摘要
Friction factor estimation is essential in fluid flow in pipes calculations. The Colebrook equation, which is a referential standard for its estimation, is implicit in friction factor, f. This implies that f can only be obtained via iterative solution. Sequel to this, explicit approximations of the Colebrook equation developed using analytical approaches have been proposed. A shift in paradigm is the application of artificial intelligence in the area of fluid flow. The use of artificial neural network, an artificial intelligence technique for prediction of friction factor was investigated in this study. The network having a 2-30-30-1 topology was trained using the Levenberg-Marquardt back propagation algorithm. The inputs to the network consisted of 60,000 dataset of Reynolds number and relative roughness which were transformed to logarithmic scales. The performance evaluation of the model gives rise to a mean square error value of 2.456 × 10<sup>–15</sup> and a relative error of not more than 0.004%. The error indices are less than those of previously developed neural network models and a vast majority of the non neural networks are based on explicit analytical approximations of the Colebrook equation.
Friction factor estimation is essential in fluid flow in pipes calculations. The Colebrook equation, which is a referential standard for its estimation, is implicit in friction factor, f. This implies that f can only be obtained via iterative solution. Sequel to this, explicit approximations of the Colebrook equation developed using analytical approaches have been proposed. A shift in paradigm is the application of artificial intelligence in the area of fluid flow. The use of artificial neural network, an artificial intelligence technique for prediction of friction factor was investigated in this study. The network having a 2-30-30-1 topology was trained using the Levenberg-Marquardt back propagation algorithm. The inputs to the network consisted of 60,000 dataset of Reynolds number and relative roughness which were transformed to logarithmic scales. The performance evaluation of the model gives rise to a mean square error value of 2.456 × 10<sup>–15</sup> and a relative error of not more than 0.004%. The error indices are less than those of previously developed neural network models and a vast majority of the non neural networks are based on explicit analytical approximations of the Colebrook equation.
