数据分解模式下PM2.5与气态污染物的组合预测研究

大气污染治理是我国实现生态文明的必经之路，制定有效性的大气治理方案，作为参考的大气污染物月均浓度预测结果是至关重要的.针对大气环境污染物月均浓度时间序列的高噪音、非平稳和非线性等特点，本文提出一种基于数据分解模式的组合预测模型.上海市的实例验证及与其他3种模型的对比研究表明：本文所提出的组合预测模型适用于政策制定所需但样本量受限的月均或年均数据预测；所提出的子序列重构的新模式比传统求和算法重构模式提高预测精度12.5%；相较于其他模型，其预测性能最优（绝对百分比误差的均值仅为9.05，且对历史拟合的皮尔逊系数均为0.90以上）.实现了对大气污染物月均浓度高精度预测，可为相关政策的制定提供科学的定量参考.

Hybrid predication of PM2.5 and gaseous pollutants under data decomposition mode

Atmospheric pollution control is the only way which must be passed to achieve ecological civilization. Therefore, the prediction with high-precision for the monthly average concentration of atmospheric pollutants is very importance for formulating effective control plans. In terms of the high noise, non-stationary and nonlinear characteristics of atmospheric pollutant, a novel hybrid prediction model under data decomposition mode is proposed. Empirical and comparison investigations show that the proposed hybrid prediction model is suitable for predicting the limited sample-size monthly or yearly average concentration which is required for policy-making. Compared with the traditional sum algorithm, the reconstruction mode improves the prediction accuracy by 12.5% with the help of the neural network method. Compared with other prediction models, its prediction performance is the best (i.e., the mean absolute percentage error is only 9.05 and the Pearson correlation coefficient of the fitting results for history are greater than 0.90). This method can own the precise prediction for the monthly average concentration of atmospheric pollutants and scientific quantitative reference value is provided for the formulation of relevant policies.