Economic and industrial advancements have led to a significant decline in air quality. Accurate air quality prediction stands as an effective approach to advance environmental management. The traditional machine learning and neural networks for air quality prediction have insufficient generalization ability and overfitting problems when dealing with high-dimensional, heterogeneous, and highly uncertain air quality data. This study integrates neighborhood rough set (NRS) based attribute reduction with a long short-term memory (LSTM) network and proposes an NRS-LSTM model for PM 2. 5 prediction. In the proposed framework, NRS is first employed to remove redundant attributes from air meteorological data, and the reduced feature subset is then used as the input of the LSTM predictor, with PM 2. 5 concentration as the output variable. Empirical evaluations were conducted on datasets from Xi’an and Baoji, China. In the three-hour prediction task for Xi’an, averaged over four representative seasonal windows, the proposed NRS-LSTM reduced RMSE from 12. 20 to 10. 55, MAE from 9. 67 to 8. 27, and MAPE from 16. 99% to 13. 94%, while increasing the average R² from 0. 8542 to 0. 8986, compared with the standalone LSTM model. In the external validation experiment on the Baoji dataset, NRS-LSTM further achieved an RMSE of 11. 45, an MAE of 9. 28, a MAPE of 12. 73%, and an R² of 0. 9658, outperforming the competing NRS-ENN, NRS-DBN, and NRS-BPNN models. These results demonstrate that the proposed model provides a more accurate and robust solution for PM 2. 5 forecasting and offers a useful quantitative tool for air pollution management and environmental decision-making.
Bao et al. (2026) studied this question.