Integrating analytical and machine learning approaches to simulate and predict dam foundation stress and river valley contraction in a large-scale reservoir

Abstract

The safety of several large-scale reservoirs all over the world has been of concern due to dam foundation stress (DFS) that gradually changes following impoundment inducing the river valley contraction (RVC). Presently, there are limited approaches for the prediction of DFS and RVC based on complex hydro-geomechanics principles. However, these approaches require extensive information that is cumbersome and time-consuming to gather and hence expensive. In this paper, five machine learning models (MLMs) for DFS and RVC prediction were established by merging innovative analytical, BP neural networks and optimized algorithm approaches. Three key influencing factors; namely: seepage, temperature, and creep are used as input information in these models. The developed MLMs were validated using well-documented case study results over nine years for Xiluodu reservoir in China. The trend-fitting effect and statistical indicators of the proposed MLMs demonstrated strong predictive ability (R² > 0.9). Among the MLMs, Generic algorithm-BP and Sparrow search algorithm-BP methods were found to be comprehensive. The predicted RVC and DFS using MLMs are consistent with the coupled multi-field analytical method from the literature and provide reliable predictions using limited information. This study serves as a valuable reference for predicting DFS and RVC of large reservoirs for ensuring long-term safety.