Event-Aware Validation of Short-Term Water-Level Forecasting in a Single Data-Scarce Gauged Catchment: Implications for Local Flash-Flood Risk Management

Overlapping forecasting windows can conceal the limited independence of event archives. We evaluated seven models using 2039 records from 41 events at Hongjiata Hydrological Station during 2016–2025. Telemetry elevations were harmonized by adding 0.665 m before 2023. Forecasts at 1, 3, 6, and 12 h used 12 or 24 h inputs. Primary validation trained on 2016–2023, validated on 2024, and tested on seven 2025 events. Five neural seeds, event-level evaluation, repeated event-disjoint partitions, and temporal and sample-support sensitivities assessed robustness. With 12 h inputs, Ridge achieved 1 h root mean square error (RMSE) of 0.107 m and Nash–Sutcliffe efficiency (NSE) of 0.908; convolutional neural network–long short-term memory (CNN-LSTM) achieved RMSE of 0.119 ± 0.012 m. The best mean-ranked 6 and 12 h configurations retained negative NSE. On matched 1 h targets from six events, CNN-LSTM strict-minus-random event-mean RMSE was 0.023 m (exploratory 95% bootstrap interval: 0.003–0.047 m). This difference varied across configurations. No 2025 observations exceeded the 31.50 m blue threshold. The results support event-aware assessment of local water-level forecasting without establishing cross-catchment generalization or warning-detection skill.