GHDFloodNet: An Advanced Model for Improved Short-Term Flood Forecasting

Accurate short-term flood forecasting is vital for effective risk management and early warning systems. However, many data-driven models struggle to generalise with limited historical data and fail to consistently capture complex temporal dependencies across varying forecasting horizons. To address these challenges, this study proposes GHDFloodNet (Generalised Hybrid Data-limited Flood Prediction Network), a hybrid deep learning framework designed for robust multi-step-ahead forecasting. GHDFloodNet integrates First-Order Model-Agnostic Meta-Learning (FOMAML) with a Temporal Fusion Transformer (TFT) to enable rapid task adaptation and effectively capture long-range temporal dependencies and variable interactions. To further enhance predictive consistency, the framework incorporates a bidirectional Long Short-Term Memory (BiLSTM) network augmented with an additive attention mechanism and static feature fusion as a core learner within a meta-ensemble architecture. Bayesian hyperparameter optimisation within an AutoML framework identifies optimal model configurations, while a dedicated data handling layer with real-time augmentation improves stability under non-stationary conditions. The framework was evaluated for multi-horizon water level forecasting across four lead time ranges (1–6 h, 6–12 h, 12–24 h, and 24–48 h) using rainfall and lagged water level observations as primary inputs. Experimental results demonstrate that GHDFloodNet achieves robust, nearly invariant error distributions across the full 1–48 h forecast window, reporting an MSE of 0.53–0.55, RMSE of 0.72–0.74, and MAE of 0.35–0.36. Furthermore, the model exhibits stable goodness-of-fit, with R2 and NSE values consistently ranging from 0.44 to 0.47 across all lead times, significantly outperforming conventional baselines, which typically exhibit pronounced error escalation at longer horizons. Overall, GHDFloodNet demonstrates that horizon-independent forecast reliability can be architecturally engineered, offering critical value for operational flood forecasting where consistent performance across all lead times outweighs peak short-range precision.