<p>Accurate streamflow prediction is of great significance for water resources planning and management. While the application of time series decomposition techniques has significantly improved the accuracy of streamflow series prediction, high-frequency components remain a key factor limiting the prediction accuracy. In this study, a new hybrid model combining improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) and support vector machine regression (SVR) based on the combination range of parameters optimization, i.e., ICEEMDAN-C-SVR, was proposed. The ICEEMDAN-C-SVR model predicted high-frequency components using a narrow parameter optimization range and low-frequency components using a wide parameter optimization range. Three meta-heuristic algorithms (sparrow search algorithm, northern goshawk optimization algorithm, and egret swarm optimization algorithm) were used to obtain the parameters of SVR. The performance of the proposed model was evaluated using the streamflow series at Jingle and Zhaishang stations in the Fenhe River Basin. Compared with the ICEEMDAN-SVR model with a single parameter optimization range, the ICEEMDAN-C-SVR model improves the Nash-Sutcliffe efficiency by 3.46% and 4.74%, respectively. Meanwhile, the ICEEMDAN-C-SVR model makes the prediction results more stable, and the results obtained with different meta-heuristic algorithms converge. Overall, the developed ICEEMDAN-C-SVR model achieves high accuracy and robustness in streamflow prediction. It also avoids prediction discrepancies caused by different MHAs in SVR model construction. The results provide positive support for water resource management.</p>