Abstract: In view of the strong nonlinearity and non-stationarity of monthly runoff in the upper reaches of Yangtze River, a hybrid model integrating the chaos theory and an ensemble AdaBoost.RT extreme learning machine is proposed for monthly runoff analysis and prediction. Firstly, the chaotic characteristics of monthly runoff in watershed were researched and revealed based on parameter identification of the runoff system. The optimal delay time and embedding dimension of the monthly runoff time series are deduced. Secondly, with the time series of the reconstructed phase space matrix as input variables, an improved AdaBoost. RT algorithm based on self-adaptive dynamic threshold was incorporated to improve the performance of extreme learning machine. Finally, the optimal chaotic ensemble learning model for monthly runoff prediction was obtained. Results showed that the proposed model could evidently improve the generalization and stability of single extreme learning machine model, and thus achieve better prediction performance.

Key words: runoff forecasting, upper reaches of Yangtze River, chaotic dynamic characteristics, phase space reconstruction, extreme learning machine, integrated prediction