基于时序分解和神经网络的PM2.5浓度预测研究——以沈阳市为例

PM2.5 Concentration Prediction Based on Time Series Decomposition and Neural Network——Take Shenyang as an Example

可吸入细颗粒物PM2.5,其形成与扩散既受人类生产活动影响也受季节气候条件影响,PM2.5浓度变化具有规律与随机交互的非线性特征,传统预测方法遇到较大困难.文章提出了一种基于小波分解的深度学习预测模型WD-LSTM,针对小波分解不具有自适应等局限性提出基于经验模态分解的深度学习预测改进方法EMD-LSTM,对PM2.5浓度时序数据进行经验模态分解以获其在不同时间周期尺度的子序列,最后利用LSTM对各子序列进行预测计算.文章采集了辽宁省沈阳市11个空气质量监测站2017年1月至11月7316条小时级数据,将文章提出的WD-LSTM、EMD-LSTM与LSTM、Xgboost等进行多重对比实验.结果表明,WD-LSTM和EMD-LSTM预测模型总体上具有更高的预测精度、在分站点以及分时间尺度对比中体现出更强的泛化能力,其中EMD-LSTM在高污染情况下相比其他模型表现出更好的预测准确性.

The formation and diffusion of PM2.5 is affected by both human production and seasonal climate conditions.The variation of PM2.5 concentration has the nonlinear characteristics of regular and random interaction,which makes the traditional prediction method difficult.In this paper,a deep learning prediction model WD-LSTM based on wavelet decomposition is proposed firstly,then EMD-LSTM based on EMD-LSTM is proposed for the limitation of wavelet decomposition that is not adaptive.Empirical mode decomposition is carried out on PM2.5 concentration time series data to obtain its subsequences in different time period scales,and then LSTM is used to predict each subsequence calculation.In this paper,7316 hourly data of 11 air quality monitoring stations in Shenyang,Liaoning Province from January 2017 to November 2017 are collected,and the WD-LSTM,EMD-LSTM proposed in this paper are compared with LSTM,Xgboost,etc.The results show that WD-LSTM and EMD-LSTM have higher prediction accuracy in general,and show stronger generalization ability in the comparison of sub stations and sub time scales,and EMD-LSTM has better prediction accuracy than other models in the case of high pollution.