Abstract: With the frequent occurrence of air pollution in recent years, it is urgent to establish an effective air quality early warning system. However, most of the existing researches neglect the importance of data preprocessing and air quality evaluation in the design of early warning system, leading the lack of data mining and the deviation of prediction results. A hybrid air quality early warning system is proposed, which consists of three modules namely data preprocessing, prediction and air quality evaluation, respectively. According to the characteristics of the original data, the classical empirical mode decomposition (EMD) is used to decompose the training set. The Lempel Ziv complexity algorithm is applied to identify the sequence after decomposing as high frequency and low frequency components. The data input matrix is obtained according to the average mutual information (AMI). In order to improve the prediction accuracy and stability, the extreme learning machine (ELM) is used to predict the low-frequency sequences. Extreme gradient boosting (XGBoost) algorithm is applied into high-frequency sequences with added multiple factors. Finally, in the air quality assessment module, the primary pollutants of each day is confirmed. In this paper, Nanjing air quality is taken as an example. The results show that the prediction method has higher accuracy and stronger stability than other single models. The evaluation module also provides certain air quality information, forming a complete early warning system and providing scientific basis for decision makers to control air pollution.

Key words: air quality; empirical mode decomposition; average mutual information; xgboost; extreme learning machine