Evaluation of LSTM vs. conceptual models for hourly rainfall runoff simulations with varied training period lengths

Abstract

Accurate high-resolution runoff predictions are essential for effective flood mitigation and water planning. In hydrology, conceptual models are preferred for their simplicity, despite their limited capacity for accurate predictions. Deep-learning applications have recently shown promise for runoff predictions; however, they usually require longer input data sequences, especially for high-temporal resolution simulations, thus leading to increased model complexity. To address these challenges, this study evaluates the robustness of two novel approaches using Long Short-Term Memory (LSTM) models. The first model integrates the outputs of a simple conceptual model with LSTM capabilities, while the second model is a stand-alone model that combines coarse and fine temporal inputs to capture both long and short dependencies. To ensure accuracy and reliability, we utilized a century-long meteorological dataset generated from a sophisticated physics-based model, eliminating any influence of measurement errors. The training phase employed multiple sub-periods ranging from 7- to 50-year, with a separate 50-year subset for validation. Our findings highlight the consistent improvement of both LSTM models with increasing training dataset lengths, while conceptual models show no notable enhancement beyond 15 years of training data. Both LSTM models demonstrate superior performance in capturing the reference flow duration curve, offering a promising pathway for more computationally efficient models for runoff predictions.