Traffic Hyperparameters for Long-Term Traffic Forecasting

Traffic forecasting researches promise lower error rates for longer periods. However, most studies in the literature are generally focused on short term traffic speed estimation. In this study, the parameters that affect the long-term traffic prediction are examined. Three different methods including well-known short-term predictor ARIMA, proposed Mean Filtering Estimation and CNN methods, were exploited within the scope of this study. First, the parameters of these methods, which may affect the traffic speed, are introduced. Then, we examine those parameters including window size, number of week and number of hop for the aforementioned methods to improve the success of long-term prediction. Moreover, a graph-based approach was adopted into these methods. This study reveals that long-term predictions up to seven days could be achieved with an error rate of +/- 9 km/h for 208 different locations in Istanbul.