Embedding machine learning techniques into a conceptual model to improve monthly runoff simulation: A nested hybrid rainfall-runoff modeling

Abstract

One of the frequently adopted hybridizations within the scope of rainfall-runoff modeling rests on directing various outputs simulated from the conceptual rainfall-runoff (CRR) models to machine learning (ML) techniques. In those coupled model exercises, after the parameter calibrations of the CRR models are made, their specific outputs constitute auxiliary inputs for the ML model training. However, in this parallel hybridization comprising two consecutive processes, performing the cascade calibration of CRR and ML models increases the computational complexity. Moreover, the mutual interaction between the parameters governing CRR and ML models is also not considered. In this study, to cope with the handicaps mentioned, artificial neural networks (ANN) and support vector regression (SVR) were separately embedded into a monthly lumped CRR model. The dynamic water balance model (dynwbm) was preferred as the CRR model. Then, all free parameters within these nested hybrid models were calibrated simultaneously. The ML parts within the nested schemes manipulate various output variants derived with three conceptual parameters for monthly runoff simulation. These new hybrid models equipped with an automatic calibration algorithm were applied at several locations in the Gediz River Basin of western Turkey. The performance measures regarding mean and high flows indicated that the nested hybrid models outperformed the standalone models (i.e., dynwbm, ANN, and SVR) and coupled model variants. Thus, the credibility of a novel modeling strategy, which takes advantage of the supplementary strengths of a conceptual model and different ML techniques, was demonstrated.