| dc.description.abstract | Abstract
Passenger volume prediction in urban transportation management is crucial, particularly origin-destination (OD) prediction, which helps understand passenger movement patterns and efficiently schedule train operations and vehicle allocation. Traditional methods often fail to handle holidays, special events, or unexpected incidents, resulting in low prediction accuracy. To address this, we propose a neural network prediction model that combines activity labels with historical time-series OD data, improving prediction accuracy during special events.Our model, constructed using the nn.Module in PyTorch, takes inputs such as time periods, entry and exit stations, passenger counts, workdays or holidays, anomaly event markers, activity types, and travel times between OD points. The output is the exit proportion at each station. Given that the OD proportion is a probability distribution, we use the cross-entropy loss function to enhance the model′s accuracy in predicting OD distributions for metro stations. For the learning rate algorithm, we switch from Adam to SGD and manually adjust the learning rate to avoid extreme values in the model output, significantly improving the model′s adaptability to high-dimensional data. Additionally, we propose predictive scheduling and enhanced service recommendations for the Taipei Metro, including preemptively adjusting train intervals, real-time updates of passenger flow information, and developing event cooperation strategies to boost overall operational efficiency and passenger satisfaction. Ultimately, we use ticket OD data to calculate inter-station volume. | en_US |