Comparative Evaluation Of Three Methods For Predicting Traffic Volume
Zaman, Seyed Ali (2008) Comparative Evaluation Of Three Methods For Predicting Traffic Volume. Masters thesis, Universiti Putra Malaysia.
In many places the capacity of existing road traffic system is frequently exceeded by the traffic demand. Combinations of technologies and systems that are generally called as Intelligent Transportation Systems (ITS) have the potential to perform as an influential tool to battle against congestions by increasing the effectiveness of the present surface transportation network. One of the most important issues regarding the utilization of above system is the need to forecast the traffic volume. This research presents forecasting of short-term traffic volume utilizing Artificial Neural Networks (an intelligent advanced method), ARIMA (Auto Regressive Integrated Moving Average) time series method and Historical Average along the PLUS highway in Malaysia. The study focuses on two stations at Sungai Besi and Nilai along Section 5 of the highway. Feedforward ANNs, ARIMA timeseries, and Historical Average methods are developed for these sections for single and multiple intervals in order to forecast traffic volume and compare the results. The data for this study consist of a three months period of 2006 obtained from PLUS highway authority. Twelve various NNs models are developed including Univariate and Multivariate models with a wide range of inputs. This is done to find the most effective NNs model with the highest performance in terms of traffic volume forecasting. Models were developed for all week days as well as single day’s model. Inputs of these models are mostly previous hours’ traffic volume, upstream flow, and weather information. Four time series models and one historical average model were developed for forecasting traffic volume. Time series models are developed for weekdays and holidays separately. This study proved that the architecture of ANNs model is suitable to be applied to the traffic volume forecasting problem. It also demonstrates that a successful neural network model requires considerable effort in defining the network’s parameters. Generally NNs with previous hours’ traffic volume, same hour traffic volume of same day of last weeks, and abnormal day distinguisher as input are more successful than others. The study revealed that NNs model brings the best results and consequently has the highest performance for forecasting short-term traffic volumes. NNs method also shows an acceptable level of accuracy for the case of multiple forecasting which had a low level of error raise. It can be concluded that NNs models are site specific and they perform better in sites with high level of traffic variation due to their adaptive nature. ! " # $% ! ! " !
Repository Staff Only: Edit item detail