Improving Streamflow Forecasting Efficiency Using Signal Decomposition Approaches

Abstract

This study introduces a novel approach utilizing the Maximal Overlap Discrete Wavelet Transform (MODWT) to enhance daily streamflow forecasting at two USGS stations (14211500 and 14211550) from 1998 to 2021. The MODWT is integrated with three machine learning models: Extremely Randomized Trees (ERT), Artificial Neural Networks (ANN), and Gaussian Process Regression (GPR). Autocorrelation and partial autocorrelation functions were employed to determine relevant lags and generate multiple input variables, which were then analyzed through MODWT to derive multi-resolution analysis features. The hybrid model incorporating MODWT significantly improved prediction accuracy. Among the methods, ANN with MODWT (ANN6_MODWT) demonstrated superior performance compared to standalone ANN, ERT, and GPR models. ANN6_MODWT achieved improvements of 15.60%, 24.70%, 39.74%, and 28.34% in terms of correlation coefficient (R), Nash-Sutcliffe efficiency (NSE), root mean square error (RMSE), and mean absolute error (MAE) at USGS 14211550, and 13.50%, 23.80%, 46.47%, and 34.06% at USGS 14211500. These results underscore the potential of MODWT for enhancing streamflow prediction accuracy.