Noise Removal in River Flow Forecasting

Abstract

We report a study of river flow modeling and forecast by using both noisy and thresholded discharge data as inputs to a neuro-wavelet based neural network. The data was was obtained from USGS station 04156000 Tittabawassee River at Midland, Michigan. In the neuro-wavelet network we combine wavelet analysis by using Daubechies wavelet and artificial neural networks to perform river flow forecasting of the Tittabawassee River. We obtain and compare mean squared errors, correlation coefficients, and root mean squared relative errors for three model performances. Results on the potential benefit in predictive power from denoising river flow data are presented and discussed.