A seasonal GRBF network for nonstationary time series prediction

A gradient radial basis function (GRBF) neural network has proved to be an efficient tool in dealing with nonstationary and nonlinear time series, but it has not been used to treat seasonal series. In this paper we apply a modified GRBF neural network, named seasonal GRBF, to make a one-step prediction of a nonstationary time series with the property of seasonality. The signal is the random drift of a micro-electromechanical system gyro, and is proved to be a nonstationary seasonal process. The original GRBF model is also used to do this job for a comparison. The experiments show that the GRBF model cannot cope with seasonality while the seasonal GRBF model has a good performance.