A Probabilistic Wavelet–Support Vector Regression Model for Streamflow Forecasting with Rainfall and Climate Information Input*

Abstract

Abstract It is essential to explore reliable streamflow forecasting techniques for water resources management. In this study, a Bayesian wavelet–support vector regression model (BWS model) is developed for one- and multistep-ahead streamflow forecasting using local meteohydrological observations and climate indices including El Niño–Southern Oscillation (ENSO) and the Indian Ocean dipole (IOD) as potential predictors. To accomplish this, a two-step strategy is applied. In the first step, the discrete wavelet transform is coupled with a support vector regression model for streamflow prediction. The three key factors of mother wavelets, decomposition levels, and edge effects are considered in the wavelet decomposition phase when using the hybrid wavelet–support vector regression model (WS model). Different combinations of these factors form a variety of WS models with corresponding forecasts. The second step combines multiple candidate WS models with “good” performance via Bayesian model averaging. This integrates the predictive strengths of different candidate WS models, giving a realistic assessment of the predictive uncertainty. The new ensemble model is used to forecast daily and monthly streamflows at two sites in Dongjiang basin, southern China. The results show that the proposed BWS model consistently generates more reliable predictions for daily (lead times of 1–7 days) and monthly (lead times of 1–3 months) forecasts as compared with the best single-member WS models and the adaptive neuro-fuzzy inference system (ANFIS). Furthermore, the proposed BWS model provides detailed information about the predictive uncertainty.