Abstract: Monitoring the safety of aqueducts is of great significance to ensure the stability of long-distance water delivery in the South-to-North Water Diversion Project. This paper develops a combined prediction model for aqueduct deformation based on time series decomposition and machine learning, aimed at the problem of insufficient prototype observation data mining in previous aqueduct deformation predictions of low accuracy. First, a singular spectrum analysis is used to decompose the deformation monitoring data of an aqueduct into three parts: trend, seasonal, and remainder components. Then, we adopt a kernel-based extreme learning machine to predict the seasonal and trend components, and construct a prediction model of the remainder components using the long short-term memory and phase-space reconstruction theory. These prediction results are superimposed to construct a combined aqueduct deformation prediction model through time series decomposition and machine learning. Against the deformation monitoring data from the Shuangjihe branch aqueduct, this combined model is verified. The results show it is a robust model with a prediction accuracy higher than that of conventional prediction models.

Key words: aqueduct, deformation prediction, time series decomposition, singular spectrum analysis；kernel-based extreme learning machine, long short-term memory