Enhanced-Historical Average for Long-Term Prediction

Intelligent Transportation is a promising solution that aims to maximize efficiency by utilizing the computing power provided by computers to dynamically analyze current road traffic conditions and rationally allocate traffic resources. In Intelligent Transportation, accurate and effective prediction results are reliable guarantees for subsequent policy-making and resource scheduling. However, the traffic feature prediction is complex, for it is supposed to handle the interactions among numerous nodes, the influence of nodes' historical information, and the randomness caused by accidents. For long-term traffic forecasting, most current work focuses on prediction within one week, while little research has been touched on prediction over longer horizons. Long-term forecasting suffers the problems of instability, over-learning, and lack of interpretability. Abnormal data points in the historical data can easily affect the stability of the prediction results. The deep learning models for ultra-long-time prediction problems can lead to an over-fitting phenomenon due to too few input features. Meanwhile, the blackbox model feature of deep learning also reduces the reliability of the prediction results. Therefore, this paper takes the historical average model as the base model and considers external factors, thus proposing the Enhance-Historical Average (Enhanced-HA) algorithm to improve the prediction effect without affecting the reliability and stability, obtaining significant enhancement on the SZ-Traffic dataset.