Comparison of design alternatives for real-time urban flood prediction using field-measured flood-depth data

Minute-level urban flood forecasting performance can vary with lead time, precipitation-input configuration, temporal readout, and forecast format. Using field-observed flood depths at Saebyeok Market in Busan and 1-min precipitation data from three nearby gauges, we compared design alternatives for real-time prediction using precipitation alone. Long short-term memory (LSTM) was the reference encoder, with the Transformer as a secondary comparator. Using identical target timestamps and flood-depth observations, we evaluated gauge combinations; grouped and non-grouped input representations; last-state (LAST), average (AVG), and attention-weighted (ATTN) readouts; and point and horizon forecasts. The final test set contained four flood-producing events, and 90.0% of its targets were 0 cm. Among the grouped configurations, the lowest numerical mean absolute errors (MAEs) for 1-, 10-, 30-, and 60-min point forecasts were 1.07, 1.09, 1.39, and 1.80 cm, respectively; the corresponding lowest mean MAEs for horizon forecasts were 1.21, 1.28, 1.35, and 1.54 cm. In paired moving-block bootstrap analyses under Case 4, all 95% confidence intervals included zero, indicating uncertainty in the paired MAE differences. Event-scale hydrographs showed that the reproduction of peak magnitude and recession varied among events. These findings support the joint selection of precipitation inputs, temporal readout, forecast format, and lead time for target operations.