Abstract

A deep learning-based post-processing (PP) system was developed to improve the predictive skill for ozone and particulate matter less than 2.5 μm in diameter (PM_2.5) forecasts in the Visual atmospheric Environment Utility System (VENUS) operational air quality forecasting system in Japan. The PP system uses a deep neural network trained on paired Weather Research and Forecasting (WRF)/Community Multi-scale Air Quality Modeling System (CMAQ) model outputs and ground-based observations from 2013 to 2016, with its performance evaluated using independent data from 2017. For ozone, PP removed the systematic overestimation bias in the raw forecasts, reducing the mean deviation from 16.5 to − 0.6 ppbv; increased the domain-wide correlation coefficient from 0.657 to 0.870; and decreased the root mean square deviation (RMSD) markedly from 20.7 to 8.1 ppbv. For PM_2.5, PP improved predictive skill, increasing the domain-wide correlation coefficient from 0.544 to 0.671 and reducing the RMSD from 9.0 to 6.5 µg/m³. However, the long-term decrease in observed concentrations can cause overcorrection in the fixed PP model. The PP system corrected errors arising from uncertainties in physical and chemical parameterizations directly and limited model resolution at a lower computational cost than data assimilation. Our findings demonstrate that the PP system is suitable for operational implementation and that strategies such as sequential retraining are important for including long-term concentration trends.