<p>Accurate and timely runoff prediction is essential for effective water resource management and controlling floods and droughts. However, the stochasticity of runoff due to environmental changes and human activities poses a significant challenge in achieving reliable predictions. This paper presents a multi-scale two-phase processing strategy to develop a hybrid model for runoff prediction. In the first phase of model design, the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) is utilised to identify significant frequencies in the non-stationary target data series. The inputs to the model are decomposed into intrinsic modal functions during this stage. In the second phase, the swarm decomposition (SWD) is used to decompose high-frequency components with consistently high values of time-shift multi-scale weighted permutation entropy (TSMWPE) into sub-sequences. This permits further identification and establishment of data attributes that are incorporated into the extreme learning machine (ELM) algorithm. The ELM then simulates the series of component data, creating a comprehensive tool for runoff prediction. The hybrid model demonstrates exceptional accuracy, achieving a Nash-Sutcliffe efficiency greater than 0.95 and a qualification rate exceeding 0.93. This model can be utilised in decision-making systems as an efficient and accurate solution for generating reliable predictions, particularly for hydrological challenges characterized by non-stationary data.</p>