Water quality prediction is an important basis of water environment planning, evaluation and management, and has important significance to promote the sustainable utilization of water resources and ecological development. According to the uncertainty of various factors in water quality prediction, based on unascertained measure (UM) theory, the method of changing the initial value of the nework to improve BP neural network. And analysis and forecasting were made with the quality data from the Yingluoxia Station in the Heihe River Basin during 1998-2011. To analyze the correlation degree of reference sequence and other factors with gray correlation method (volatile phenol as the reference sequence), and determine the input nodes of BP network which includes CODmn, DO, SO4 2-, Cr6+ and volatile phenol, the output nodes is volatile phenol, thereby establish UM-BP model. The analysis results show that UM -BP prediction model has a higher prediction accuracy than standard BP neural network model. Therefore, the model used in water quality prediction in the Heihe River Basin is feasible.