基于随机森林的川东北城市区县域臭氧污染预报模型研究
- Study on Ozone Pollution Forecast Model of Urban and County Area in Northeast Sichuan Based on Random Forest
- 基金项目：国家重点研发计划项目(2023YFC3709301)；四川省重点研发项目(2023YFG0129)；国家外国专家项目(G2022036008L)；四川海聚计划高端人才引进项目(25RCYJ0053)
- 刘琰琰
- 成都信息工程大学大气科学学院；成都平原城市气象与环境四川省野外科学观测研究站
- 宋龙娟
- 四川省达州生态环境监测中心站；四川师范大学化学与材料科学学院
- 摘要：基于2018~2021年川东北三城市（达州、南充、广安）的环境空气质量数据及欧洲中心细网格模式预报资料，探究了气象因子非线性交互作用对近地层臭氧（O3）浓度时空变化的影响，并在此基础上构建 川东北三城市O3预报因子库，经随机森林（RF）模型筛选最优特征因子输入模型，构建并优化了未来1~10天的逐日O3浓度预报（RF-O3）模型。研究发现：①2018~2021年川东北三城市的O3污染情况呈逐年改善态势；②RF特征因子识别结果显示川东北三城市O3污染动态预报关键特征因子以气温、云量、起报天O3日最大8h平均（MDA8_O3）浓度、海平面气压等变量为主；③RF-O3预报模型优化后，对于未来1~10天的逐日MDA8_O3浓度预报精度较好，平均绝对误差（MAE）14.35~22.29 μg·m-3、平均绝对百分比误差（MAPE）31.28~48.22%、均方根误差（RMSE）18.31~28.01 μg·m-3、拟合优度（R2）0.66~0.22，前3天R2能维持在0.50以上;④对RF-O3预报模型进行范围与级别预报检验，发现未来1~10天不同预报时效O3级别预报准确率均较高（可达93.54%~97.93%），其中优、良天预报准确率对级别预报准确率有较大贡献，分别为98.14~99.31%与75.80~91.68%。
- Abstract：Based on the ambient air quality data of three cities in Northeast Sichuan (Dazhou, Nanchong, Guang'an) and the fine-grid model forecast data of the European Center from 2018 to 2021, the effects of nonlinear interactions of meteorological factors on the temporal and spatial variations of the near-surface layer ozone (O3) concentration were investigated. On this basis, a library of O3 prediction factors was constructed for the three cities in Northeast Sichuan. The day-by-day O3 concentration forecasting (RF-O3) model for the next 1~10 days was constructed and optimized by screening the optimal feature factor input model with the Random Forest (RF) model. The results showed that:①the O3 pollution in the three cities in Northeast Sichuan showed yearly improvement from 2018 to 2021; ②The results of RF feature factor identification show that the key feature factors for O3 pollution dynamics forecasting in the three cities in Northeast Sichuan are dominated by variables such as air temperature, cloud amount , maximum daily 8h average concentration of O3 (MDA8_O3) on the starting day, and sea level barometric pressure; ③The optimized RF-O3 forecasting model has good accuracy in predicting the day-to-day MDA8_O3 concentration for the next 1-10 days, with the mean absolute error (MAE) of 14.35~22.29 μg·m-3, the mean absolute percentage error (MAPE) of 31.28~48.22%, the root mean squared error (RMSE) of 18.31~28.01 μg·m-3, and the goodness of fit (R2) of 0.66~0.22. The R2 can maintain above 0.50 in the first 3 days; ④The RF-O3 forecasting model is examined for range and level forecasts, and it is found that the O3 level forecast accuracy is high (up to 93.54%~97.93%) for different forecast times in the next 1~10 days, among which the forecast accuracies of good and excellent days contribute more to the level forecast accuracy, which are 98.14~99.31% and 75.80~91.68% respectively.
