Online data flow prediction using generalized inverse based extreme learning machine

Abstract

Accurate prediction of data flow has been a major problem in big data scenarios. Traditional predictive models require expert knowledge and long training time, which leads to a time-consuming update of the models and further hampers the use in real-time processing scenarios. To relief the problem, we combined Extreme Learning Machine with sliding window technique to track data flow trends, in which rank-one updates of generalized inverse is used to further calculate a stable parameter for the model. Experimental on real traffic flow data collected along US60 in Phoenix freeway in 2011 were conducted to evaluate the proposed method. The results confirm that the proposed model has more accurate average prediction performance compared with other methods in all 12 months.