DRIFT: Regime-Aware Self-Supervised Learning, Zero-Inflated Forecasting, and Inference Fusion for Metro OD Demand under Temporal Distribution Shift

Abstract

Accurate origin–destination (OD) demand forecasting is critical for safe and efficient metro operations, yet remains challenging due to temporal distribution shift and structural sparsity. The former destabilizes learned representations across demand regimes, while the latter biases models toward inactive OD pairs. Existing methods typically address these issues in isolation, limiting robustness in real-world settings. This paper proposes DRIFT, a unified framework that jointly handles representation instability and zero-inflated demand. DRIFT integrates three components: (1) regime-consistent self-supervised pre-training for stable spatio-temporal representations; (2) a zero-inflated AutoGate predictor that separates zero occurrence from flow magnitude; and (3) an adaptive inference-time fusion strategy to improve robustness under distribution mismatch. Experiments on 109 months of Taipei Metro OD data (28 stations, three regimes) under a strict no-look-ahead protocol show that DRIFT achieves MAE = 23.69, RMSE = 38.60, MAPE = 11.70%, and Peak Recall = 0.993, outperforming the strongest baseline by up to 9.3%. Ablation studies confirm that each component contributes distinct functionality, and that the core representation and sparsity-aware designs alone, without the inference-time fusion, already surpass all baselines. These results indicate that jointly modeling regime consistency and sparsity is effective for long-horizon OD forecasting.