Reliable drought prediction is crucial for early warning and decision-making regarding drought-related disasters. Flash droughts, characterized by their rapid onset and severe impacts, are challenging to predict and present critical challenges to water resource management. Specifically, streamflow flash droughts (SFDs) are significant as they directly influence water supply for irrigation, industry, and domestic use. Traditional hydrologic models and drought early warning systems are primarily designed to capture slowly evolving drought conditions and often fail to represent rapid transitions in streamflow that characterize the onset of SFDs. Despite increasing attention to the identification and characterization of SFDs, their prediction remains unexplored. This study evaluates the ability of different deep learning architectures, especially the Temporal Fusion Transformer (TFT), in comparison to the widespread baseline Long Short-Term Memory (LSTM) model and a two-source LSTM incorporating static catchment attributes, to predict SFDs based on streamflow percentiles. The model training and performance evaluation were conducted using hydroclimatic datasets from 671 Catchment Attributes and Meteorology for Large-sample Studies (CAMELS) catchments in the contiguous United States. Performance was evaluated using standard accuracy and event detection metrics. Results show that the TFT achieved the highest overall performance, with a median KGE of 0.87, RMSE of 0.13, MAE of 0.10, and a correlation coefficient of 0.91, outperforming the LSTM with static features (median KGE = 0.81) and the baseline LSTM (KGE = 0.78). In terms of detecting SFDs, the TFT showed the highest median detection rate and lowest miss rate across all models. These improvements were consistent across most catchments and hydroclimatic regions, indicating robust performance gains relative to LSTM-based approaches, although some regions with complex hydrologic behavior remained challenging for all models. While model performance varied across hydroclimatic regions, particularly in data-sparse or hydrologically complex areas, the integration of static attributes and attention-based mechanisms consistently improved predictive skill. These findings demonstrate that Transformer-based models can improve prediction of rapid streamflow changes associated with the onset of SFDs; however, the results are based on deterministic predictions and do not explicitly quantify predictive uncertainty. These findings contribute new insights into the spatial patterns and physical drivers of model performance under non-stationary climate conditions.
Bakar et al. (Wed,) studied this question.