| Estimation of traffic volume plays a key role in reducing traffic congestion, enhancing control of transportation facilities and improving traffic safety. Traffic volume predictions can be applied to operational investigations of intersection, freeway, tunnel, toll booth and even parking lots. From the operational time frame consideration, forecasting techniques may be divided into the categories of long term, intermediate term and short term. In a single-series model, traffic volume forecasting is based on past values and is not based on any other data series. In a causal model, traffic volume forecasting is involved in other data series which are considered as "drive forces" of traffic volume variation.; This thesis discusses applications of decomposition models, Winters' models, ARIMA (Box-Jenkins) Models and neural network models to traffic volume forecasting problems. As a working example, the hourly traffic volume data collected on IR 271 and IR 90 in Cuyahoga county (Cleveland) from June 4 through 9, 1997 are used to build forecasting models and to check the accuracy of models. Decomposition models, Winters' models, ARIMA models and neural network models are examined for traffic volume forecasting in this thesis. For selected forecasting models and their corresponding mean absolute percentage errors, the radial basis function neural network model performs better than any other model. In structural types of models, multiplicative models have better forecasting performance than additive models. The latter shows even negative forecasting traffic volumes which are significantly different from actual observations. Although it does not have negative forecasting values for ARIMA (2,0,0) model, high mean absolute percentage error (MAPE) still indicates its poor prediction performance. The predictions generated by a neural network model are not only the best fit to actual time series, but also catch every turning point of the time series. Invoked by a searching procedure for extreme values, the 0.618 method is used to determine the optimal number of hidden neurons for a radial basis function neural network. Applying the 0.618 method to the working example obtains a minimum MAPE value of 0.2168 (the best accuracy). This study indicates that increasing number of inputs does not improve forecasting accuracy of test data, although it does enhance MAPE for training data. Also, in terms of forecasting accuracy in the example, the multivariate time series model does not perform better than univariate time series models. Finally, the strategy of updating forecasting models in real time operation is discussed and four rules are given for different situations. |
