This study utilizes the characteristics of railway reservation data and proposes a novel model based on the concept of curve similarity. The proposed model considers mainly temporal features hidden in the reservation data and establishes four modules. The similarity evaluation module is responsible for identifying similar booking curves in the historical database; the sample selection module decides which and how many samples should be selected for computing predictions; the prediction module integrates the essential information of the selected samples and generates forecasts; and the learning module searches for parameters applied throughout the whole forecasting procedure. The established model is compared with three benchmark models to verify the model validity. Empirical results show that, on average, the proposed similarity-based forecasting model can improve at least 9% of mean square errors (MSEs) over the benchmark models.
2016 Eastern Asia Society for Transportation Studies