Enhancing monthly lake levels forecasting using heuristic regression techniques with periodicity data component: application of Lake Michigan

Abstract

This study investigates the accuracy of three diferent techniques with the periodicity component for estimating monthly lake levels. The three techniques are multivariate adaptive regression splines (MARS), least-square support vector regression (LSSVR), and M5 model tree (M5-tree). Data from Lake Michigan, located in the USA, is used in the analysis. In the frst stage of modeling, three techniques were applied to forecast monthly lake level fuctuations up to 8 months ahead of time intervals. In the second stage, the infuence of the periodicity component was applied (month number of the year, e.g., 1, 2, 3, …12) as an external subset in modeling monthly lake levels. The root-mean-square error, mean absolute error, and coefcient of determination were used for evaluating the accuracy of the models. In both stages, the comparison results indicate that MARS generally outperforms LSSVR and M5-tree. Further, it has been discovered that including periodicity as an input to the models improves their accuracy in projecting monthly lake levels.