Leveraging Recurrent Neural Networks for Flood Prediction and Assessment

Abstract

Recent progress in Artificial Intelligence and Machine Learning (AIML) has accelerated improvements in the prediction performance of many hydrological processes. Yet, flood prediction remains a challenging task due to its complex nature. Two common challenges afflicting the task are flood volatility and the sensitivity and complexity of flood generation attributes. This study explores the application of Recurrent Neural Networks (RNNs)—specifically Vanilla Recurrent Neural Networks (VRNNs), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU)—in flood prediction and assessment. By integrating catchment-specific hydrological and meteorological variables, the RNN models leverage sequential data processing to capture the temporal dynamics and seasonal patterns characteristic of flooding. These models were employed across diverse terrains, including mountainous watersheds in the state of South Carolina, USA, to examine their robustness and adaptability. To identify significant hydrological events for flash flood analysis, a discharge frequency analysis was conducted using the Pearson Type III distribution. The 1-year and 2-year return period flows were estimated based on this analysis, and the 1-year return flow was selected as a conservative threshold for flash flood event identification to ensure a sufficient number of training instances. Comparative benchmarking with the National Water Model (NWM v3.0) revealed that the RNN-based approaches offer notable enhancements in capturing the intensity and timing of flood events, particularly for short-duration and high-magnitude floods (flash floods). Comparison of predicted disharges with the discharge recorded at the gauges revealed that GRU had the best performance as it achieved the highest mean NSE values and exhibited low variability across diverse watersheds. LSTM results were slightly less consistent compared to the GRU albeit achieving satisfactory performance, proving its value in hydrological forecasting. In contrast, VRNN had the highest variability and the lowest NSE values among the three. The NWM model trailed the machine learning-based models. The study highlights the efficacy of the RNN models in advancing hydrological predictions.