Abstract: Back analysis of rheological parameters is usually necessary in the prediction of long-term deformation of rock-fill dams. In the case of the back analysis of transient deformation mechanical parameters and rheological parameters at the same time, the back analysis is inefficient owing to a large number of parameters and required training samples and complicated network structure. According to the monitoring data of a rock-fill dam, we classified the settlement into transient deformation and rheological settlement, and then employed BP neural network to increase training samples gradually and to train the network circularly, and back-analyzed the transient mechanical parameters and rheological parameters respectively in two steps. The settlement curves of forward analysis by using back-analyzed parameters were consistent with the monitored settlement curves. It's predicted that settlement of the rock-fill dam tends to be stable three years after the normal storage level was reached, and the predicted maximum settlement is in accordance with the monitored data. The results indicate that the number of training samples is small, the inversion is efficient and the deformation prediction is reliable.

Key words: rock-fill dam, rheology, BP neural network, two-step circular back analysis, deformation prediction