Abstract: Complex landforms in Karst area such as Karst pipes, fissures and Karst caves leads to unclosed valley and ground water exchange. The total flow at the outlet section is not in absolute linear relation with precipitation owing to the storage adjustment of groundwater reservoir. To overcome the low precision of rainfall runoff forecasting, we established a conventional BP network model and a modified BP network model for runoff forecasting. The average daily precipitation, average daily evaporation, and  runoff in the earlier stage in the basin upstream of Qixingguan Station at Liuchonghe basin were taken as influencing factors. In the modified model, SPASS software was employed to select the influencing factor numbers and adjust the initial weights in the input layer. Logarithm processing is also performed to deal with daily runoff data. It’s found that the modified BP model can increase the precision of large flood and small flood forecasting.

Key words: BP neural network, runoff forecasting, the Karst area, logarithm processing