Abstract: Previous prediction models were limited by their inadequate consideration of temperature hysteresis effects and crack influences of an arch dam, and suffer from overly complex, redundant displacement factors and low prediction accuracy. To achieve accurate predictions of displacement in the arch dams with significant cracks, this paper develops a novel predictive method. First, we construct a displacement monitoring model for the dams, accounting for temperature hysteresis effect and crack influences. Then, a gradient boosting regression tree (GBRT) is used for feature selection among influencing factors, eliminating irrelevant variables; Kernel principal component analysis (KPCA) is applied to extract features from the retained temperature hysteresis and crack factors, so as to construct a displacement prediction dataset. Finally, we construct a displacement prediction model by integrating the salp swarm algorithm with the kernel extreme learning machine (SSA-KELM). Engineering case results demonstrate feature selection and feature extraction effectively mitigate the interference of irrelevant variables and reduce data dimensions, thereby improving prediction accuracy significantly. Compared with other benchmark models, SSA-KELM that presents the highest prediction accuracy and stability is a new viable approach for predicting displacement in arch dams with cracks.

Key words: arch dam displacement prediction, hysteresis effect, crack influence, feature selection, kernel principal component analysis, machine learning