I’m benchmarking an LSTM against a calibrated HEC-HMS on a 210 km² piedmont catchment, targeting 3‑hour lead using 5‑min NEXRAD and a USGS 15‑min gage; so far the LSTM yields KGE 0.76 vs 0.63 in WY2019–2023 storms. For those doing short‑lead flash‑flood prediction, which predictors or feature engineering steps gave you the biggest gain?
What moved the needle for me was adding a single ‘decayed rainfall’ feature as an antecedent wetness proxy — an exponential moving sum of the last 6–48 h of 5‑min rain with τ tuned by CV — which bumped KGE by about 0.05–0.1 on 3‑h leads. Small caveat: it underperforms if you don’t mask snow‑affected periods; did you try different τ windows per season?
But have you tried integrating recent rainfall intensity as a feature? It’s made a difference in my work with 3-hour forecasts. @jlee927, curious how your decayed rainfall feature plays into that?
I’ve found that incorporating a moving average of recent storm totals can really enhance the model’s responsiveness for those 3-hour forecasts. For example, adding a 12-hour sum seemed to help in catching the peaks during rapid events. Have you considered tuning the LSTM with that kind of summation approach?