Abstract: Deformation monitoring data of a rockfill dam are a time series that can be mined using a time series prediction model for analysis of its variation trend. This paper presents a new method for rockfill dam deformation prediction. First, we use a seasonal-trend decomposition procedure based on loess (STL) to decompose the deformation monitoring data of a rockfill dam into three parts: secular trend, seasonal variation, and irregular variation. Then, an empirical mode decomposition (EMD) method is used to stabilize the irregular variation. Finally, we adopt a long short-term memory (LSTM) technique to predict the decomposed sequences and a Bayesian optimization method to optimize the parameters. To evaluate the accuracy of this method, we numerically simulate the Shuibuya concrete faced rockfill dam for different training time, prediction time, and numbers of outliers; and compare it with other time series prediction models. The results show our new method is more accurate and applicable to evaluating rockfill dam performance.

Key words: rockfill dam, deformation prediction, time series decomposition, empirical mode decomposition, LSTM, Bayesian optimization