A Temporal Fusion Transformer Framework for 24-Hour Ozone Forecasting Across the Continental United States

Abstract

Ground-level ozone (O3), a harmful air pollutant, poses a significant public health threat, making accurate 24-hour forecasting a critical priority. Predicting ozone concentrations across the continental United States is challenged by complex atmospheric dynamics and the failure of conventional models to capture long-range temporal dependencies. This research introduces a novel forecasting framework that applies a Temporal Fusion Transformer (TFT) to predict hourly ozone concentrations 24 hours in advance. The model is trained on a fused dataset combining ground-truth measurements from approximately 8,000 U.S. AirNOW stations with meteorological and chemical data from the Environmental Protection Agency (EPA) organization. The TFT architecture is uniquely suited for this task, integrating static (e.g., station location) and dynamic (e.g., meteorology, precursor gases) inputs. The model's performance was benchmarked against a Community Multiscale Air Quality (CMAQ) model, achieving a Root Mean Square Error (RMSE) of 6.9 ppb and a Mean Absolute Error (MAE) of 5.3 ppb. This represents a 24.2% and 32.1% improvement over the baseline, respectively. The model’s interpretable attention mechanism identified ozone in parts per billion (O₃), nitrogen dioxide in parts per billion (NO₂), carbon monoxide in parts per million (CO), and organic carbon in micrograms per cubic meter (OC) as the most influential predictors. The TFT model establishes a new state-of-the-art for 24-hour ozone forecasting in the U.S., validating the application of advanced transformer architectures to complex environmental science problems and providing a powerful tool to support more timely public health warnings.