摘要
PM(2.5)是衡量空气污染物浓度的核心指标。通过挖掘PM(2.5)历史数据的时序特性,完成对未来PM(2.5)浓度值的精确预测具有较强的学术意义和应用价值。然而,原始PM(2.5)浓度值时间序列数据相关性对模型的预测精度产生了较大的影响。为了解决这个问题,文中提出一种基于补充总体经验模态分解-皮尔逊相关分析(CEEMD-Pearson)和深度长短期记忆神经网络(Long Short Term Memory,LSTM)混合模型的PM(2.5)浓度预测方法。该方法利用补充总体经验模态分解(Complementary Ensemble Empirical Mode Decomposition,CEEMD)对PM(2.5)浓度历史数据进行不同频率的分解,增强数据中体现的时序特性。然后通过Pearson相关性检验方法对分解后的不同频率子波(IMFs)进行筛选,将筛选后的增强数据输入到多隐含层的深度LSTM网络的输入层进行训练并预测。实验数据表明,CEEMD-LSTM混合模型的预测精度为80%,但是该模型在训练次数为7000次左右才收敛;而经过Pearson二次筛选后的模型在训练800次左右就已经收敛,并且精度提升到87%;CEEMD-Pearson与深度LSTM神经网络混合模型的训练效果最优,在训练650次左右就已经收敛,并且预测精度达到了90%。实验结果说明,CEEMD模态分解方法可以展现出历史数据中的隐藏时序特性,结合Pearson相关性分析进行的二次筛选可有效地提升模型训练的收敛速度和预测精度。因此,基于CEEMD-Pearson和深度LSTM的混合模型可以获得最佳的训练效果、最快的收敛速度以及最精准的预测结果,可以有效解决PM(2.5)浓度预测问题。
PM2.5 is well-known as the key indicator for measuring the concentration of air pollutants.It is of great significance for both academic study and applications to make accurate prediction of future PM2.5concentration values by excavating the time series characteristics of PM2.5historical data.However,the correlation of time series data of the original PM2.5concentration value has great influence on the prediction accuracy of the model.In order to solve this problem,a PM2.5concentration prediction me-thod based on CEEMD-Pearson and deep LSTM hybrid model was proposed in this paper.The CEEMD modal decomposition me-thod is adopted to decompose the PM2.5concentration historical data at different frequencies,and to enhance the timing characte-ristics of the data.Then,the Pearson correlation test method is used to screen the different frequency IMFs after decomposition,and the filtered enhancement data is input to the input layer of the deep LSTM network of multiple hidden layers for training and prediction.Experimental data shows that the prediction accuracy of the CEEMD-LSTM hybrid mo-del is 80%.However,the model converges after 7000 training times.While by means of the secondary screening of Pearson correlation test,the model converges after 800 training times,and the prediction accuracy is improved to 87%.At last,the hybrid model combines CEEMD-Pearson with deep LSTM neural network has the best training effect.It converges after 650 training times,and the prediction accuracy reaches 90%.Experimental results show that the CEEMD modal decomposition method can show the hidden time series characteristics in historical data.The secondary screening combined with Pearson correlation analysis can effectively improve the convergence speed and prediction accuracy of model training.Therefore,based on the CEEMD-Pearson and deep LSTM hybrid models,the best training result,the fastest convergence speed and the most accurate prediction result can be obtained,which can effectively solve the PM2.5concentration prediction problem.
作者
丁子昂
乐曹伟
吴玲玲
付明磊
DING Zi-ang;LE Cao-wei;WU Ling-ling;FU Ming-lei(College of Sciences,ZheJiang University of Technology,Hangzhou 310023,China)
出处
《计算机科学》
CSCD
北大核心
2020年第S01期444-449,共6页
Computer Science
基金
浙江省科技厅“一带一路”专项(2015C04005)。