DeCyReM: De-Cycled Residual Memory with Utility-Gated Retrieval for Traffic Flow Forecasting

Accurate traffic-flow forecasting remains challenged by abrupt and irregular states even after dominant periodic patterns are captured. Existing predictors model the resulting difficult errors implicitly through their parameters and cannot explicitly reuse specific historical errors at inference. We find that multi-horizon base-prediction errors recur conditionally across similar de-cycled node-local states. Based on this observation, we propose De-Cycled Residual Memory with Utility-Gated Retrieval (DeCyReM). It constructs queries from de-cycled residuals, spatial context, and temporal information, retrieves historical base errors from a node-local difficult-and-diverse memory, and uses a utility gate to control their contribution as forecast corrections. Across PeMS03, PeMS04, PeMS07, and PeMS08, DeCyReM records the numerically lowest reported value in seven of the twelve dataset–metric combinations and ranks within the top three in eleven. Ablation and mechanism analyses show that retrieval improves error alignment and mainly benefits difficult samples; similarity alone cannot determine correction utility, whereas gate values track helpful retrievals. Case and robustness analyses further characterize successful correction, safe attenuation, and mismatch-induced failure, identifying reliable key–value correspondence as essential. Overall, DeCyReM converts reusable historical error experience into selective forecast corrections, providing an error-centric approach to forecasting abrupt and irregular traffic states.