Prediction of Water Level using Monthly Lagged Data in Lake Urmia, Iran

Prediction of water level fluctuations in lakes is a necessary task in hydrological and limnological studies. Lake Urmia, a hyper-saline lake in the North Western part of Iran, is dealing with a gradual atrophy. In this study, parametric and nonparametric models are used for predicting monthly water level fluctuations in Lake Urmia. Eleven previous water levels in the form of monthly lagged data are used as the known independent variables of the model while lake water level at the twelfth month is considered as the unknown dependent variable to be predicted. Parametric models used in the modelling are multi-linear regression (MLR), additive and multiplicative non-linear regression (ANLR and MNLR) and decision tree (DT) while feed forward back propagation neural network (FFBP-NN), generalized regression neural network (GR-NN) and radial basis function neural network (RBF-NN) are used to represent the non-parametric approach. Monthly water level data in Lake Urmia observed for 1966–2010 are used for the case study. Four criteria, coefficient of determination, Lin’s concordance correlation coefficient, performance index and root mean square percentage error are used in comparison of the models. The first two are considered for the success of the models while the last two for the failure. Success criteria are given a grade between 0 and 10, failure criteria receive a grade from −10 to 0. The summation of the grades is taken as the total grade of each model. It is found that regression models and FFBP-NN are superior to GR-NN, RBF-NN and DT in predicting monthly lake water level.