基于多阶段聚类的PM_(2.5)质量浓度预测及对比研究

Prediction and Comparative Study of PM_(2.5) Concentration Based on Multi-stage Clustering

本文提出了一个基于多阶段聚类的深度神经网络(Deep Neural Network,DNN)预测模型,用于多步骤PM_(2.5)质量浓度预测。建议的模型包括分解聚类和预测。在聚类部分中,第1阶段采用的是HDBSCAN(Hierarchical Density-based Spatial Clustering of Applications with Noise,HDB)密度聚类来剔除噪点,在此基础上,再进行第2阶段聚类。第2阶段聚类采用的是Kmeans、Agglomerative、高斯混合以及BIRCH聚类算法(Balanced Iterative Reducing and Clustering Using Hierarchies)4种聚类算法。在预测部分中,使用了DNN作为预测器,选取了深圳市11个空气质量监测站的2015全年逐时数据来验证模型的有效性。实验结果表明,基于多阶段聚类的预测模型适合PM_(2.5)质量浓度的多步高精度预测,性能优于无聚类预测模型以及单阶段聚类预测模型。

A deep neural network(DNN)prediction model based on multi-stage clustering is proposed for multi-step PM_(2.5)concentration prediction.The proposed model includes decomposition,clustering and prediction.In the part of clustering,the first stage uses HDBscan density clustering to eliminate the noise,and then carries on the second stage clustering.In the second stage,Kmeans,AHClomerative,Gaussian mixture and birch clustering algorithms are used.In the prediction part,the deep neural network(DNN)is used as the predictor,and the hourly data of 11 air quality monitoring stations in Shenzhen are selected to verify the effectiveness of the model.The experimental results show that the prediction model based on multi-stage clustering is suitable for multi-step high-precision prediction of PM concentration,and its performance is better than DNN model and single-stage clustering prediction model.