基于CEEMDAN-SE和LSTM神经网络的PM10浓度预测

针对PM₁₀浓度时间序列具有明显的非线性和波动性特征，提出一种基于自适应噪声的完整集成经验模态分解（complete ensemble empirical mode decomposition with adaptive noise，CEEMDAN）-样本熵（sample entropy，SE）和长短期记忆神经网络（long short-term memory，LSTM）的组合预测模型。首先利用CEEMDAN-SE将原始PM₁₀浓度时间序列分解为若干个复杂度差异明显的子序列；然后针对各子序列的内在特性结合气象因素分别建立适当参数空间的LSTM预测模型；最后将预测结果进行叠加得到最终预测结果。以唐山市4个空气质量监测站的实测PM₁₀浓度数据进行模型验证分析，结果表明:所提预测模型对比其他几种预测模型显示出较高的预测精度，以及良好的普适性。

PREDICTION OF PM10 CONCENTRATION BASED ON CEEMDAN-SE AND LSTM NEURAL NETWORK

In view of the nonlinear and volatility characteristics of PM₁₀ concentration time series, this paper presented a prediction model of PM₁₀ concentration based on complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN)-sample entropy (SE)-long short-term memory (LSTM). The original PM₁₀ concentration time series were decomposed into several sub-sequences with obvious complexity differences by CEEMDAN-SE. Then, an appropriate LSTM prediction model was built by adding meteorological parameters to each different sub-sequence. The final results were got by adding the prediction results. The data of four monitoring stations in Tangshan was used to implement simulation experiment, and the results confirmed that the proposed prediction model showed high prediction precision, and good universality, comparing with other prediction models.