Abstract: There are some ecological problems in semi-arid areas of the Loess Plateau，such as abominable ecological environment, heavy soil and water loss，unbalanced annual rainfall. Meanwhile rainwater is very limit to collect and store by using rainwater—harvesting engineering and the amount of supplementary irrigation was 18 mm each time. The purposes of this paper were to study the influence degree about supplementary irrigation on maize yield in very independent growth period and find out sensitive period of under supplemental irrigation with water harvesting, to response relationships between yield and water under different stages by mathematical model. Taking ETmin/ETa as relative evapotranspiration (deformed based on Jensen Model), the effects of some factors on model were eliminated by using relative yield and relative evapotranspiration. Relative yield and evapotranspiration were taken as input and output sample of BP Neural Network model, through a large number of compared training and analyzing, to establish model of maize response to water under supplemental irrigation with water harvesting. Meanwhile, compared and verified among the measured value, BP Neural Network model and the deformed Jensen model were coupled in this model. The results showed that simulation accuracy of the BP Neural Network model was higher obviously than the deformed Jensen model, which could reflect sensitivity of water demand in each growing stage and forecast maize yield. The disadvantages of the model were also discussed in the paper.