Multi-step forecasting of short-term traffic flow based on Intrinsic Pattern Transform

Multi-step forecasting is an essential but tricky aspect of Intelligent Transportation Systems (ITS). Existing models generally yield unreliable results as the forecasting horizon increases due to the decay of temporal dependence. This paper presents a novel module named Intrinsic Pattern Transform (IPT) to uncover the intrinsic traffic pattern and captures long-term temporal dependence. Specifically, Empirical Mode Decomposition (EMD) is adopted to separate the traffic flow into multiple Intrinsic Mode Functions (IMFs). The mean instantaneous frequencies extracted from each IMFs via Hilbert transform indicate practical implications of traffic flow composition. We replace priori-based frequency with mean instantaneous frequencies to reconstruct long-term trends using Fourier Transform. Applying IPT to raw traffic flows successfully extracts traffic patterns, such as daily and rush hour patterns, which provides a novel perspective to understand the traffic evolution trend better. We validate IPT and IPT-based models by conducting experiments on two real-world datasets. It is experimentally demonstrated that introducing IPT for the stand-alone model does not impair single-step prediction performance, and error of multi-step prediction reduce by 0.44–5.38 MAE/step. An in-depth analysis of the robust and residual distribution demonstrates that the IPT exhibits high tolerance to noise while suppressing the generation of outliers. Comparison experiments with other baseline models demonstrate that our approach has better performance and three times lower time complexity for multi-step prediction.