[Objective] The objective of this paper is to explore the characteristics and forecasting methods of ice regime of upper reach of Wanjiazhai dam and Lamawan reach in Togtoh County after Wanjiazhai reservoir, and provide a basis for ice prevention and scheduling. [Methods] The processes of freezing-up and thawing of the river reach were analyzed based on the field observations from 1998 to 2015, including the dates and durations of both freezing-up and thawing and their relationships to the affecting factors like air temperature, flow rate, ice thickness, and number of days of frozen. The significant correlation coefficients were selected and BP neural network model was applied to predict the durations of the freezing-up, durations of thawing, and dates of thawing of the upstream reach respectively. [Results] The accuracy of the predicting results were up to 93% compared with the observed date. [Conclusion] Neural network model is suitable for forecasting freezing-up duration, breaking-up duration and breaking-date in upper reach of Wanjiazhai dam, and is more suitable for solving the problem of ice regime influenced by complex factors.