Rainfall prediction using optimally pruned extreme learning machines

Abstract

Rainfall impacts local water quantity and quality. Accurate and timely prediction of rainfall is highly desirable in water management and hydrogeologic hazards mitigation, which is critical for the environmentally sustainable development. Previous rainfall prediction processes are complex and computationally costly for its intrinsic high uncertainty and variability. In this paper, a data-driven approach is applied to predict the rainfall based on historic data via time-series modeling and optimally pruned extreme learning machine (OP-ELM). The rainfall datasets collected from six counties within Three Georges Reservoir are utilized as case studies in this research, first, the rainfall data is pre-processed with outlier removal and missing value imputation, and the monthly ahead difference is computed based on the instant average of monthly rainfall. Next, the autocorrelation function and partial autocorrelation function were computed and the Ljung–Box test statistic is utilized to explore the significance of the historic lagged-series. All the statistically significant historic lags are selected as inputs for prediction algorithms. Last, an OP-ELM algorithm is developed to predict the monthly rainfall with tenfold cross validation. Four activation functions: the sigmoid, sine, hardlim, and radial basis function, are considered in the OP-ELM. The prediction performance is evaluated with metrics including mean absolute error, mean absolute percentage error, root mean square error, and max error rate. Overall, the computational results indicate the proposed framework outperforms the other benchmarking machine learning algorithms through six case studies in Three Gorges Reservoir, China.