Lake level forecasting with radial based neural networks

Abstract

Many water resources across the world have increasing and decreasing water levels. The
change of water level in lakes, which is one of the water resources, is associated with climate
change and the effects of climate change can be seen at lake levels the fastest. Lake MichiganHuron studied in this study is an 8 km wide body of water formed by the confluence of Lake
Michigan and Huron. Lake Michigan-Huron is the largest freshwater lake in the world. The aim
of this study is to estimate the water level of Lake Michigan-Huron in the USA. For this purpose,
radial-based artificial neural networks were used. In the forecast model, lake levels in the past
months and periodicity number were used as input data. The lake water level (m) data used
has a record length of 104 years (1918-2021). All data is divided into 4 parts (M1, M2, M3 and
M4). 75% of all data was used for the training phase (M1+M2+M3) and 25% for the testing
phase (M4). The test sections were changed from M1 to M4 so that the training and testing
rates remained constant. Mean absolute error (MAE), root-mean-square error (RMSE) and
coefficient of determination (R2) were used as evaluation criteria. As a result, it is seen that
the models make very good predictions in all data sets and in the training-test phases.
However, according to the test results, the data set that gives the most successful results is the
M1 package and the input set using data that has been lag time for 7 months.