Short-term traffic flow forecasting considering operational characteristic differences

To address accuracy degradation in short-term traffic flow forecasting caused by spatiotemporal differences in traffic operational characteristics, this study proposes a framework integrating DTW-enhanced K-means clustering and a Feature-Augmented LSTM (FA-LSTM). Traditional models have two flaws: they use Euclidean distance for clustering (failing to capture traffic time-series shape similarity) and standard LSTM ignores irregular inter-arrival times/abrupt feature variations. To solve these, we first replace K-means’ Euclidean distance with DTW to cluster detectors with homogeneous temporal profiles. Then, we design FA-LSTM, embedding inter-arrival times and feature differences into its input and gating mechanism to enhance sensitivity to traffic dynamics. Experiments on PEMS04 data show the cluster-wise strategy reduces average RMSE, MAE, and MAPE by 1.03, 0.74, and 3.3% vs. a global model. It also improves peak-period (morning/evening rush) fitting, providing reliable support for urban traffic management.