| This dissertation has established a data-driven prediction model for traffic data and explored related data processing techniques for its adaptive and real-time application. This model can approximate the nonlinear, non-stationary, time-dependent and spatiotemporal-correlated multivariate relationship for traffic prediction in real time and adaptively, built up from continuously incoming real-world traffic data.; First, a nonparametric model using the multivariate local linear regression is employed to conduct real-time traffic prediction, which can provide higher accuracy than existing nonparametric models and also offer a closed-form expression for prediction intervals under certain assumptions. A nonparametric procedure using bootstrap to obtain prediction intervals is also proposed. Then, the adaptive algorithm for parametric models is extended to this nonparametric model and the feedback available in short-term forecasting can be used. Next, the joint spatiotemporal prediction is studied to utilize the spatial correlation. In addition, the wavelet preprocessing is implemented to improve prediction further by virtue of multiscale analysis. Several case study results are presented and discussed for all models. Finally, concluding remarks and suggestions for future research are presented. |
