Abstract: The multi-step forecast accuracy of daily runoff time series is improved. Based on the wavelet packet decomposition（WPD）method,the elephant group optimization（EHO）algorithm and the extreme learning machine（ELM）and deep extreme learning machine（DELM）two predictors,the WPD-EHO-ELM,WPD-EHO-DELM daily runoff time series mixed forecast model,and applied to the multistep forecast of daily runoff time series at Jingdong Hydrological Station in Yunnan Province. Firstly,the two-layer WPD is used to decompose the daily runoff time series data into 4 sub-sequence components to reduce the complexity and instability of the daily runoff series data.Secondly,when the delay time is 1,the Cao method is used to determine each sub-sequence component. Finally,this paper introduces the principle of EHO algorithm,uses EHO to optimize ELM,DELM input layer weights and hidden layer biases,and establishes WPD-EHOELM and WPD-EHO-DELM models to perform multi-step prediction of each sub-sequence component. And the prediction results are added and reconstructed to obtain the final multi-step forecast results of daily runoff. And the WD-EHO-ELM and WD-EHO-DELM models are constructed based on the wavelet（WD）decomposition and the undecomposed EHO-ELM and EHO-DELM models as comparative analysis models. The results show that:（1）WPD-EHO-ELM and WPD-EHO-DELM models have an average absolute percentage error of daily runoff forecast from 1d to 5 d for an example forecast period of ≤9.44%,pass rate ≥89.2%,and certainty coefficient ≥0.99,The accuracy grades are all Grade A,and the forecasting effects are better than other models such as WD-EHO-ELM. Among them,the forecast period is1～3 d. The average absolute percentage error of daily runoff forecast is less than or equal to 1.81%,the pass rate is 100%,and the certainty coefficient is greater than or equal to 0.999 6. The forecast effect is the most ideal.（2）WPD-EHO-ELM and WPD-EHO-DELM models can give full play to the advantages of WPD decomposition,EHO algorithm and ELM and DELM networks,and show high forecast accuracy and stability. The forecast accuracy increases with the number of days in the forecast period.（3）Models and methods can provide a new approach for realizing the multi-step forecasts and precise forecasts of daily runoff time series.