A Comparative Study of a Two-Dimensional Slope Hydrodynamic Model (TDSHM), Long Short-Term Memory (LSTM), and Convolutional Neural Network (CNN) Models for Runoff Prediction

Abstract

Accurate runoff prediction in complex slope catchments remains challenging due to terrain heterogeneity and dynamic rainfall interactions. This study conducts a systematic comparison between a physics-based Two-Dimensional Slope Hydrodynamic Model (TDSHM) and data-driven deep learning models (LSTM and CNN) for runoff forecasting under variable rainfall conditions. Using 214 rainfall–runoff events (2013–2023) from the Qiaotou watershed in Nanjing, China, the TDSHM integrates rainfall momentum, wind effects, and hydrodynamic principles to resolve spatiotemporal flow dynamics, while LSTM and CNN models leverage seven hydrological features for data-driven predictions. Results demonstrate that the TDSHM achieved superior accuracy, with a mean relative error of 10.77%, Nash–Sutcliffe Efficiency (NSE) of 0.801, and Mean Absolute Error (MAE) of 3.17 mm, outperforming LSTM (24.38% error, NSE = 0.751, MAE = 4.61 mm) and CNN (28.10% error, NSE = 0.506, MAE = 6.82 mm). The TDSHM’s explicit physical interpretability enabled precise simulation of vegetation-modulated runoff processes, validated against field observations (92% predictions within ±15% error). While LSTM captured temporal dependencies effectively, CNN exhibited limitations in sequential data processing. This study highlights the TDSHM’s robustness for scenarios requiring mechanistic insights and the complementary role of LSTM in data-rich environments. The findings provide critical guidance for flood risk management, soil conservation, and model selection trade-offs between physical fidelity and computational efficiency.