【摘要】 针对现有机器学习模型预测PM_2.5浓度存在模型过于复杂、没有考虑时空信息和缺失值填补不准确而导致模型性能下降的问题，利用随机森林取代统计学方法填补缺失值，并纳入时空因素提升模型精度.建立了综合遥感数据、气象及协同污染物数据，适用于沿海城市的PM_2.5浓度预测模型（K-means-RF-XGBoost模型）,模型预测耗时仅为BP神经网络的4%.利用2019年大连市实时监测数据对模型PM_2.5浓度预测进行训练和测试，结果表明，建立的K-means-RF-XGBoost模型预测PM_2.5浓度有很高的准确性，与没有考虑时空信息的同种模型相比均方根误差(e_rms)降低了约48%,决定系数（R²）提升了约10%;能有效地预测高PM_2.5浓度并适用于波动范围大的情况，如春季模型在测试集中R²可达0.935;同时在日级预测上表现优异，R²可达0.819.该研究为沿海城市PM_2.5浓度预测提供了新思路.更多还原

【Abstract】 In response to the problem of performance decrease of existing machine learning model for predicting PM_2.5 concentration because that the model is too complex, and does not consider spatio-temporal information and effective missing values imputation is not accurate, random forest is used instead of statistical methods to fill in missing values, and spatio-temporal factors are incorporated to improve model accuracy. Combining remote sensing data, meteorological and collaborative pollutant data, a model（K-means-RF-XGBoost model） suitable for PM_2.5 concentration prediction in coastal cities is established, with a prediction time of only 4% of that of BP neural networks. The prediction of PM_2.5 concentration of the model is trained and tested using real-time monitoring data from Dalian in 2019. The results show that the established K-means-RF-XGBoost model has high accuracy in predicting PM_2.5 concentration, and compared to the same model without considering spatio-temporal information, the root mean square error(e_rms) decreases by about 48%, and coefficient of determination（R²） increases by about 10%. It effectively predicts high PM_2.5 concentrations and is suitable for large fluctuation ranges, such as an R² of 0.935 is achieved in the testing set for the spring model. At the same time, it performs well in daily prediction, with an R² of 0.819. This study provides a new idea for predicting PM_2.5 concentration in coastal cities.更多还原