Multi-Model Ensemble and Reservoir Computing for Efficient River Discharge Prediction in Ungauged Basins

Despite the critical need for accurate flood prediction, water resource management, and climate impact planning, many regions—particularly in Asia, Africa, and South America—face a significant lack of river discharge observation. Although numerous hydrological and machine learning models have been proposed, it is still a grand challenge to achieve rainfall-runoff modeling which is accurate, interpretable, and computationally cheap even under conditions with limited river discharge observation data. We address this challenge by proposing a novel method that leverages multi-model ensemble and reservoir computing (RC). First, we applied Bayesian model averaging (BMA) to 43 “uncalibrated” catchment-based conceptual hydrological models. Second, we trained RC to correct errors in the BMA predictions of river discharge. Since training RC is intrinsically a linear regression to determine the weights of its output layer, there are no iterative computations in the whole process of our proposed method, which significantly enhances computational efficiency. Third, based on both the weights of BMA and RC obtained in gauged river basins, we inferred the corresponding weights for ungauged river basins by linking catchment attributes to these weights. We evaluated this method in 87 ungauged river basins in Japan and found that it achieved a median Kling-Gupta Efficiency (KGE) of 0.55 and a median Nash-Sutcliffe Efficiency (NSE) of 0.52. These results reveal that individual conceptual hydrological models do not necessarily need to be calibrated when an effectively large ensemble is assembled and combined with machine-learning-based bias correction. Furthermore, by leveraging the relationship between observed data and catchment attributes, our method enables river discharge prediction in ungauged basins, making it applicable to a wide range of regions.