Bias Correction of the CAMx Air Quality Model Using Deep Learning: A Comparative Evaluation of Station-Specific and Multi-Station Training Approaches

Introduction

Numerical air quality models such as CAMx are essential tools for operational forecasting, yet they often exhibit systematic biases relative to ground-level observations. Machine learning techniques offer a promising approach for post-processing bias correction. However, the choice of training strategy, whether to train models independently at each monitoring site or to pool data across multiple stations, remains an open question, particularly under data-limited conditions.

Methods

In this study, two bias correction approaches were developed and compared for the CAMx model over the city of Thessaloniki, Greece, using hourly data from 2019. The first approach trains separate models for each of the eight monitoring stations (station-specific), while the second merges data from all stations into a single unified model (multi-station). Three deep learning architectures were evaluated: a hybrid LSTM-CNN, a Bidirectional LSTM (BiLSTM), and a Transformer. Input features included CAMx pollutant concentrations, WRF meteorological variables, cyclical temporal encodings, and lag features. Predictions were generated for four pollutants: NO₂, O₃, PM₁₀ and PM_2.5. Model performance was assessed using RMSE and the Pearson correlation coefficient (R).

Results

Both approaches substantially reduced CAMx biases, with RMSE reductions of 62–77% and Pearson R improvements of 96–175% relative to raw CAMx output. The multi-station approach outperformed the station-specific approach in 87–100% of cases, depending on the architecture. The Transformer benefited most from the multi-station strategy, achieving a 31–44% additional RMSE reduction compared to its station-specific counterpart, which is expected given that Transformer architectures typically require larger volumes of training data.

Conclusions

Pooling multi-station data for bias correction effectively overcomes the limitation of restricted training data volume and enables deep learning models to capture broader spatiotemporal patterns, offering a robust strategy for improving operational air quality forecasts.