Abstract: Previous models for intelligent prediction of rockfill dam deformation, lacking attention to the uneven distribution of deformation time series over multiple measuring points, are limited to low accuracy. This paper develops a rockfill dam deformation prediction model, CTSA-ConvLSTM, to combine a convolutional neural network (CNN), the attention mechanism, and a long short-term memory (LSTM) neural network. This model extracts the temporal and spatial characteristics of deformation and generates different weights for the measurements taken at different instants and different locations, so that it realizes the adaptive learning of global deformation patterns of a rockfill dam. In the case study of the Shuibuya dam, the model is verified against the deformation data from all the measuring points at the maximum dam section. It performs better than Holt-Winters and other conventional time series prediction models, and its prediction accuracy is higher than that of a LSTM-based deformation model developed by the authors. By extracting the spatiotemporal characteristics of monitoring data through deep learning, it improves the accuracy and provides a new idea for improving dam safety monitoring models.

Key words: deformation prediction of rockfill dam, spatial and temporal correlation, convolutional neural network (CNN), attention mechanism, convolutional long short-term memory network (ConvLSTM)