Deep Learning for Short-Term Traffic Flow Prediction in Smart Cities

Abstract

Accurate short-term prediction of traffic flow is a cornerstone of intelligenttransportation systems in smart cities, supporting signal control, routeguidance, and congestion management. Traffic flow is nonlinear and depends on complex temporal patterns, which classical statistical models capture only partially. This paper applies deep learning, in the form of a long short-term memory network, to short-term traffic flow prediction using historical detector data. The network learns temporal dependencies from past flow observations to forecast flow over short horizons, and it was compared with historical-average, autoregressive, and support-vector-regression baselines. The deep learning model produced the most accurate forecasts, reducing the root-mean-square error to 21.4 and the mean absolute percentage error to 9.3 percent, clearly outperforming the baselines across the prediction horizons considered. The predicted flow tracked the observed diurnal pattern closely, including the morning and evening peaks. The results demonstrate that recurrent deep learning is an effective tool for short-term traffic prediction and a valuable component of smart-city traffic management. KEYWORDS: traffic prediction, deep learning, long short-term memory,smart cities, intelligent transportation, time-series forecasting