SRAF-ID: A Sensor-Reliability-Aware Framework for Robust Traffic Speed Forecasting Under Missing and Faulty Sensor Observations

Reliable traffic speed forecasting depends on trustworthy historical road-sensor observations, yet deployed sensors may exhibit missing values, outages, noise, calibration drift, and stuck readings. Existing forecasting models are commonly evaluated on cleaned inputs, whereas conventional imputation optimizes historical reconstruction rather than downstream prediction. This study presents the Sensor-Reliability-Aware Framework with Identity-Preserved Design (SRAF-ID), a prediction-oriented speed-channel repair front-end trained end to end using only future forecasting loss. The final model requires no controlled fault-location labels during training or inference. SRAF-ID constructs same-sensor temporal and mask-aware graph-neighborhood candidates, combines them through learned two-way softmax fusion, and preserves node-identity and temporal-context features. On raw-time-disjoint 70%/10%/20% splits of the Metropolitan Los Angeles (METR-LA) and California Performance Measurement System Bay Area (PEMS-BAY) datasets, ten-seed matched stress tests cover six window-level controlled perturbations. SRAF-ID reduces faulty-average mean absolute error from 5.12 to 4.82 on METR-LA and from 1.99 to 1.94 on PEMS-BAY, corresponding to relative reductions of 5.7% and 2.4%, respectively. It achieves a lower mean MAE in all 12 dataset-fault comparisons and a lower faulty-average MAE in all ten seeds on both datasets; the clean-input MAE also decreases. Checkpoint-only tests retain positive all-sensor and affected-sensor mean gains in all eight localized dataset-condition pairs, whereas unseen 0.75-standard-deviation global drift produces small adverse means with paired intervals crossing zero. The evidence therefore supports fault-label-free robustness under the defined stress protocols while leaving field-recorded event continuity and fault frequency for external validation.