LLM-Explained Graph Traffic Forecasting for DOT Corridor Operations: Full Empirical Evaluation on METR-LA and PEMS-BAY with Crash Evidence

Abstract

Short-term corridor forecasting becomes useful to transportation agencies only when the prediction layer and the explanation layer are evaluated together. This paper studies an LLM-explained graph traffic forecasting pipeline on METR-LA and PEMS-BAY and augments the high-error analysis with public crash records. We used 12 historical 5-minute observations to predict the next 12 steps, corresponding to 5-60 minute traffic states, and report the 15-, 30-, and 60-minute horizons with valid-speed masked MAE, RMSE, and MAPE. The comparison includes Persistence, HistoricalAverage, SharedGRU, GraphTemporalCNN, GraphTemporalTransformer, and an ensemble. On METR-LA, the ensemble achieved the best average 12-step MAE at 3.961 mph. On PEMS-BAY, Persistence achieved the best average 12-step MAE at 2.175 mph, while GraphTemporalCNN reduced the 60-minute error relative to Persistence and produced the lowest average RMSE. The explanation layer then localized high-error windows to key sensors and matched the most relevant windows against independent collision records. A METR-LA episode on 2012-06-16 14:10-15:10 matched a nearby Los Angeles collision report at 14:10, and a PEMS-BAY episode on 2017-06-13 18:30-19:55 matched a CCRS rear-end crash on I-280/SR-85 at 18:35. The resulting briefs preserve the forecast residuals, sensor IDs, and external evidence needed for corridor review.