基于特征气象因素和机器学习的内蒙古乌海市沙尘天PM10预测

Prediction of PM10 on dust days in Wuhai City of Nei Mongol based on characteristic meteorological factors and machine learning

关键词:  LSTM  GRU  XGBoost  SHAP  沙尘天  PM₁₀  气象因子

Key words:  LSTM  GRU  XGBoost  SHAP  dust days  PM₁₀  meteorological feature

摘要:

PM₁₀ 是降低能见度和危害人体健康的主要污染物之一，其在沙尘天浓度急剧升高，带来显著公共卫生风险。文章基于内蒙古自治区乌海市2019—2023年逐小时监测数据，构建长短期记忆网络 （LSTM）、 门控循环单元 （GRU） 和极端梯度提升 （XGBoost） 模型，仅以气象因子作为输入项，比较3个模型在沙尘天与非沙尘天的PM₁₀预测性能，并引入夏普利加法解释法（SHAP）量化气象因子贡献。结果表明， 3个模型均可实现有效预测 （R² ：0.813—0.816），其中GRU模型在各项误差指标上表现最优。沙尘天的模型拟合度相较于非沙尘天显著提高 （R² ≈0.92），但峰值浓度预测结果仍存在系统性偏差，误差较非沙尘天增加近1倍。SHAP结果表明，大气压力是首要驱动因子，极大风速在沙尘条件下的重要性显著增强，相对湿度对PM₁₀表现出抑制效应，而温度则表现出促进效应。研究验证了纯气象数据驱动PM₁₀短时预测的可行性，也为缺乏历史污染物数据的地区提供了快速预警方案，并为未来利用卫星遥感监测数据和多源气象信息改进预测提供了参考。

Abstract:

Background, aim, and scope PM₁₀ is a critical air pollutant that adversely affects human health and poses significant public health risks, particularly during dust storm events when its concentrations tend to spike abruptly. As a city located in the arid and semi-arid region of Nei Mongol, Wuhai City is particularly vulnerable to dust storms, which frequently trigger severe PM₁₀ pollution episodes. Effective prediction of PM₁₀ levels, especially during these high-impact events, is therefore crucial for timely public warnings and air quality management in this region. Based on hourly monitoring data in Wuhai City from 2019 to 2023, this study evaluated the predictive performance of Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Extreme Gradient Boosting (XGBoost) models driven exclusively by meteorological inputs. The models’ predictive performance for PM₁₀ was compared under dust and non-dust conditions. Additionally, SHapley Additive exPlanations (SHAP) values were employed to quantify the contribution of meteorological drivers. The scope of this study is defined by its deliberate limitation to meteorological data, which tests the sufficiency of these variables for prediction and enhances model portability in data-scarce regions. Materials and methods This study utilized hourly PM₁₀ and meteorological observation data from Wuhai City, spanning 2019 to 2023. The meteorological features included atmospheric pressure, temperature, relative humidity, and wind-related variables. Three models were developed for LSTM, GRU, and XGBoost. Model performance was evaluated using the coefficient of determination (R² ), mean absolute error (MAE), and root mean square error (RMSE). The SHAP framework was applied to enhance model interpretability and quantify the contribution of each meteorological feature to the predictions. Results All three models demonstrated robust predictive skill (R² : 0.813—0.816), with the GRU performing best across all error metrics. The model’s fit on dust days improved significantly (R² ≈0.92) compared to that on non-dust days. However, systematic underestimations of predicted peak concentrations persisted, with the error magnitudes nearly doubling. This highlights a critical limitation in forecasting the severity of extreme pollution events. Discussion Atmospheric pressure was the most influential predictor across all weather conditions. The importance of instantaneous maximum wind speed spiked during dust days, directly reflecting the physical mechanism of local dust entrainment. In addition, relative humidity exhibited a suppressive effect on PM₁₀ by enhancing particle deposition and reducing suspension, whereas higher temperatures promoted PM₁₀ concentrations through boundary layer development and increased turbulence, which facilitated particle lifting and dispersion. These findings highlight that both large-scale meteorology and local meteorological extremes jointly regulate PM₁₀ variability, with their effects being especially pronounced under dust conditions. Conclusions This study confirms that machine learning models driven solely by meteorological data are a viable option for PM₁₀ prediction, particularly for early warning systems in data-scarce regions. Recommendations and perspectives To improve the prediction of extreme PM₁₀ peaks, future research should focus on integrating multisource data. Incorporating satellite remote sensing data and more advanced meteorological parameters could provide crucial information to better quantify dust emission and transport, thereby improving the reliability of forecasts for severe air pollution events.