Prediction of Intake Vortex Risk by Nearest Neighbors Modeling

Abstract

Vortex formation at intakes can cause damage, clogging, reduced flow efficiency, and even loss of life. For practical prediction of vortex risk, engineers often compare expected design parameters with published data by using parameter proximity to evaluate the relative risk of vortex formation. Unfortunately, this procedure is ill-defined, and the resulting risk estimates are highly subjective. In response, a formal equivalent of the data proximity procedure was developed by implementing the nearest neighbors algorithm on available experimental and field data. This database was partitioned and the machine learning parameters adjusted to obtain a stochastic model with maximum predictive accuracy. Unlike the flow parameters and submergence, the approach geometry was not found to be a significant factor in the model, although this may be attributable to data noise and range of tested values. The final model, which excluded the channel approach geometry, fit all vertical intake vortex formation data to within 0.1% error and perfectly fit the horizontal intake data. Probability charts generated from the model show regions of vortex formation and problems more numerous and larger on average than regions of low vortex probability, thus validating consideration of potential vortex formation risk for conservative intake design.