A Novel Stochastic Tree Model for Daily Streamflow Prediction Based on A Noise Suppression Hybridization Algorithm and Efficient Uncertainty Quantification

Streamflow prediction is one of the critical components of hydrological interactions and a vital step for integrated water resources management for different water-related sectors. Accurate streamflow prediction can provide significant information about flood mitigation, irrigation operation, and land use planning. The study aims to improve data quality and prediction accuracy by remarkably reducing improper noise in streamflow data. In the present study, daily streamflow prediction for Haji Arab station, Gazvin (Iran) from 1969-2020 is conducted for different time scales of 1-week, 2-weeks ahead. First, observed data was analyzed and cleaned with preprocessing techniques to model and predict the streamflow. Due to non-linearity, complexity and erroneous noise of streamflow data, Complete Ensemble Empirical Mode Decomposition with Adaptive Noise method (CEEMDAN) was applied on input data to extract oscillations and noise resulting the decomposed streamflow stationary components. In the next step, streamflow data were modeled with some tree methods (M5 tree, RF, REP tree), and the methods were hybridized with the CEEMDAN method (CEEMDAN-M5 tree, CEEMDAN-RF, CEEMDAN-REP tree). Different quantitatively and visually based criteria metrics such as mean absolute error (MAE), root mean square error (RMSE), Nash Sutcliffe coefficient (NSE), Legate-McCabe index (LMI), and Willmott's Index of the agreement (WI) were applied for model validation. Results revealed that, on the weekly scale, the hybrid CEEMDAN-RF model (NSE:0.924, LMI:0.811, and WI:0.905) outperformed all benchmarked standalone and hybrid models. On the fortnight scale, the hybrid CEEMDAN-M5 tree model (NSE:0.725, LMI:0.504, and WI:0.728) demonstrated superior performance compared to the other models. Preprocessing techniques enhanced the modelling prediction power up to 20% accuracy.