Sensitive time series prediction using extreme learning machine

Inspired by a multi-granularity and fractal theory, this work mainly focuses on how to conceive a training and test dataset at different levels under a small dataset in a complex real-time application. Such applications do not purely pursue most accurate values, but a low-cost(sub-optimal) solution may be popular during a timely prediction on those sensitive time series. Then a chaotic system is experimented and analysed in detail for three gap-sampling schemes, namely, microscope, middle scale and macro scope. At the same time, the influence of different activation functions on the accuracy and speed of their network model is discussed. The efficiency of sensitive time series using Extreme Learning Machine (ST-ELM) is examined on six widely used datasets (Abalone, Auto-MPG, Body fat, California Housing, Cloud and Strike). The simulations show that the suggested ST-ELM can improve the existing performance when dealing with the idle spectrum prediction of cognitive wireless network.