Distant and Recent Historical Data Fusion for Improving Short- and Medium-Term Traffic Forecasting

Abstract

Traffic became a major issue in large and crowded metropolitan cities and might cause people to waste in the order of days within a year. It is notable that traffic speed estimation problems were addressed in three main horizons: short term, medium term, and long term. In this paper, we both introduce a novel network feeding strategy improving short- and medium-term traffic forecasting and define the aforementioned horizons by evaluating the prediction results up to 6 h. We combined the advantages of both distant and recent historical data by developing two different Recurrent Neural Network (RNN)-based methods, H-LSTM and H-GRU, that employ Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks. The proposed Historical Average Long Short-Term Memory (H-LSTM) model demonstrates superior performance compared to traditional methods, as it is capable of integrating both the typical long-term traffic patterns observed in a specific location and the daily fluctuations, such as accidents, unanticipated events, weather conditions, and human activities on particular days. We achieve up to 20% improvement, especially for rush hours, compared to the traditional approach, i.e., exploiting only recent historical data. H-LSTM could make predictions with an average of ±7.5 km/h error margin up to 6 h for a given location.