Predicting Water Level Fluctuations in Lake Van Using Hybrid Season-Neuro Approach

Abstract

Predicting water level fluctuations in a lake is crucial in terms of sustainable water supply planning, flood control, management of water resources, shoreline maintenance, sustainability of ecosystem, and economic development. This study developed a new predictive model based on the season algorithm (SA) and multilayer perceptron (MP) methods to improve prediction accuracy and extend water-level lead-time prediction. For the first time, the additive season algorithm (ASA) was used as an alternative data preprocessing technique for predicting water levels, and its performance was compared with that of wavelet transform (WT). The prediction accuracy and longer lead-time predictions were improved by the hybrid additive season algorithm–multilayer perceptron (ASA-MP) model. The results indicated that the hybrid additive season algorithm–multilayer perceptron model can be used to predict monthly water levels accurately with up to a 12-month lead time. The combined wavelet–multilayer perceptron (W-MP) model can be used to predict monthly water levels up to 6 months in advance with good agreement, whereas the standalone multilayer perceptron (MP) model can be used to predict the water levels up to 2 months in advance. The combined additive season algorithm–multilayer perceptron model outperformed the W-MP and standalone MP models based on the mean squared error (MSE) and the Nash-Sutcliffe coefficient of efficiency (CE) as performance evaluation criteria. As opposed to the ASA, the WT method has serious drawbacks and complicated mathematical processes. Furthermore, the prediction performance of the W-MP model was not satisfactory. The results of this study indicated that the ASA-MP model outperformed the W-MP model at all lead times. This implies that the season algorithm can effectively eliminate periodicity and trend cycles from original data better than the wavelet transform. In addition, the MP model should be combined with a preprocessing technique for more-accurate and longer lead-time predictions.