Short-Term Traffic Flow Forecasting by Selecting Appropriate Predictions Based on Pattern Matching

Abstract

Forecasting short-term traffic flow is one critical component in traffic management to improve operational efficiency. Data driven method, which trains the predictor with historical data across a given past period, have been proved to perform well. However, days which experience significantly different traffic flow patterns, negatively influence forecasting results. This paper proposes an advanced method, making use of appropriate prediction based on pattern matching. First, historical data is divided into several groups, according to their patterns, by clustering algorithms. Then the predictor is trained for each group based on a convolutional neural networks and long-short-term-memory model. For each time point, the degree of similarity between the target day and each group is measured, and the predictor trained by the group possessing the highest degree of similarity is selected to be appropriate. Based on a case study from Seattle, we show that selecting an appropriate predictor can significantly improve the accuracy of predictions. In addition, we demonstrate that the new method can, in general, outperform alternative methods in terms of prediction accuracy and stability.