Improving WRF-Chem PM_2.5 predictions by combining data assimilation and deep-learning-based bias correction.

Area of Science:

• Atmospheric Science
• Environmental Modeling
• Data Assimilation

Background:

• Numerical model simulations are crucial for predicting PM2.5 concentrations.
• Data assimilation (DA) and bias correction (BC) are established methods to enhance model accuracy.
• Simultaneous application of DA and BC for PM2.5 forecasting requires further investigation.

Purpose of the Study:

• To compare the effectiveness of DA and BC in improving PM2.5 predictions.
• To develop and evaluate a novel scheme combining DA and BC simultaneously.
• To assess the performance of the combined approach across different geographical regions.

Main Methods:

• Four parallel experiments were conducted using the WRF-Chem model during winter 2019.
• Experiments included a control run, DA using GSI, deep-learning-based BC, and a combined DA-BC approach.
• Performance was evaluated using root-mean-square error (RMSE) for PM2.5 predictions.

Main Results:

• Both DA and BC individually improved PM2.5 predictions within the first 24 hours.
• The combined DA-BC method demonstrated superior performance compared to individual DA or BC.
• The combined approach reduced RMSE by 38.90%–48.86% compared to the control, with significant regional improvements.

Conclusions:

• Simultaneous data assimilation and bias correction offer the most effective strategy for enhancing PM2.5 forecasting accuracy.
• The integrated DA-BC method significantly optimizes air quality predictions across diverse urban agglomerations.
• This study highlights the synergistic benefits of combining DA and BC for operational PM2.5 prediction systems.