Accurate and reliable runoff prediction is a powerful guarantee for the scientific scheduling and efficient utilization of water resources. Data-driven hydrological models weaken the physical processes of the hydrological cycle and establish the mathematical relationship between input and output through training, providing a solution for runoff prediction in watersheds without underlying surface data. Taking the Nanguo River, a first-level tributary of the Lancang River, as an example, this paper utilizes Principal Component Analysis (PCA) to conduct dimensionality reduction on the sample data.Based on the Long Short-Term Memory (LSTM) model, the daily runoff and daily precipitation of the previous L days before the current moment t are taken as the inputs of the model, and the daily runoff at moment t + K is taken as the output, to predict the daily runoff of the Nagouba Hydrological Station with different lead times ranging from 1 to 5 days. The results show that the prediction accuracy of the model decreases continuously with the extension of the lead time. When the lead time is 1-d, the Nash-Sutcliffe Efficiency (NSE) coefficients in both the verification period and the training period is greater than 0.80. The prediction performance is better than that of the three data-driven models of Back Propagation Neural Network (BP), Support Vector Machine (SVM), and Random Forest (RF), providing a new idea for the daily runoff prediction of watersheds without underlying surface data and a new ideas and reference cases for achieving accurate prediction of medium and long-term runoff.