To improve the accuracy of daily runoff time series prediction and improve the prediction performance of the regularized extreme learning machine (RELM), the optimization performance of the improved dung beetle optimization (IDBO) algorithm and improved dwarf mongoose optimization (IDMO) algorithm was compared and verified, and the WPT-IDBO-RELM and WPT-IDMO-RELM models for daily runoff time series prediction were proposed based on wavelet packet transform (WPT). The daily inflows of the Mudihe Reservoir and Malutang Power Station in Yunnan Province were predicted. The results show that the average absolute percentage errors of the WPT-IDBO-RELM and WPT-IDMO-RELM models in predicting daily runoff for the Mudihe Reservoir are 1.048% and 1.015%, respectively, and 1.493% and 1.478% for Malutang Power Station, which are better than other comparative models. The optimization performance of the IDBO and IDMO algorithms on standard test functions and instance objective functions is better than that of comparative algorithms. The better the optimization performance of the IDBO and IDMO algorithms, the better the hyperparameters of the RELM, and the higher the prediction accuracy of the WPT-IDBO-RELM and WPT-IDMO-RELM models. WPT can decompose the daily runoff series into subseries components that have stronger regularity and are fewer in number, significantly reducing model complexity and computational scale while improving prediction accuracy.