Acceleration-enriched input preprocessing for traffic speed forecasting: a dual-channel framework

Most spatiotemporal traffic forecasting models rely exclusively on speed time series with unified normalization, ignoring the kinematic relationship between speed and acceleration. This paper presents an empirical study of acceleration-based preprocessing, examining whether explicitly derived and filtered acceleration signals—used as auxiliary input features—improve speed forecast accuracy. We evaluate causal Savitzky–Golay (SG) filtering applied to finite-difference acceleration derived from 5-minute loop-detector data, combined with dual-channel normalization that applies separate statistics to speed and acceleration inputs. The preprocessing pipeline operates entirely at the input level, requiring no architectural modifications. We benchmark the approach on five representative spatiotemporal architectures—DCRNN, AGCRN, STGIN, Graph WaveNet, and the transformer-based STAEformer—across the METR-LA and PEMS-BAY benchmarks. The proposed acceleration preprocessing consistently improves accuracy for all five architectures, with MAE reductions of roughly 8–38% over speed-only baselines—largest at short horizons and smaller but consistent at the 60-minute horizon. The benefit of SG filtering is horizon- and dataset-dependent: it is most valuable at longer horizons and on the smoother PEMS-BAY network, whereas raw acceleration can suffice at short horizons on METR-LA; the transformer-based model benefits from acceleration at all horizons without requiring SG filtering. The framework is architecture-agnostic and composable with existing forecasting systems.

Actions (login required)

View Item