A neural network approach to modeling and predicting intercity passenger flows

The predictive ability and accuracy of traditional methods of flow modeling are often limited due to methodological assumptions or data requirements. Neural network (NN) models do not require any methodological or distributional assumptions and have been used recently as alternatives to conventional interaction models in transportation research. The predictive ability of NNs with large databases and the temporal stability of these models have not been fully assessed. This study evaluates the predictive ability of NNs with a database of Amtrak passenger flows with special characteristics, i.e., a very large database with many zero values; it also evaluates the contribution of various combinations of NN inputs and the temporal stability of the NN models derived.;A NN is a dynamic computing system consisting of processing units that interact with each other based on a structure similar to the human nervous system that learns from experience. The most commonly used neural network model is the back-propagation method. A feed-forward back-propagation type NN was successfully trained and was applied to predict Amtrak passenger flows between 97 stations. The study also discusses the training techniques used in handling the data set with many zero values.;The study found that the NN outperformed regression methods and predicted as well as a fully-constrained gravity model by producing a smaller root mean squared error (RMSE), extracting essential patterns in the data, while predicting zero and large flows well. A flow assignment using the NN predicted flows yielded satisfactory results in terms of R2 (coefficient of determination), the standard error of estimate, the flow patterns, the average link flows in the system, and the total passenger miles traveled (PMT).;The study found that the magnitude of the NN training and trained parameters positively correlate with the importance and contributions of the input variables to the NN. The accuracy of the NN model was found to be stable with temporal data. The model predicted flows followed the seasonal pattern of the data and the RMSEs were improved by 38% to 51% over the RMSEs from the regression model. The findings are useful to transportation modelers, researchers and practitioners.