An Artificial Neural Networks Approach to Forecast Short-term Railway Passenger Demand

This paper experiences a three-phrase back-propagation neural network approach to forecast short-term railway passenger demand. The first phase involves the selection of variables, the size of training data set, and the modification of stochastic outliers, under a specific origin-destination (O/D) pair of a given train service. In the second phase, in order to verify the robustness of developed approach, we construct two aggregated models, in which each model applies different temporal aggregations of demand. In the third phase, we construct three integrated models by considering multiple train services simultaneously to enhance the future application. The approach shows encouraging results and most forecasting performances are under 20% of mean absolute percentage error (MAPE). In addition, the approach is able to forecast railway passenger demand effectively under various scenarios of train services. The outcomes of the models can offer detailed demand prediction for railway operation planning, such as train scheduling and seat allocations.