Abstract: To improve the multi-step forecast accuracy of runoff time series, a novel model was established combining wavelet packet transform（WPT）, hunter-prey optimization（HPO） algorithm and extreme learning machine（ELM）, which was then applied to the multi-step forecast of monthly and daily runoff time series of the Nankang River Hydrological Station in Yunnan Province. The principle of HPO algorithm is introduced, and 6 typical functions are selected to simulate and verify HPO under different dimensional conditions. Then the data of runoff time series is decomposed into 4 subsequence components using double-layer WPT, so as to reduce the complexity and instability of the runoff sequence data. The ELM input layer weights and hidden layer biases are optimized to establish a WPT-HPO-ELM model for the prediction of monthly and daily runoff in multiple steps. The results show that the HPO algorithm has good optimization accuracy and global search ability; the WPT-HPO-ELM model performs ideally at the forecast of 1-3 months monthly runoff, with the mean absolute percentage error≤2.43%, the pass rate≥99.2%, and the coefficient of certainty ≥0.999; it can also present a satisfactory result at the forecast of the monthly runoff with a forecast period of 4-6 months, with the mean absolute percentage error≤15.0%, the pass rate ≥73.3%, and the coefficient of certainty≥0.99; however, it performed poorly when the forecast period is ≥7 months. It has an ideal forecasting effect for the daily runoff with a forecast period of 1-3 d, with the mean absolute percentage error ≤1.23%, the pass rate=100%, and the coefficient of certainty≥0.999; it also performs satisfactorily when forecast period is 4-7 d, with the mean absolute percentage error≤ 15.3%, the pass rate≥ 73.0%, and the coefficient of certainty≥ 0.94; whereas when the forecast period is ≥ 8 d, the forecast effect is poor. The WPT-HPO-ELM model can give full play to the advantages of WPT, HPO and ELM, showing high forecast accuracy and stability; however, the forecast error increases with the increase of the forecast period. The proposed model and method can provide a new approach for the multi-step forecasting of runoff time series.