基于逆方差多模型融合的空气质量指数预测方法

空气质量预测对合理制定环境治理政策具有重要意义。针对目前单体预测模型存在模型不稳定和泛化能力不强的问题,提出基于逆方差权重分配方法融合3种单体模型的空气质量指数(air quality index, AQI)预测方法。首先,以北京市为例,构建空气质量指数预测数据集;其次,分别构建长短期记忆网络(LSTM)、门控循环单元(GRU)、双向长短期记忆网络(Bi-LSTM)、自回归积分滑动平均模型(ARIMA)和多元线性回归(MLR)5种模型对数据集进行预测,并对比以上模型的预测结果;最后,在多模型融合方法中,选择逆方差法计算预测精度较高的3种单体模型的权重,根据算得权重构建逆方差融合预测模型。与预测精度较高的3种单体模型以及加权平均融合预测模型相比,逆方差融合预测模型对空气质量指数的预测精度R²分别提高3.9%、3.4%、1.6%和0.5%,达到0.933。结果表明：逆方差融合预测模型综合了各单体预测模型的优点,能够提高AQI预测精度。

AN AIR QUALITY INDEX PREDICTION METHOD BASED ON INVERSE VARIANCE MULTI-MODEL FUSION

The prediction of air quality is of great significance for formulating environmental governance policies. Aiming at the problems of instability and weak generalization ability of the single model method, a multi-model fusion prediction method of air quality index (AQI) based on the inverse variance weight distribution method and fusion of three single model methods was proposed. Firstly, taking Beijing as an example, the air quality index prediction dataset was constructed. Secondly, five models, LSTM, GRU, Bi-LSTM, ARIMA and MLR, were constructed to predict the dataset, and the prediction results of these models were compared. Finally, in the multi-model fusion method, the inverse variance method was used to calculate the weight of three monomer models with high prediction accuracy, and the inverse variance fusion prediction model was constructed according to the calculated weight. Compared with the three monomer models with higher prediction accuracy and the weighted average fusion prediction model, the prediction accuracy, R2 of the inverse variance fusion prediction model for air quality index was improved by 3.9%, 3.4%, 1.6% and 0.5% respectively, reaching 0.933. The results showed that the proposed inverse variance fusion prediction model integrated the advantages of each monomer prediction model, which could improve the prediction accuracy of air quality index.