Design of Dynamic Neural Networks to Forecast Short-term Railway Passenger Demand

This paper develops two dynamic neural network structures to forecast short-term railway passenger demand. The first neural network structure follows the idea of autoregressive model in time series forecasting and forms a nonlinear autoregressive model. In addition, two experiments are tested to eliminate redundant inputs and training samples. The second neural network structure extends the first model and integrates internal recurrent to pursue a parsimonious structure. The result of the first model shows the proposed nonlinear autoregressive model can attain promising performance and most cases are fewer than 20% of Mean Absolute Percentage Error. The result of the second model shows the proposed internal recurrent neural network can perform as well as the first model does and keep the model parsimonious. Short-term forecasting is essential for short-term operational planning, such as seat allocation. The proposed network structures can be applied to solve this issue with promising performance and parsimonious structures.