Combination of dynamic TOPMODEL and machine learning techniques to improve runoff prediction

Abstract

TOPMODEL has been widely employed in hydrology research, undergoing continuous modifications to broaden its practical applicability and enhance its simulation accuracy. To encompass spatial discretization, diffusion-wave characteristics, depth-dependent flow velocity, and flux estimation in the unsaturated zone, a generalized dynamic TOPMODEL is developed by introducing a greater number of physical parameters. The present study aims to evaluate the optimal combination of these parameters within the dynamic TOPMODEL framework using machine learning techniques to improve the accuracy of runoff predictions and bolster the model's reliability. An innovative training method is suggested to elevate the model's performance by integrating the Long Short-Term Memory (LSTM) algorithm and a topological classification, which relies on the evolving spatial distribution of runoff conditions during floods. The research findings show that the proposed methodology achieves the lowest mean relative error (MRE) at 0.106, the highest Pearson correlation coefficient (PC) at 0.938, and the highest coefficient of determination (R-2) at 0.906 among the three dynamic TOPMODEL types adopted in this study. The effective implementation of a case study in a river basin showcases the feasibility of the proposed method in conjunction with dynamic TOPMODEL and underscores the importance of employing the suggested training procedure.