Improving Streamflow Prediction Using Multiple Hydrological Models and Machine Learning Methods

Streamflow prediction is crucial for flood monitoring and early warning, which often hampered by bias and uncertainties arising from nonlinear processes, model parameterization, and errors in meteorological forecast. We examined the utility of multiple hydrological models (VIC, H08, CWatM, Noah-MP, and CLM) and machine learning (ML) methods to improve streamflow simulations and prediction. The hydrological models (HMs) were forced with observed meteorological data from the India Meteorological Department (IMD) and meteorological forecast from the Global Ensemble Forecast System (GEFS) to simulate flood peaks and flood inundation areas. We used Multiple Linear Regression, Random Forest (RF), Extreme Gradient Boosting (XGB), and Long Short-Term Memory (LSTM) for the post-processing of simulated streamflow from HMs. Considering the influence of dams is crucial for the effectiveness of HMs and ML methods for improving streamflow simulations and predictions. In addition, ML-based multi-model ensemble streamflow from HMs performs better than individual models, highlighting the need for multi-model-based streamflow forecast systems. The post-processing of streamflow simulated by the hydrological models using ML significantly improved overall streamflow simulations, with limited improvement in high-flow conditions. The combination of physics-based hydrological models, observed climate data, and ML methods improve streamflow predictions for flood magnitude, timing, and inundated area, which can be valuable for developing flood early warning systems in India.