基于二次分解和深度学习的PM_(2.5)集成预测方法

PM_(2.5) Integration Prediction Method Based on Two-Layer Decomposition and Deep Learning

考虑到PM_(2.5)浓度序列具有非线性、非平稳性以及波动性强等特征,通过将不同的模态分解技术与常用机器学习模型、神经网络模型进行组合对比分析,在“分解—聚类—集成”的研究范式下,提出一种融合二层分解技术及弹网正则化长短期记忆神经网络(ELSTM)构建的CEEMDAN-VMD-K-ELSTM组合模型,并利用北京市日均PM_(2.5)浓度数据进行实证检验。研究结果表明,基于“分解—聚类—集成”研究范式的组合模型在RMSE和MAE模型评价指标中,显著优于已有的组合模型。

Considering that PM_(2.5) concentration sequence had the characteristics of non-linearity,nonstationarity and volatility,the paper firstly made a combinatorial comparative analysis of various model decomposition technique,common machine learning model and neural network model,and then proposed the CEEMDAN-VMD-K-ELSTM combined model based on the combination of two-layer decomposition technology and the elasticized regularized long and short-term memory neural network(ELSTM)under the research paradigm of"decomposition-cluster-integration".Furthermore,the paper made an empirical test using the data of daily average PM_(2.5) concentration in Beijing.The results showed that,the combination model based on the"decomposition-cluster-integration"research paradigm was significantly better than existing combination models after the test of the evaluation indicators of both RMSE and MAE models.