# Ensemble neural network models for stability prediction and optimization of hydraulic structures considering uplift pressure and exit gradient

## TL;DR

This paper introduces a new method combining neural networks and genetic algorithms to optimize the stability of hydraulic structures by predicting and minimizing uplift pressure and seepage.

## Contribution

The novel contribution is an ensemble model integrating FFNN, XGBoost, SVM, and GA for stability optimization of hydraulic structures.

## Key findings

- Optimal cutoff wall inclination angle of 165° minimizes uplift pressure and exit gradient.
- Seepage discharge optimal angles vary from 60° to 120° depending on position.
- Ensemble model achieved high R-squared values (0.94–0.99) with low standard deviations across 5-fold cross-validation.

## Abstract

This study aims to develop a novel ensemble modeling approach that integrates artificial neural networks with finite element analysis to optimize the stability of hydraulic structures, particularly through the design of cutoff wall configurations. The research investigates the effects of varying cutoff wall positions and inclination angles on key parameters such as uplift pressure, seepage discharge, and exit gradient. Numerical simulations were performed using Geostudio SEEP/W to analyze seepage patterns across multiple configurations. The proposed methodology combines a Feed-Forward Neural Network (FFNN), XGBoost Regressor, and Support Vector Machine (SVM) with a Genetic Algorithm (GA) to create a predictive optimization framework. The findings reveal that the optimal cutoff wall inclination angle for minimizing both uplift pressure and exit gradient is 165° across all positions, while for seepage discharge, the optimal angle varies by position, ranging from 60° to 120° and increasing incrementally by 15° from upstream to downstream. The ensemble model demonstrated robust predictive performance across 5-fold cross-validation trials, achieving mean R-squared values of 0.99 ± 0.01 for uplift pressure, 0.94 ± 0.02 for seepage discharge, and 0.97 ± 0.01 for exit gradient. The small standard deviations indicate consistent performance across different data partitions, validating model stability and generalizability. The Genetic Algorithm results closely aligned with the numerical model outputs, validating the robustness of the proposed framework. This study introduces a significant improvement over traditional analytical methods by providing an integrated approach that enhances the safety and efficiency of hydraulic infrastructure design, particularly under complex conditions where conventional techniques may fall short.