Abstract: In order to weaken the nonlinearity and non-stationarity of the runoff time series and improve the accuracy of daily runoff forecasts in different forecast periods, a new VMD-PSO-LSTM multi-step forecasting combined model is proposed. First, the variational modal decomposition (VMD) method is used to decompose the original daily runoff sequence into a set of sub-sequences, and the long and short-term memory (LSTM) model parameters are optimized by the particle swarm optimization algorithm (PSO), and each sub-sequence is established. In the PSO-LSTM model, the prediction results of each component are reconstructed to integrate the prediction results. The VMD-PSO-LSTM model is applied to the multi-step prediction of daily runoff at Huayuankou and Lijin Stations on the lower reaches of the Yellow River. Nash sutcliffe efficiency coefficient, correlation coefficient and root mean square error are used to evaluate the prediction results of the model. The results show that the E_NS and R of the two stations are above 0.90 when the forecast period is 1 d, 2 d and 3 d. The comparison with the prediction results of CEEMD-PSO-LSTM and PSO-LSTM models shows that this model can effectively improve the accuracy of multi-step daily runoff prediction. It is an efficient and stable runoff prediction model, and provides a new method for multi-step daily runoff.