Data-intelligence-driven flood forecasting and early warning in a small river basin: A BiLSTM model integrated with...

Reliable flood forecasts must be translated into timely, accountable warnings, particularly in small mountainous basins with short concentration times and limited historical records. This study links a Bidirectional Long Short-Term Memory (BiLSTM) rainfall-runoff model with Jiangxi Province's hierarchical "631" response mechanism. Twenty flood events from the Xiangshui River Basin were divided chronologically into 14 training events and six independent test events. The final network contains four stacked BiLSTM layers and a fully connected output layer; it does not contain convolutional or pooling layers. Against LSTM, GRU and Xin'anjiang baselines, BiLSTM obtained a mean Nash-Sutcliffe efficiency of 0.90 and a mean denormalised RMSE of 11.2 m3/s. All six paired NSE differences favoured BiLSTM, but the small test set precludes confirmatory statistical inference; results are therefore reported with event-level differences and leave-one-event-out sensitivity rather than interpreted from the Wilcoxon p-value alone. SHAP marginal attribution identified Qingxi, Yingfang and Zhongcun as the most influential stations and the two most recent input lags as the dominant time steps. A marginal-consistent station-lag diagnostic further localised the strongest first-order contributions to upstream rainfall at t-1 and t-2, while explicitly not being treated as a SHAP interaction estimate. Forecast skill declined from NSE = 0.90 at 1 h to 0.58 at 6 h, indicating that quantitative precipitation forecasts are required for operational 3-6 h prediction. The revised integration framework therefore assigns observation-driven BiLSTM forecasts to the short-lead "1" stage and QPF-informed forecasts to the "3" and "6" stages. The findings support a focused, auditable pathway from data-driven prediction to warning action while defining the limits imposed by event scarcity and missing future rainfall forcing.