Predictive Memory and Horizon-Conditioned Fusion for Multi-Step Traffic Flow Forecasting

Multi-step traffic flow forecasting requires integrating recent observations with historical context, yet determining how that context should shape feature representations and contribute to different prediction steps remains challenging. We propose a memory-conditioned spatio-temporal framework, termed ForePatch-DST, that incorporates predictive memory into both feature encoding and horizon-conditioned fusion. A causal patch encoder trained to predict the next patch summarizes long historical sequences into node-wise memory, which conditions recent-input encoding through a gated residual connection. Learnable horizon queries assign node-specific weights to this memory, parallel temporal and spatial representations, and a periodic reference from the corresponding target interval in the previous week. Experiments on PEMS03, PEMS04, PEMS07, and PEMS08 use the most recent 12 observations to predict the next 12 steps, supplemented by long historical and periodic inputs. Compared with the published results for MDEformer, ForePatch-DST achieves mean absolute errors that are 2.15%, 3.88%, 0.83%, and 4.16% lower on the four datasets, respectively. Ablation results support the contributions of predictive memory and horizon-conditioned fusion within the evaluated framework. These findings indicate that integrating historical context into feature construction and prediction-step-dependent source fusion can improve multi-step traffic flow forecasting.