Accurate and timely runoff prediction is a powerful basis for important measures such as water resource management and flood and drought control, but the stochastic of runoff brought by environmental changes and human activities poses a significant challenge to obtaining reliable prediction results. This paper develops a secondary decomposition hybrid mode. In the first stage of model design, the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) is utilized to discover the significant frequencies in the predicted non-stationary target data series, where the inputs to the model are decomposed into intrinsic modal functions. In the second stage, the swarm decomposition (SWD) is required for decomposing the high-frequency components whose time-shift multi-scale weighted permutation entropy (TSMWPE) values remain calibrated to be high into sub-sequences, and further identifying and establishing the data attributes that will be incorporated into the extreme learning machine (ELM) algorithm in order to simulate the respective series of component data aggregated into a comprehensive tool for runoff prediction. The hybrid model shows superior accuracy, with the Nash-Sutcliffe efficiency exceeds 0.95 and qualification rate greater than 0.93, which can be used for decision-making system design as an efficient and accurate model for generating reliable predictions, especially for hydrological prediction problems characterized by non-stationary data.