The Effect of Nonstationary in the Precipitation Time Series Predictions

In this paper, the problem of non-stationary in the precipitation time series predictions is studied. The author used empirical mode decomposition (EMD) and neural network methods to make predictions of the non-stationary time series. The result shows that these methods can improve the prediction accuracy when they were used to predict the precipitation of Beijing, Tianjin, and Shijiazhuang. However the low accuracy was found in the precipitation prediction of Nanjing. The author studied four groups of prediction data with space time-index plot method and showed that prediction error depends on geometrical structure of the attractor in the data.