基于STL-分解与LSTM-RF组合模型的PM_(2.5)浓度预测

PM_(2.5) Concentration Prediction Based on the Combination of STL Decomposition and LSTM-RF Ensemble Model

细微颗粒(PM_(2.5))与大气环境、人体健康息息相关。为了能及时、准确的估算出PM_(2.5)浓度及污染等级,本研究分别构建了STL-LSTM-RF的PM_(2.5)浓度预测模型,实验区域选择上海市,选取2021年逐小时PM_(2.5)浓度数据、其余空气污染物数据,和欧洲中期天气预报中心3km*3km再分析数据开展PM_(2.5)1-6H的预测实验。结果表明STL-LSTM-RF模型相比LSTM拟合效果在各个时间尺度上均有提升,在1-6小时内的预测上成绩良好,同时捕捉突变数据的技术也有提高。由于数据滞后值的增加,LSTM的预报效能有明显减弱的倾向,而STL-LSTM-RF模拟效能降低的水平则要明显小于LSTM模拟。综上所述,STL-LSTM-RF模拟可以做到更有效、精确的预报PM_(2.5)浓度,也可以为民众生活出行、企业生产生活、政府部门决策等领域提供技术依据,因此具备了很大的使用价值。

Fine particulate matter(PM_(2.5))is closely related to atmospheric environment and hu-man health.In order to estimate PM_(2.5) concentration and pollution level in a timely and accurate manner,this study constructed STL-LSTM-RF models for PM_(2.5) concentration prediction.The experimental area selected was Shanghai,and hourly PM2.5 concentration data for 2021,as well as data on other air pollutants,and 3km*3km reanalysis data from the European Centre for Me-dium-Range Weather Forecasts,were used to conduct PM_(2.5) prediction experiments for 1-6 hours.The results show that the STL-LSTM-RF model has improved fitting effect compared to LSTM at all time scales,achieving good performance in predicting PM_(2.5) concentration for 1-6 hours and improving the ability to capture sudden changes in data.Due to the increase in lagged values of data,the predictive efficiency of LSTM has significantly decreased,while the decrease in simulation efficicncy of the STL-LSTM-RF model is significantly smaller than that of LSTM.In summary,STL-LSTM-RF simulation can achieve morc effective and accurate predic-tion of PM_(2.5) conccntration,and can also provide technical support for people's daily travel,cn-terprise production and life,and government decision-making.Therefore,it has great practical value.