Traffic speed prediction with neural networks

Official URL: http://risc01.sabanciuniv.edu/record=b1655696 (Table of Contents)

With the increasing interest in creating Smart Cities, traffic speed and flow prediction have attracted more attention in contemporary transportation research. Neural networks have been utilized in many recent studies to tackle this problem; yet, these methods have focused on the short-term traffic prediction while longer forecast horizons are needed for more reliable mobility and route planning. This work aims at filling this gap by trying to address the mid-term forecasting as well as the short-term. The study employs feedforward neural networks that combine different time series forecasting techniques such as na¨ıve, moving average and exponential smoothing where the predicted speed values are fed into the network as inputs. We train our neural networks and select the hyperparameters of the network structures to minimize the error; thus, yielding the best possible setup for further forecast input. In our experimental study, we analyzed two nearly 20-km multi-segment routes from the city of Istanbul in Turkey. The speed data on these routes are collected by GPS for every minute for a 5-month horizon. Our computational tests showed that forecasts are more successful when performed on a route with shorter segments as well as a combination of conventional predictive methods are input to a neural network. We also discovered that depending on the characteristic of the analyzed road, it is possible to utilize the information from neighboring segments.