Development of Smart Self-Adaptive Hydraulic Structures for Flood Discharge Control and Cavitation Mitigation Using Machine Learning

Abstract

Flood discharge control and cavitation damage remain essential problems in the operation and safety of hydraulic structures under extreme hydrological conditions. We propose the design of a smart self-adaptive hydraulic model integrating hydraulic modeling and machine learning principles to effectively control flood discharge and deal with cavitation in a smart structure, using the power-efficient flow of hydraulic fluids. Unsteady hydraulic behavior was simulated to obtain a detailed set of data such as discharge, pressure, velocity, aeration, and turbulence conditions. Three ML models (Support Vector Machine (SVM), k-Nearest Neighbors (KNN) and Convolutional Neural Network (CNN)) have been trained and tested for cavitation risk prediction and adaptive control support. Findings confirm that the proposed smart self-adaptive method considerably surpasses the conventional hydraulic control. During flood peak, the smart system managed to obtain a maximum outflow of approximately 280 m³/s that closely followed the inflow peak of 300 m³/s while conventional control was limited to 235–240 m³/s, reducing discharge mismatch by approximately 15–20%. In contrast to the negative oscillations of −4.0 m to 6.4 m under conventional operation, water levels in the downstream were substantially stabilized, staying within 9.5–13.0 m. The risk of cavitation was significantly reduced, with peak risk indices falling from 0.80–0.85 to 0.63–0.66, equivalent to a reduction of nearly 20–25%, with an increase in the minimum pressure by 20–25 kPa along the spillway surface. Energy dissipation efficiency could be increased from the unstable values of 43–45% with the conventional system to well-steady between 60–75% with smart control, yielding improvements of up to 15%. Overall, we improved the structural safety index by about 0.10–0.15 along the flood event. The best performing ML model was CNN, with about 92% accuracy, a very low prediction error (RMSE ≈ 0.11), and strong robustness to noise (accuracy reduction ≈4.1%); outperforming SVM and KNN.