<p>The effective prediction and simulation of nonstationary hydrological time series play a critical role in the rational allocation of limited water resources. To enhance forecasting accuracy and provide valuable technical support for dispatching operations, this study presents an integrated hydrological time series prediction model that combines mode decomposition, machine learning, and metaheuristic algorithms. First, the adaptive chirp mode decomposition is used to extract components of varying resolutions from the nonstationary hydrological time series. Next, a bidirectional gated recurrent unit is selected as the predictor to capture the complex relationships between inputs and outputs for each component, with computational parameters optimized using the marine predators algorithm. Finally, the predicted outcomes for all components are integrated to generate the final forecasting results. Real-world runoff data from multiple hydrological stations validate the effectiveness of the proposed model. Extensive experiments, evaluated using multiple metrics, demonstrate that the model consistently outperforms traditional approaches in a range of scenarios. Thus, a reliable machine learning tool is offered for accurate hydrological forecasting.</p>