Abstract
A precise streamflow forecast is crucial in hydrology for flood alerts, water quantity and quality management, and disaster preparedness. Machine learning (ML) techniques are commonly employed for hydrological prediction; however, they still face certain drawbacks, such as the need to optimize the appropriate predictors, the ability of the models to generalize across different time horizons, and the analysis of high-dimensional time series. This research aims to address these specific drawbacks by developing a novel framework for streamflow forecasting. Specifically, a hybrid ML model, WKELM-R, is developed to predict streamflow based on daily discharge and precipitation. The model combines ridge regression (RR), locally weighted linear regression (LWLR), and kernel extreme learning machine (KELM) to enhance multi-step-ahead predictions by accounting for both linear and nonlinear characteristics. In data preprocessing, this study applies multivariate variational mode decomposition (MVMD) for decomposition to handle non-stationarity and complexity, Boruta-XGBoost for feature selection to select the optimal inputs and decrease the dimension, and gradient-based optimizer (GBO) for adjustment of model parameters to overcome the need to optimize the appropriate predictors. To demonstrate the ability to handle real-world conditions and different time horizons, WKELM-R was applied to a watershed in North Dakota, USA to forecast discharge for three different time horizons. The results were compared with those from the existing standalone and hybrid models by multi-criteria decision-making (MCDM), demonstrating the efficacy and unique capabilities of the new hybrid model in streamflow forecasting (for the testing level at t + 3: R = 0.992, RMSE = 0.426, NSE = 0.983; at t + 7: R = 0.997, RMSE = 0.249, NSE = 0.994; at t + 14: R = 0.996, RMSE = 0.304, NSE = 0.991).