Improving the GFS Model Using Observational Data and Transformer Based Deep Learning

Accurate next day precipitation guidance is vital for urban flood risk management and water resources operations, yet raw numerical forecasts often suffer from bias and event detection errors. This study proposes a Transformer based post processing approach that fuses GFS outputs with station observations to improve daily rainfall prediction for Houston, Texas. The model is an encoder only Transformer with two task specific heads: a classification head that estimates the probability of rain occurrence and a regression head that predicts rainfall amount. At inference, a hard gating rule retains the amount prediction only when the estimated probability of rain exceeds a fixed threshold, thereby suppressing spurious light rain outputs. Data consist of daily records from the William P. Hobby International Airport station spanning January ۲۰۱۰ to September ۲۰۲۴. Features include core meteorological variables, a compact set of physically motivated composites, and calendar indicators. The series is partitioned chronologically into training, validation, and testing subsets with seventy, ten, and twenty percent of the data respectively. On the held out test set, the hard gated Transformer improves upon raw GFS across accuracy and association measures: MSE decreases from ۰.۳۵۳۱ to ۰.۱۹۳۲, MAE from ۰.۲۱۲۲ to ۰.۱۷۱۲, and RMSE from ۰.۶۰۳۱ to ۰.۴۲۶۳; R۲ rises from ۰.۴۷۰۸ to ۰.۵۶۱۴ and CC from ۰.۷۱۰۶ to ۰.۷۶۹۷. The false alarm ratio drops from ۰.۵۲۷۹ to ۰.۱۹۳۳, while the probability of detection decreases modestly from ۰.۶۳۹۶ to ۰.۶۰۹۶, reflecting a more conservative and operationally reliable prediction. These results demonstrate that a multi task Transformer with hard gating offers a principled and effective calibration layer for enhancing GFS precipitation guidance in a flood prone coastal metropolis.

کلیدواژه ها:

GFS ، Precipitation ، Transformer ، Deep learning ، Rainfall prediction