Abstract
The possibility of making short-term prediction of rainfall is studied by investigating the existence of chaotic behavior in the rainfall data series. The minimum number of variables essential and the number of variables sufficient to model the dynamics of the rainfall process are identified. The behavior of rainfall over different record lengths is studied. The effects of the data size and the delay time on the correlation dimension estimate are also analyzed. Daily rainfall data of different record lengths from each of six stations in Singapore are analyzed. The correlation dimension method, the inverse approach of the nonlinear prediction method, and the method of surrogate data (to detect nonlinearity) are used in the analysis. The results indicate that the rainfall data exhibit nonlinear behavior and possibly low-dimensional chaos, which imply that short-term prediction based on nonlinear dynamics might be possible. The minimum number of variables essential is identified as 3 and the number of variables sufficient lies in the range between 11 and 18. The results also indicate that the attractor dimensions of data of longer record lengths are greater than that of data of shorter record lengths. The study suggests that a minimum of similar to 1,500 data points is required for the computation of the correlation dimension. Recommendation on the selection of the delay time is also provided.