Equivalent Discharge Coefficient of Side Weirs in Circular Channel—A Lazy Machine Learning Approach

• Published in: Water (Basel) Vol. 11; no. 11; p. 2406
• Main Authors: Granata, Francesco, Di Nunno, Fabio, Gargano, Rudy, de Marinis, Giovanni
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.11.2019
• Subjects: Algorithms; Artificial intelligence; Discharge coefficient; Gene expression; Hydraulics; k nearest neighbor; k-star; Learning algorithms; Machine learning; machine learning algorithms; Model accuracy; Prediction models; Rapid flow; side weir; Supercritical flow; Support vector machines; Variables; Weirs
• DOI: 10.3390/w11112406
• ISSN: 2073-4441

Side weirs have been widely used since ancient times in many hydraulic works. Their operation can be analyzed following different approaches. However, almost all possible analysis approaches require knowledge of the discharge coefficient, which depends on several geometric and hydraulic parameters. An effective methodology for predicting discharge coefficient can be based on machine learning algorithms. In this research, experimental data obtained from tests carried out on a side weir in a circular channel and supercritical flow have been used to build predictive models of the equivalent discharge coefficient, by which the lateral outflow can be estimated by referring only to the flow depth upstream of the side weir. Four models, different in the input variables, have been developed. Each model has been proposed in 5 variants, depending on the applied algorithm. The focus is mainly on two lazy machine learning algorithms: k Nearest Neighbor and K-Star. The 5-input variables Model 1 and the 4-input variables Model 2 noticeably outperform the 3-input variables Model 3 and Model 4, showing that a suitable characterization of the side weir geometry is essential for a good accuracy of the prediction model. In addition, under models 1 and 2, k Nearest Neighbor and K-Star, despite the simpler structure, provide comparable or better performance than more complex algorithms such as Random Forest and Support Vector Regression.