Abstract: The medium and long term forecast of runoff in areas with little data is related to the medium and long term power generation of power plants, and also has a strong guiding effect on the short term economic operation of power plants. The Yalu River basin is an important clean energy base in Northeast China. Since North Korea controls more than half of the area of the Yalu River basin, it is difficult to share its runoff data with China, which brings certain obstacles to the medium-and long-term runoff forecast of the Yalu River basin. Taking the inflow runoff of Shuifeng Reservoir in Yalu River Basin as the research object, six model methods, namely phase space reconstruction model（local method, global method）, LSTM model, wavelet analysis-LSTM model, coupled phase space reconstruction（local method, global method） and wavelet analysis model, were used to forecast the inflow runoff of Shuifeng reservoir in the medium-and long-term. The accuracy of the prediction results of the above six models was compared by mean absolute error, mean absolute percentage error and qualification rate. The results show that the coupled phase space reconstruction（global method） and wavelet analysis model are used to forecast the annual runoff. In the monthly scale runoff forecast, the results of coupled phase space reconstruction（local method）, wavelet analysis model and wavelet analysis-LSTM model are better from January to May, while the coupled phase space reconstruction（global method） and wavelet analysis model have obvious advantages from June to December. In the 1-year forecast period, the wavelet analysis-LSTM model has a good effect. For the ten-day runoff prediction, the effect of wavelet analysis-LSTM model is better in the 1-day forecast period, and that of coupled phase space reconstruction（global method） and wavelet analysis model in the 3-day forecast period has obvious advantages. The study will support the formulation of medium-and long-term operation plans for Shuifeng Reservoir and downstream power plants.