Prediction model of PM_2.5 mass concentrations based on fuzzy time series and support vector machine

To solve the instability of multiple-factor regression models and the existence of over-learning and local minima of neural network models in predicting PM_2.5 mass concentration,a method was proposed by combining support vector machine with fuzzy granulation of time series to predict the variation trend and range of PM_2.5 mass concentration. According to the daily periodic variation of PM_2.5 in different seasons,a 24-h pattern was determined to be the window length of granulating. Feature extraction of data samples proceeded by a triangular membership function was applied to support vector machine inputs for regressive modeling,and the optimum parameters of models were selected by grid search based on k-fold cross validation. Then a time series prediction model was established by using 1-h PM_2.5 mass concentration obtained by Wanliu monitoring station at Haidian district of Beijing in 4 seasons from March 2013 to February 2014,and its resolving was realized by LIBSVM tool in MATLAB platform. The results show that the prediction model of PM_2.5 mass concentration based on fuzzy granulation of time series can solve the instability caused by uncertain factors in mechanism modeling and get a good fitting effect on fuzzy granulation parameters.