Abstract: Aimed at the limitation of the traditional intelligent black box model that cannot explain the deformation mechanism of arch dams, we apply the Shapley Additive Explanation (SHAP) theory and deconstruct a machine learning deformation prediction model for a super high arch dam, focusing on analysis of the influence of water pressure, temperature and aging on the radial horizontal displacement of its different parts. From the deformation monitoring data of the dam, we construct a Light Gradient Boosting Machine (LightGBM) black box prediction model that uses SHAP to eliminate the factors with multicollinearity, and analyze the contribution of different influencing factors to model deformation prediction for the whole factor set and a single sample. In a case study of the dam abutment of a super high arch, we examine the relationship of the influencing factors versus its radial horizontal displacements at dam foundation, arch crown, and other dam parts. We find the aging factor has a greater influence on the displacements at the higher elevations close to the arch crown. The temperature factor mainly affects the displacements near the arch crown, and the water pressure factor mainly affects those at higher elevations; while neither of both factors has a considerable effect on the displacements at the measuring points in the dam foundation and the rock mass deep into the abutment. Our method overcomes the shortcoming of poor visibility and unclear internal mechanism of the previous intelligent 'black box' deformation prediction model. Thus, this interpretable model yields the relevant laws that help working performance analysis and operation management of super high arch dams.

Key words: Hydro-engineering, super-high arch dams, monitoring models, SHAP interpretability, LightGBM algorithm, deformation prediction