With the rapid development of Artificial Intelligence (AI) technology, an increasing number of intelligent algorithms have been used for simulating and forecasting hydrological process, among which the Long Short-Term Memory (LSTM) network is widely studied. The training of artificial intelligence networks often entails a large amount of training data, which contradicts the limitation of hydrological data. In this study, the effect of training data amount on the performance of LSTM network for runoff simulation are evaluated. First, the runoff series of 130 years are randomly generated by K-Nearest Neighbour (KNN) algorithm and SWAT model. The K-Nearest Neighbour (KNN) algorithm is employed for generating the meteorological data series based on the observed data, and the SWAT model is used to obtain the runoff series with the generated meteorological data series. Then, the LSTM models are developed and evaluated, with the 5-year, 10-year, 20-year, 40-year and 80-year data series of rainfall and runoff as training data respectively, and the 50-year data serves as validating data. The results obtained in Yalong River, Minjiang River and Jialing River show that (1) increasing the training data amount can effectively reduce the over-fittings of LSTM network; (2) increasing the training data amount can also improve the prediction accuracy and stability of LSTM network.