摘要
在空气污染指数的监测中,传统单项预测方法不能反映PM_(2.5)质量浓度与复杂因素的非线性关系,文章提出一种基于赋权K近邻(K-nearest neighbor,KNN)算法的长短期记忆(long short-term memory,LSTM)神经网络模型来预测PM_(2.5)质量浓度。首先利用相关性分析提取与PM_(2.5)相关性较大的空间因素,并对每个时间节点选取K个近邻,赋予相应权重来表现不同的影响力度;然后通过重构原始数据K倍维度的新数据集,进行LSTM神经网络模型的监督学习训练,提取时间序列的特征和固有的长期依赖关系,最后实现PM_(2.5)日值质量浓度不同未来时刻的预测。实验结果表明,文中提出的赋权KNN-LSTM预测模型具有可行性和有效性,和其他模型相比,表现出较高精度的预测性能。
In the monitoring of the air pollution index,traditional single prediction method cannot reflect the non-linear relationship between PM_(2.5) mass concentration and complex factors.Therefore,a long short-term memory(LSTM)neural network model based on the weighted K-nearest neighbor(KNN)algorithm is proposed to predict PM_(2.5) concentration.Firstly,the correlation analysis is used to extract the spatial factors that are more correlated with PM_(2.5) and K neighbors are selected for each time node,the corresponding weights are given to show different influence forces.Then,based on the reconstruction of the new data set with K dimensions of the original data,the LSTM neural network model is trained with supervised learning to extract the characteristics of time series and the inherent long-term dependence,and finally the prediction of daily PM_(2.5) mass concentration at different future times is realized.The experimental results show that the proposed weighted KNN-LSTM prediction model is feasible and effective,and has higher precision compared with other models.
作者
刘晴晴
陈华友
LIU Qingqing;CHEN Huayou(School of Economics, Anhui University, Hefei 230601, China;School of Mathematical Sciences, Anhui University, Hefei 230601, China)
出处
《合肥工业大学学报（自然科学版）》
CAS
北大核心
2021年第12期1689-1697,共9页
Journal of Hefei University of Technology：Natural Science
基金
国家自然科学基金资助项目(71871001,71771001)。
