A hybrid CNN-XGBoost-Temporal Fusion Transformer framework for explainable multi-variable hydroclimatic extremes forecasting under CMIP6 climate scenarios
Abstract
Hydroclimatic extremes are projected to intensify across monsoon-dependent river basins, yet reliable forecasting remains limited by three persistent gaps: deep learning models rarely integrate spatial and temporal processing in a unified pipeline, almost always target a single variable, and offer limited interpretability. This study introduces CXTF-Net, to the best of our knowledge the first three-stage hybrid framework for explainable, multi-variable, multi-horizon hydroclimatic forecasting. It combines a CNN-BiLSTM encoder for joint spatial-temporal feature extraction, an XGBoost feature-selection gate that reduces 28 candidate predictors to the most informative 12, and a Temporal Fusion Transformer (TFT) decoder that simultaneously produces calibrated probabilistic forecasts of streamflow, the 12-month Standardised Precipitation Index (SPI-12), a compound flood risk index, and basin water deficit. Applied to a 43-year multi-source record (1980–2022) for the Mahanadi River Basin, India, CXTF-Net achieved an NSE of 0.934, a KGE of 0.921, and an RMSE of 18.4 m 3 /s for 1-day-ahead streamflow forecasting, outperforming seven benchmarks including LSTM (NSE = 0.852), Transformer (NSE = 0.889), and CNN-LSTM (NSE = 0.901) at horizons up to 30 days. NSE exceeded 0.79 at the 30-day horizon across all four target variables. SHAP attribution and attention rollout jointly identified SST anomalies, ENSO Niño 3.4, and antecedent soil moisture as the three dominant hydroclimatic drivers, with physically consistent horizon-dependent transitions from land-surface to ocean-forced predictability. CMIP6 SSP5-8.5 projections indicate a statistically significant decline in mean annual basin rainfall ( − 0.72 mm yr − 1 ) and a near-tripling in severe drought frequency by 2080–2100. CXTF-Net provides an interpretable, transferable framework for climate-risk decision-making in monsoon-driven basins.