Runoff forecasting plays an increasing importance in the optimal management of water resources. If the one-step forecasting of a runoff time series is already a challenging task, performing multi-step ahead forecasting in this kind of data is more complicated. However, little is known about how different multi-step-ahead forecasting strategies would perform in forecasting runoff time series. Several multi-step-ahead forecasting strategies have been developed and can be categorized in five types: Iterated (or Recursive), Direct, DirRec (combining Recursive and Direct), MIMO (Multi-Input Multi-Output) and DirMO (combining Direct and MIMO). This paper aims at answering the research question: which strategy for multi-step ahead forecasting using deep learning LSTM predictors brings out the best performance for runoff data. Hence, we performed experiments with these different strategies on three run-off datasets collected at three hydrological stations in the two rivers of Vietnam. The experimental results show that the DirMO and MIMO strategies are better than all the other multi-step ahead forecasting strategies for runoff data using LSTM model.