Observation-Corrected Rainfall-Forecast Disaggregation for LSTM Water-Level Forecasting in an Urban Stream: A Controlled Assessment with Perfect, Synthetic, and Actual Nowcasts

To test how forecast-rainfall errors should be handled in small-stream water-level forecasting, we evaluated the two-stage disaggregation in Korean Patent 10-2619596: uniform 10 min splitting of hourly forecasts (stage 1) and splitting only the remainder after subtracting rainfall observed in the current hour (stage 2). Twenty-one configurations, including ratio correction, were fed to an observation-trained LSTM for Busan’s Oncheon Stream (five seeds; pre-specified criteria). Corrections used only rainfall observed up to the time of each water-level forecast, and verdicts rest on actual nowcasts (2022–2026) released before that time; perfect (observed future rainfall) and synthetic forecasts are simulations only. In a comparison between modeling approaches trained with and without a forecast channel, the forecast-input model with actual nowcasts did not outperform the no-forecast model from 30 min onward, whereas perfect forecasts reached a persistence skill score (PSS) of 0.84–0.86 at 180 min. Stage 2’s remainder-disaggregation variants failed the event-window accuracy criterion (ΔPSS vs. stage 1 ≥ −0.05; worst −0.75) in every run, while ratio correction met it (worst −0.019 to −0.009). Alarm verdicts rest on 5–21 crossings, differed among runs, and changed for 13–16 of 20 forecast-input configurations when six questionable training events were excluded. At the reference threshold, stage 1’s improved detections (2.0–2.8-fold versus no forecasts) were accompanied by increased false alarms 1.10–1.18-fold, which the criteria do not penalize. Post hoc, forecast-trained stage 1 outperformed the no-forecast model by 0.06–0.17 PSS at 180–240 min (one test period). These single-basin results favor ratio correction over remainder disaggregation and, tentatively, forecast training, but do not establish operational forecasting benefits.