A comparison of artificial intelligence approaches in predicting discharge coefficient of streamlined weirs

In the present research, three different data-driven models (DDMs) are developed to predict the discharge coefficient of streamlined weirs (C-dstw). Some machine-learning methods (MLMs) and intelligent optimization models (IOMs) such as Random Forest (RF), Adaptive NeuroFuzzy Inference System (ANFIS), and gene expression program (GEP) methods are employed for the prediction of C-dstw. To identify input variables for the prediction of C-dstw by these DMMs, among potential parameters on C-dstw, the most effective ones including geometric features of streamlined weirs, relative eccentricity (lambda), downstream slope angle (beta), and water head over the crest of the weir (h(1)) are determined by applying Buckingham pi-theorem and cosine amplitude analyses. In this modeling, by changing architectures and fundamental parameters of the aforesaid approaches, many scenarios are defined to obtain ideal estimation results. According to statistical metrics and scatter plot, the GEP model is determined as a superior method to estimate C-dstw with high performance and accuracy. It yields an R-2 of 0.97, a Total Grade (TG) of 20, RMSE of 0.032, and MAE of 0.024. Besides, the generated mathematical equation for C-dstw in the best scenario by GEP is likened to the corresponding measured ones and the differences are within 0-10%.