To deal with the highly random and unstable sequence of Air Quality Index(AQI)，an EMDWTS twolayer decomposition and unstructured data based combined forecast model is proposed.Firstly，the Baidu index keywords are filtered and the corresponding data is extracted，after which the locally linear embedding (LLE) is applied to reduce the dimensions.Secondly，the empirical modal decomposition (EMD) and reconstruction are carried out on AQI historical sequence and dimensionlowering results.Then，the wavelet transform (WT) is adopted to decompose and reconstruct the gained highfrequency sequence.After reconstruction，the Holt exponential smoothing，support vector regression (SVR) and artificial neural network (ANN) are used to forecast the results of twolayer decomposition and the original low frequency and trend sequence.Subsequently,the forecast results are integrated by BP neural network.Eventually， the gained forecast results above are added up and the final prediction results of AQI are obtained.The comparative experiment’s results demonstrate that the forecast model aforesaid can make full use of a variety of data information，and the prediction accuracy is quite high.