VELOCITY DISTRIBUTION IN A CONTROLLED HYDRAULIC JUMP IN A COMPOUND CHANNEL: AN EXPERIMENTAL AND MACHINE LEARNING (ML) STUDY

Abstract

The experimental study of hydraulic jump characteristics is always a difficult task, particularly regarding the instantaneous velocity, which is difficult to measure experimentally because of its fluctuation and the presence of air bubbles, mainly at the level of the hydraulic jump. However, it is possible to experimentally measure velocities in two-phase flows (air‒water) using an intrusive measuring instrument such as a Microreel. This work will experimentally examine two different hydraulic jumps with two flow rates (26.78 l/s and 17.82 l/s) controlled by two thin sills (H_S = 14 cm and H_S= 15 cm) in a channel composed of a trapezoidal shape with a rectangular base. The results of the measurements of the velocities (relative velocity ) as a function of the positions (relative positions ). The relationship  was analyzed by multiple regression and given a second-degree polynomial function. The coefficients of determination of this function are very high (R² approximately 0.9), which describes a strong relationship between speeds and distances. These results allow us to reflect on the application of learning techniques (ML) to predict relative velocities at any position of the hydraulic jump. Three learning techniques were tested: RBFNN, MPLNN, and GRNN. However, the results prove that the RBFNN has a very accurate speed productivity and records the lowest deviations from the experimental results. Its performance indicators are RMSE = 0.0163 and MAE = 0.0085 for level 1, RMSE = 0.0138 and MAE = 0.0089 for level 2, and RMSE = 0.0166 and MAE = 0.0099 for level 3. However, the MLPNN model gives competitive results to the RBFNN precisely at level 4, where it registers the smallest deviation (RMSE = 0.0457, MAE = 0.0308).