空气质量预测的可解释机器学习建模方法  

An Interpretable Machine Learning Modeling Approach for Air Quality Prediction

作　　者：钟琳 颜七笙 ZHONG Lin;YAN Qi-sheng(School of Science,East China University of Technology,Nanchang Jiangxi 330013,China)

机构地区：[1]东华理工大学理学院,江西南昌330013

出　　处：《计算机仿真》2023年第10期511-518,共8页Computer Simulation

基　　金：国家自然科学基金资助项目(71961001)。

摘　　要：为提高空气质量指数的预测精度,提出了差分进化算法(Differential Evolution, DE)优化极限学习机(Extreme Learning Machine, ELM)的空气质量指数预测模型,并在随机森林(Random Forest, RF)对输入特征选择时,利用LIME(Local Interpretable Model-agnostic Explanations)对单个样本进行可解释性分析,以解决特征选择时的可解释性不足的问题。即首先运用随机森林和三次十折交叉验证选取最优特征,再将筛选后的特征作为DE-ELM模型的输入进行空气质量预测。实验结果表明,与未筛选输入变量的DE-ELM、ELM、DE-BP相比,筛选了输入变量的DE-ELM模型有着较高的精度,其均方误差RMSE分别下降了54.55%、5.74%、28.21%,拟合优度决定系数R2为0.9235,与其它模型相比分别提高了12.08%、1.39%、3.60%。LIME模型很好地解释了特征选择时各个特征的贡献率,提高了训练模型的可信度。In order to improve the prediction accuracy of air quality index,a differential evolution algorithm(DE)is proposed to optimize the air quality index prediction model of extreme learning machine(ELM).When Random Forest(RF)selects input features,Local Interpretable Model-agnostic Explanations(LIME)were used to analyze the interpretability of a single sample to solve the problem of insufficient interpretability during feature selection.That is,first using random forest and three ten-fold cross-validation to select the optimal features,and then using the filtered features as the input of the DE-ELM model for air quality prediction.The experimental results show that compared with DE-ELM,ELM,and DE BP models without screening input variables,the DE ELM model with screening input variables has higher accuracy,with a mean square error reduction of 54.55%,5.74%,and 28.21%,respectively.The coefficient of goodness of fit determination is 0.9235,which is 12.08%,1.39%,and 3.60%higher than other models.The LIME model effectively explains the contribution rate of each feature during feature selection,improving the credibility of the training model.

关 键 词：空气质量预测 随机森林 差分进化算法 极限学习机 可解释模型 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]