Short time traffic speed prediction using data from a number of different sensor locations

In this study we predict traffic speed on Istanbul roads using RTMS (remote traffic microwave sensor) speed measurements obtained from the Istanbul Municipality Web site from 327 different sensor locations. We do speed predictions 5 minutes to an hour ahead and use SVM (support vector machine) and kNN (k nearest neighbor) methods for speed prediction. First of all, for speed prediction at a certain sensor location, we compute the most important past speed measurements for better accuracy using feature selection methods. We also find out which other sensors could be used to predict the speed at a certain sensor location and show that especially for nearby/correlated sensors, it is possible to get better results using related sensor measurements in addition to the sensor being predicted. We also show that only using the correlated sensors, it is possible to get good accuracy. This result could be very useful when a sensor breaks down or needs to be calibrated. In all our experiments, we find out that SVM produces better results than kNN.