Development of Advanced Data-Intelligence Models for Radial Gate Discharge Coefficient Prediction: Modeling Different Flow Scenarios

This research aims to predict a radial gate's discharge coefficient (C_d) under free and submerged flow conditions using several machine learning (ML) algorithms. Several parameters are used to develop the learning process of the ML algorithms, including the gate radius (R), gate opening height (W), depth of water upstream (Y_o), depth of water downstream (Y_T), trunnion pin height (h), and width (B). For this purpose, various new versions of ML models have been developed, such as the bagging regression tree (BGRT), bidirectional recurrent neural network (Bi-RNN), bidirectional long short-term memory (Bi-LSTM), Light Gradient Boosted Machine (LightGBM) Ensemble, Multiple Additive Regression Trees (MART), and Neural Regression Forests (NRFs). This study was extended to examine the sensitivity of the adopted predictors for C_d prediction. The adopted ML models generally achieved good and acceptable predictability. In quantitative metrics, C_d was accurately predicted using the Bi-LSTM model with a minimum value of mean absolute percentage error (MAPE = 2.245) and maximum Willmott index (WI = 0.861) over the testing phase for the free-flow condition. For the submerged flow condition, the BGRT model attained the best results, with (MAPE = 2.899) and (WI = 0.900).