Evaluating the Performance of Advanced Machine Learning Models in Predicting the Urmia Lake Water Level

Lakes play an important role in the hydrological cycle and predicting the water level in them can provide vital information for the future management of lakes and their ecosystems. To this aim, in this research two models were developed and run for 1-, 2- and 3-month water level forcasts for the Lake Urmia in north-west of Iran; the tree post-pruning using the REPT method and the combined REPT model with the ROF-REPT. The water level time series data from 2001 to 2020 were divided into two categories for model building (from 2001 to 2014) and for model validation (from 2015 to 2020). Different input scenarios were developed and evaluated to find the most effective input scenario of climate variables. Finally, the developed models were evaluated through visual and quantitative criteria. The results showed that the combined ROF-REPT model has a higher performance than the single REPT model for all forecasts of 1-, 2- and 3-months. Nash-Sutcliffe Efficiency was obtained between 0.45 and 0.87 for single models and between 0.53 and 0.95 for combined models. Also, it was shown that the developed models are able to predict the water level up to 3 months ahead.