As a critical input, high-speed railway (HSR) passenger demand is a significant guide for railway planning and operation. Accordingly, the necessity of accurately forecasting short-term demand has become a pressing problem; it has increasingly attracted attention as a research interest. In this paper, a novel hybrid model, specifically, the SSA–WPDCNN–SVR model, is proposed for short-term HSR passenger demand forecasting that explicitly considers the relevance of neighbour time data. The model consists of three procedures: (i) the decomposition of the original time series into the principal component (PC) and several detailed components (DCs) by the singular spectrum analysis (SSA) method, which is adopted as the signal processing procedure; (ii) the forecasting of PC by means of the designed convolutional neural network with the wavelet packet decomposition (WPDCNN) through the transformation of the one-dimensional time series into weekly cross-correlation matrices (i.e., image-like two-dimensional data); (iii) the forecasting of DCs by the support vector regression (SVR) method. The case studies of three typical origin–destination pairs in an HSR line are considered which demonstrates that the proposed model not only extracts the fluctuation characteristics of passenger flows, but also outperforms several other existing models with its higher short-term HSR demand forecasting accuracy.

CICTP