Forecasting of lake water level based on a hybrid model of innovative gunner algorithm

Abstract

The increase in water consumption along with climate change makes freshwater more vulnerable to pollution. Conserving the dwindling freshwater resources is one of the major problems faced by water management experts. Hence, forecasting lake water level (WL) fluctuations is crucial for managing water resources systems. As part of this study, the standalone support vector regression (SVR) model along with the hybrid SVR model combined with the algorithm of innovative gunner (AIG) was employed for WL multistep-ahead (1-3 months) forecasting of Michigan and Superior Lakes. The hybrid model takes advantage of the unique strength of individual models and it can be an effective way to improve forecasting performance. The average mutual information (AMI) and Taylor diagram techniques were applied for WL of lakes up to a lag of 20 months to determine the most effective inputs for model development. Multistep-ahead WL forecasting models of each lake were built based on the training subset (70%; from 1918 to 1991) and the testing subset (30%; from 1992 to 2021). The models were validated statistically and visually in the testing stage. Finally, based on the results, error values were found to be small for AIG-SVR models for both selected stations: Lake Michigan: 0.039 m < RMSE < 0.140 m and Lake Superior: 0.037 m < RMSE < 0.133 m. The test results also revealed that AIG-SVR model has boosted the accuracy of the results. The great potential of the AIG algorithm based on SVR as a novel model was proved in estimating accurate multistep lake WL values with high generalization ability and low variance compared to the SVR model for both lakes.