Missing Traffic Speed Data Imputation Using Road Segment Characteristics for Long-Term Traffic Speed Prediction

Abstract

Traffic speed estimation has become one of the challenging difficulties, especially for metropolitan cities in the last decade. On the other hand, sensor failures are still an important problem for data-intensive tasks such as traffic-oriented problems in smart cities. Thus, the absence of traffic speed data considerably decreases the performance of speed estimation algorithms. Following this drawback, we bring a new insight into missing traffic speed data imputation and propose several techniques to fill the gaps caused by traffic speed sensor failures. Our methodology exploits the traffic characteristic of similar segments and replaces the missing values with the existing values of k closest segments with similar characteristics for the corresponding timestamp. The introduced method, a variant of the k-Nearest Neighbor algorithm, achieves a great performance in producing real-like values, especially for long-term traffic speed estimation. Test results show that 3.5% error minimization is possible for predicting traffic speed from 1 to 7 days ahead. We also proved that missing value ratios below 50% could be healed with a negligible prediction error.