Taking Manshuihe watershed, Beijing as the research object, the characteristics and abrupt change point of the annual precipitation and runoff coefficient from 1956 to 2016 were analyzed. Principal Component Analysis was used to analyze the influence of hourly rainfall and annual rainfall on annual runoff coefficient based on runoff coefficient under the underlying surface condition of 2000-2016. A linear principal component regression model and BP neural network model based on LM Algorithms for annual runoff coefficient and main rainfall factors was established. The results showed that annual runoff coefficient had been declining significantly in the past 61 years, and had three sharp downward trends from 1956 to early 1970s, late 1970s to late 1980s, 2000 to present. The short-term rainfall had a great influence on annual runoff coefficient under the current underlying surface. The correlation coefficient, Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) were 0.99, 0.002 6 and 0.005 respectively between annual runoff coefficient predicted by neural network model and measured value. Compared with regression model, the neural network model gave a better result in simulating annual runoff coefficient.