Fine particulate matter (PM2.5) poses a significant threat to public health in China, necessitating accurate prediction for both public health warnings and emission control strategies. Traditional PM2.5 prediction methods fall into two categories: numerical models (CTMs like WRF-Chem, CMAQ, and GEOS-Chem) and statistical models. CTMs, while providing physically based explanations, suffer from uncertainties in emission rates, meteorological data, and simplified chemical parameterizations. Data assimilation techniques help mitigate these uncertainties but come at the cost of computational efficiency. Statistical models, initially employing linear regression, have evolved to incorporate machine learning (ML) algorithms, addressing non-linear relationships but often falling short in capturing complex spatiotemporal correlations inherent in regional air pollution. Deep learning (DL) networks, with their ability to handle non-linear spatiotemporal correlations, offer a promising alternative. Existing DL models for air quality forecasting often focus on urban-scale, next-day predictions. This research proposes PPN, an advanced spatial-temporal DL model designed for short-range (0-72h) PM2.5 forecasting on a regional scale. PPN integrates the strengths of DL networks, CTMs, and data assimilation techniques, aiming to achieve improved accuracy and efficiency in regional PM2.5 concentration prediction.

The authors review existing PM2.5 prediction methods, categorizing them into numerical and statistical approaches. Numerical models like WRF-Chem, CMAQ, and GEOS-Chem, while providing a mechanistic understanding, suffer from uncertainties related to emissions, meteorology, and chemical parameterizations. Data assimilation techniques are used to improve these models, but computational efficiency remains a concern. Statistical methods, starting with linear regression and progressing to ML algorithms (like XGBoost and Random Forest), have shown promise, but struggle with spatiotemporal correlations. Deep learning (DL) has emerged as a powerful tool, particularly CNN-RNN architectures like CNN-LSTM and GRU-CNN, which have demonstrated success in air quality forecasting. However, many DL models focus on short-term predictions at the urban scale, and don't fully incorporate the dynamics of CTMs and data assimilation. This paper aims to address this gap.

The PPN model employs an encoder-decoder architecture with PredRNN as its backbone, leveraging convolutional layers for spatial feature extraction and LSTM layers for temporal dependencies. The model incorporates meteorological variables, emission data, and derived parameters (changes in SLP, temperature at 850 hPa, annual mean PM2.5, and preceding PM2.5 observations) as input features. These features are categorized into local and non-local layers based on their direct impact on PM2.5, mimicking the processes within CTMs. The encoder phase incorporates multiple preceding PM2.5 observations to improve the initial forecasting field, similar to four-dimensional data assimilation (FDDA) in CTMs. The decoder phase utilizes meteorological variables, emissions, and the previous timestep's PM2.5 forecast. A weighted mean square error (WMSE) loss function is introduced to mitigate the biases potentially introduced by interpolating PM2.5 observations from monitoring stations to a 9km grid using the Inverse Distance Weighted (IDW) method. This weighted loss function gives greater importance to grids closer to monitoring stations. The model was trained and validated using data from 2020-2021, with January and June 2022 serving as the testing datasets. The Beijing-Tianjin-Hebei (BTH) region was selected as the study area due to its high population density and significant air pollution challenges. The meteorological data were obtained from WRF simulations driven by ERA5 reanalysis data. Emission data were from the MEIC inventory. The performance of PPN was compared with three other ML models (Random Forest, XGBoost, and MLP) and the WRF-Chem model with data assimilation.

The PPN model demonstrated strong performance in forecasting PM2.5 concentrations over the BTH region, achieving an R² of 0.7 and RMSE of 17.7 µg m⁻³ for the January 2022 test dataset. Spatial analysis revealed the model's ability to capture the higher PM2.5 concentrations in the south and lower concentrations in the north, reflecting the influence of emission sources and meteorological factors. Temporal analysis showed better performance within the first 24 hours, with R² values of 0.58-0.74 and RMSE of 12-17 µg m⁻³. The performance remained stable for longer forecast lead times (24-72 hours). Comparisons with other ML models (Random Forest, XGBoost, and MLP) highlighted PPN's superior performance due to its ability to capture spatiotemporal correlations. Sensitivity tests showed that the weighted loss function significantly improved the model's accuracy, especially in clean and heavily polluted conditions, reducing the tendency of underestimating high values and overestimating low values. Incorporating preceding PM2.5 observations in the encoder phase also improved the accuracy, especially within the 24-hour forecast horizon. Finally, comparison with WRF-Chem demonstrated PPN's superior predictive accuracy across all forecast lead times, with lower RMSE values (by 1-35%, with 10 cities showing >10% reduction).

The superior performance of the PPN model compared to traditional CTMs and other ML models highlights the power of integrating deep learning techniques with NWP data and data assimilation principles. The ability of PPN to capture complex spatiotemporal patterns in PM2.5 concentrations is a significant advancement in air quality forecasting. The weighted loss function and the incorporation of preceding PM2.5 observations are key contributors to the improved accuracy. The model’s success in capturing the spatial distribution of PM2.5, particularly distinguishing between northern and southern regions, demonstrates the model’s potential in regional pollution management. The findings suggest that DL-based approaches offer a valuable and efficient tool for early warning systems and effective pollution control strategies.

This study presents PPN, a novel deep learning model for regional PM2.5 forecasting that outperforms existing methods. Key innovations include a weighted loss function and the incorporation of preceding PM2.5 observations. Future research could explore the integration of additional data sources (e.g., satellite data), the incorporation of more complex transport processes in the model structure, and the expansion of the model to other regions and pollutants.

The model's performance might be influenced by the accuracy of the input data, particularly the meteorological data and emission inventories. The spatial extent of transport processes considered in the model is limited due to computational constraints. This could affect the accuracy of long-range transport predictions, especially during rapidly changing meteorological conditions. While the WMSE loss function addresses interpolation biases, residual biases may still exist in areas with sparse monitoring stations.