基于多变量LSTM与注意力机制的成都市空气质量指数预测研究

Research on Air Quality Index Prediction in Chengdu Based on Multivariate LSTM with Attention Mechanism

摘要: 针对成都市空气质量波动剧烈、极端污染频发且现有研究缺乏分层对比的现状，本文基于2020~2025年逐日空气质量监测数据，构建了单变量LSTM、多变量LSTM及融合注意力机制的多变量LSTM三种递进式预测模型，并引入AQI > 150的极端污染专项误差评价。实验以21天为时间步长进行单步预测。结果表明：多变量LSTM引入PM_2.5、PM₁₀、O₃、SO₂、CO、NO₂六项污染物后，MAE、RMSE、R²、MAPE均显著优于单变量模型；融合注意力机制进一步将R²提升至0.7981、MAPE降至23.63%，极端污染预测误差较单变量模型降低约25%。三种模型均有效收敛且无过拟合，注意力机制在多变量基础上显著增强了对关键时段与峰值的捕捉能力。综合来看，融合注意力机制的多变量LSTM模型在常规预测与极端污染场景中均表现出较好的适用性，可为成都市空气质量预测与重污染天气应对提供一定的参考。

Abstract: In response to the severe fluctuations of air quality and frequent extreme pollution events in Chengdu, as well as the lack of layered comparative studies in existing research, this paper constructs three progressive prediction models based on daily air quality monitoring data from 2020 to 2025: univariate LSTM, multivariate LSTM, and multivariate LSTM with attention mechanism. A specific error evaluation for extreme pollution (AQI > 150) is also introduced. The experiment uses a 21-day time step for one-step-ahead prediction. The results show that after introducing six pollutants (PM_2.5, PM₁₀, O₃, SO₂, CO, NO₂), the multivariate LSTM significantly outperforms the univariate LSTM in terms of MAE, RMSE, R², and MAPE. The attention mechanism further improves R² to 0.7981 and reduces MAPE to 23.63%, while the extreme pollution prediction error decreases by about 25% compared with the univariate model. All three models converge effectively without overfitting, and the attention mechanism significantly enhances the ability to capture key periods and peaks on top of the multivariate model. Overall, the multivariate LSTM model with attention mechanism demonstrates promising applicability in both routine prediction and extreme pollution scenarios, and may serve as a useful reference for air quality forecasting and heavy pollution response in Chengdu.