Evaluation of Two Modeling Methods for Generating Heavy-Truck Trips at an Intermodal Facility by Using Vessel Freight Data

Abstract

A methodology for building a truck trip generation model by use of artificial neural networks from vessel freight data has been developed and successfully applied to five Florida seaports. The backpropagation neural network (BPNN) algorithm was used in the design. Although the methodology was sound, a new model had to be developed for each of these intermodal facilities. Lead and lag variables were necessary input variables for most models to account for commodities stored on port property before export or pickup after import. Other modeling techniques were researched, and a fully recurrent neural network (FRNN) trained by the real-time recurrent learning algorithm was selected to develop a model for Port Canaveral and compare with a BPNN model. FRNN is dynamic in nature and was found to relate to the storage time of the commodities to truck trip generation. A developed Port Canaveral BPNN model was successfully validated at the 95% confidence level with collected field data. It was applied to conduct a short-term forecast of the port's truck traffic for 5 years. The average annual growth of trucks based on the estimated freight activity under the BPNN model was 5.07%. The Port Canaveral FRNN model adequately estimated the current conditions but failed to forecast truck growth. The FRNN model required more data for forecasting than backpropagation. However, when more consecutive data are available for training, FRNN may produce more accurate results.