Comparing strategies for training LSTM models for street-scale urban flood prediction in Norfolk, Virginia

Norfolk, Virginia, United States For real-time urban flood prediction at the street-scale, both speed and accuracy are critical. Among deep learning algorithms, Long Short-Term Memory (LSTM) networks are effective for time series prediction. However, the optimal approach for training LSTM models for accurate prediction remains debated in the hydrology literature. Using a dataset of 40 flood-prone streets, this study explores strategies for training LSTM models for street-scale flood prediction. Three experiments were designed to (1) compare different data-grouping approaches across global, clustered, and street models, (2) assess how the availability of water-depth information influences prediction accuracy, and (3) test the scalability of the models for newly added streets. Grouping streets into hydrologically similar clusters enhanced prediction accuracy over the global model for the test events, while street models achieved the lowest errors in most cases. This suggests that the uniqueness of street-scale flooding dynamics in urban environments requires hyper-focused model training. When testing model performance with varying water-depth inputs, LSTM models trained on streets experiencing a wide range of flood depths performed well for streets with shallow water depths. The cluster models outperformed the global model in predicting flooding on newly added streets. This suggests that flooding behavior is better captured when the training dataset consists of hydrologically similar streets rather than a diverse set. • Strategies for training LSTM models for street-scale flood prediction are explored. • Training using only local data for a single street outperforms all other models. • Training using hydrologically similar streets outperforms training using all streets. • Training using streets experiencing deep flooding can still predict shallow flooding. • Urban flood dynamics require localized and specific training data.