Hourly PM2.5 concentration forecasting based on feature extraction and stacking-driven ensemble model for the winter of the Beijing-Tianjin-Hebei area

摘要

Precise forecasting of hourly PM2.5 concentration is essential for its monitoring and controlling, especially for the winter of the Beijing-Tianjin-Hebei area as severe haze episode occurs frequently. Therefore, this paper explores an hourly PM2.5 concentration forecasting technique based on the Stacking-driven ensemble model and two kinds of input selection methods. Firstly, the partial autocorrelation function (PACF) is employed to select features of time-lagged factors, meanwhile, Spearman correlation coefficient is introduced to extract latent features of exogenous factors, jointly determining the input of the forecasting model. Subsequently, a two-layer Stacking-driven ensemble model is formed as the prediction model to enhance the feature representation and information utilization capacities. Among this ensemble model, BPNN, IBPNN, and ELM are used as base-model while LSSVR is the meta-model. Moreover, four-fold cross-validation is carried out in each base-models to enhance the generalization performance of the model. Finally, the output of each base-model is new input for meta-model to acquire ultimate forecasting values. Case study of hourly PM2.5 concentration forecasting in the Beijing-Tianjin-Hebei area during the winter substantiates that: (1) application of input selection methods is beneficial to satisfactory forecasting results; (2) the forecasting performance of the Stacking-driven ensemble model is far better than any single models composing of it; (3) the proposed model can tackle highly complicated and extremely high concentration PM2.5 data, which is feasible to PM2.5 concentration forecasting during the winter; (4) the proposed model with small forecasting error, strong generalization performance, as well robust forecasting ability is potential in the early air warning systems.