Application of a Novel Grid-Based Method Using a Wavelet Artificial Neural Network System for Predicting Water Quality Profiles in Deep Lakes: Effects of High and Low Frequency Wavelet Decomposed Components

A method employing artificial neural networks (ANNs) coupled with stationary wavelet transforms (SWTs) was used to estimate water temperature profiles in Boulder Basin, Lake Mead, from May 2011 through December 2014. Surface temperature measurements and stepwise predictions with depth were used to estimate water temperatures through the entire water column. Comparing different modeling scenarios revealed that vertical mixing mode within the water column influenced whether high or low frequency SWT components generated the most accurate water temperature estimates. Rapid temporal and spatial variations in certain parts of the water column increased prediction errors. SWT decomposition revealed that numerical errors in estimated water temperature signals tended to accumulate in specific SWT sub-signals. Excluding those sub-signals from the system improved method performance. ANNs using specific decomposed parts of the input temperature data yielded the best performance, resulting in a coefficient of determination, R2 > 0.96 and maximum relative error of 0.68%.