A Multi-model Framework for Streamflow Forecasting Based on Stochastic Models: an Application to the State Of Ceará, Brazil

Reliable long-term (decadal scale) streamflow prediction would provide significant planning information for water resources management, particularly in areas marked by significant variability at those time scales. In this study, a multi-model for prediction using four models that incorporate preprocessing methods along with data-driven forecast models coupled using the least absolute shrinkage and selection operator (LASSO) regression method is proposed. Models utilized complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and wavelet transform (WT) as the decomposition methods and autoregressive (AR) and hidden Markov models (HMM) as the predictive method. The model is evaluated in a comparative analysis with a variety of models previously proposed for hydrological time series prediction. We compare the predictive skill of alternative data-driven models for average annual streamflow (3 ~ 15 years) prediction. Results indicate that the multi-model performed better than the other models, presenting lower values of MAE and RSME. This multi-model can be a reliable tool for forecasting, which can be explored for hydrological data that have remarkably nonlinear and nonstationary features.