edge computing principles. This collaborative approach between the cloud and end devices enabled the realization of an intelligent ranking for product listings. Finally, a random selection of the passengers’ travel records from the last five years was taken to test the accuracy and performance of the proposed model, as well as to validate the intelligent ranking of the remaining tickets. The results indicate that, on the one hand, an intelligent recommendation system based on knowledge distillation and edge computing successfully achieved the concurrency and timeliness of the existing remaining ticket queries. Simultaneously, it guaranteed a certain level of accuracy, and reduced computing resource and traffic load on the cloud, showcasing its potential applicability in highly concurrent recommendation service scenarios.

In order to improve the prediction accuracy of railway passenger traffic, an ARIMA model and FSVR are combined to propose a hybrid prediction method. The ARIMA prediction model is established based on the known railway passenger traffic data, and then, the ARIMA prediction results are used as the training set of the FSVR method. At the same time, the air price and historical passenger traffic data are introduced to predict the future passenger traffic, to realize the mixed prediction of railway passenger traffic. The case study demonstrates that the hybrid prediction method can effectively improve the prediction performance of railway passenger traffic. Compared with the single ARIMA method, the hybrid prediction method improves the delay of the prediction results. Compared with the FSVR prediction result, the hybrid prediction method greatly reduces the errors in the extreme points of passenger traffic and long-term prediction. The relevant research results of this paper provide a useful reference for the prediction of railway passenger traffic.

Forecasting the passenger flow of high speed railway is a hot but difficult research topic in the field of railway transportation for a long time. It is the basis of the revenue management and the rapid response mechanism for the railway sector. With the increasingly fierce competition of aviation and railway, it brings forward new challenges to the forecasting. This paper proposes a new method of passenger flow forecasting of the high speed railway in competitive market based on XGBoost model. The ticket price of the train is also taken into account in the method. First, the flight booking info data is preprocessed and statistics features are calculated to complement the missing data. Then, the railway booking info and flight booking info merge into a large training data and prediction model is constructed by using XGBoost algorithm. Applying this model for prediction, the passenger flow in competitive market can be obtained. Experimental results show that the proposed model has higher prediction accuracy than traditional models.

Federated continual learning (FCL) must preserve useful historical knowledge while learning from evolving and statistically heterogeneous client streams. However, indiscriminate replay can retain client-specific bias, noise, or task conflicts and thereby cause negative transfer. In this paper, we propose Adaptive Multi-Factor Accurate Forgetting for Heterogeneous Federated Continual Learning (AMAF-FCL), a selective memory-management framework that jointly assesses replay reliability and adapts the influence of generated historical features. AMAF-FCL achieves this goal by (1) modeling historical knowledge in feature space with a conditional real-valued non-volume-preserving (RealNVP) normalizing flow; (2) combining class-conditional likelihood, predictive uncertainty, and global distribution consistency; and (3) adjusting replay weights according to client–global heterogeneity. The local objective combines a classification loss for learning the current task, a reliability-weighted replay loss for retaining useful historical knowledge, and a feature-distillation loss for limiting drift in the feature representation. In the EMNIST long-task-pool (EMNIST-LTP) benchmark, each client learns six two-class tasks drawn from a client-specific set of handwritten letters. AMAF-FCL achieves 51.2% average accuracy and 7.6% average forgetting, improving over the likelihood-based AF-FCL baseline by 3.7 and 1.5 percentage points, respectively. In the cross-domain digit-and-fashion setting, it reports 71.2% average accuracy and 6.8% average forgetting; with four noisy clients, it obtains 56.1% average accuracy and 9.8% average forgetting. These results indicate that multi-factor reliability assessment and heterogeneity-aware adaptive forgetting improve the balance between useful knowledge retention and harmful-knowledge suppression in heterogeneous FCL.