Abstract: To improve the accuracy of multi-step prediction of daily runoff, reduce the computational scale of the model, and enhance the performance of the Coati Optimization Algorithm（COA） and Hybrid Kernel Extreme Learning Machine（HKELM）, a Multi Pole Wavelet Packet Transform（MWPT）-Improved COA（ICOA） algorithm-HKELM daily runoff time series prediction model is proposed. Firstly, using MWPT, the daily runoff time series data is decomposed into 1 low-frequency component and 2 high-frequency components, and a HKELM is constructed by combining local Gaussian radial basis function kernel and global polynomial kernel function; Secondly, the principle of COA algorithm is briefly introduced, and by improving COA based on strategies such as Circle mapping, we propose the ICOA algorithm. The ICOA algorithm is simulated and verified through 8 typical functions, and is compared with the basic COA algorithm, Whale Optimization Algorithm（WOA）, and Grey Wolf Optimization Algorithm（GWO） to verify the optimization performance of the ICOA algorithm; Finally, using ICOA to optimize HKELM hyperparameters（regularization parameters, kernel parameters, weight coefficients）, a MWPTICOA-HKELM model is established, and MWPT-COA-HKELM, MWPT-WOA-HKELM, MWPT-GWO-HKELM, Wavelet Packet Transform（WPT）-ICOA-HKELM, Wavelet Transform（WT）-ICOA-HKELM, and MWPT-ICOA-BP models are compared and analyzed.The models are validated through multi-step prediction examples of daily runoff time series from Jingdong and Baobian hydrological stations in Yunnan Province from 2016 to 2020. The results show that:（1） ICOA has a good improvement effect, and the simulation accuracy is better than COA, WOA, and GWO algorithms.（2） The MWPT-ICOA-HKELM model has better prediction performance than other comparative models, with the best single step prediction performance for instances, better results with 3 and 5 steps ahead, and worse results with 7 steps ahead. The prediction accuracy decreases with the increase of prediction step size.（3） Optimizing HKELM hyperparameters using ICOA can significantly improve HKELM prediction performance, and the hyperparameter optimization effect is better than COA, WOA, and GWO algorithms.