Suspended particulate matters (PM10) is considered as a harmful air pollutant. Many models attempt to predict numerical levels of PM10 but a simple, clearly defined classification of PM10 levels is more readily comprehensible to the general public rather than a numerical value. However, PM10 prediction model (e.g., support vector machine (SVM)) often suffers from data imbalance problem in the training dataset that results in failure to forecast the minority class of severe cases. In this thesis, a warning system using extreme learning machine (ELM), compared with SVM, was constructed to forecast the class of PM10 level: Good, Moderate, and Severe. An imbalance strategy called prior duplication was also applied to improve the forecast of minority class. The experimental comparisons between ELM and SVM demonstrate that ELM produces superior accuracy relative to SVM in forecasting minority class (Severe) of PM10 level with or without the imbalance strategy. Furthermore, the experimental results show that the required training time and model size in the ELM model are much shorter and smaller than those of SVM respectively, leading to a more efficient and practical implementation of prediction model for large dataset. However, ELM is a batch learning algorithm and many time-series problems such as air pollution index forecast require online sequential learning rather than batch learning. Therefore, this thesis proposes a new method called meta-cognitive online sequential extreme learning machine (MCOS-ELM) that aims to alleviate data imbalance problem and sequential learning at the same time. Under the application of real air pollution data forecast, the proposed MCOS-ELM was compared with retrained ELM and OS-ELM over accuracy and time measures. Experimental results show that MCOS-ELM has the highest efficiency and best accuracy for predicting the minority class (i.e., the most important but with fewest training samples) of air pollution level.