A Regionalized Attention-LSTM Framework for Multi-Step Probabilistic Flash-Flood Forecasting in Ungauged Catchments

Flash-flood warning in ungauged catchments needs forecasts that are sub-daily, multi-step, interpretable and accompanied by uncertainty, yet calibration-based models cannot be applied where streamflow records are absent. This paper presents a methodological framework, together with a pre-specified evaluation protocol, for regionalized, multi-step, probabilistic flash-flood forecasting in ungauged catchments. No empirical results are reported. The framework couples a stacked encoder LSTM conditioned on static catchment attributes, an additive temporal attention layer, and a sequence-to-sequence decoder that issues 1–6 h discharge or stage forecasts from radar- and satellite-derived precipitation estimates, numerical weather prediction (NWP) forcings and antecedent-wetness indices. Training has two stages: the encoder and attribute embedding are pretrained at daily resolution on gauged donor catchments from CAMELS-IND, using observed streamflow only, and are then adapted at hourly resolution on the subset of catchments with sub-daily telemetry. Transferability is to be tested by withholding every catchment of the Periyar River basin (Western Ghats, Kerala) from all training stages. Monte Carlo dropout provides ensemble forecasts from which predictive intervals are derived. The protocol specifies an ablation spanning attention, attribute conditioning and pretraining, a benchmark suite from persistence to recurrent models, deterministic, event-detection and probabilistic metrics, six testable hypotheses with decision rules, a statistical analysis plan, and a data-audit gate. Fixing designs, splits and success criteria before target-catchment evaluation is intended to make later validation reproducible and to separate genuine transfer skill from tuning effects. Limitations, notably the daily-to-hourly data asymmetry and reservoir regulation, are discussed.