Prediction model of non-stationary time series parameters for a complex blending process

Considering the difficulty of on-line measurement and the large time-lagging in a complex blending process, a hybrid prediction model is proposed to effectively realize the prediction of non-stationary time series parameters with large fluctuation. Firstly, by wavelet decomposing, the original time series is decomposed into different frequency subseries according to scale. Then, according to the characteristics of each subseries, the ARMA model, BP neural network model and Holt-Winters no seasonal model are respectively used to build the prediction model for the high frequency subseries and the low frequency subseries. Finally, the prediction results of each subseries are synthetized to obtain the prediction value of original time series. The prediction results show that the proposed model has the great advantage for the prediction of non-stationary time series with large fluctuation of the process industry.