A decomposition-assisted deep ensemble framework for interpretable total nitrogen forecasting across multiple reservoir systems

Abstract

Forecasting total nitrogen (TN) across multiple reservoirs remains challenging because nutrient dynamics are nonlinear, non-stationary, and influenced by interacting hydrological, climatic, and biogeochemical processes. This study proposes a CEEMDAN-enhanced Deep Hybrid Ensemble Framework (CEEMDAN-DHEF) for monthly TN forecasting across eight reservoirs in Central China. Using a 2,304-record monthly reservoir panel spanning 2000–2023, the framework integrates water-quality variables, meteorological drivers, temporally lagged predictors, interaction terms, and CEEMDAN-derived multiscale descriptors within 12-month forecasting sequences. Five complementary neural learners a TFT-inspired temporal-attention network, Long Short-Term Memory, Convolutional Neural Network, Gated Recurrent Unit, and Deep Neural Network were trained on 12-month sequences, with their outputs integrated through constrained convex weighting and Ridge stacking calibrated exclusively on chronologically ordered development data. Under the independent chronological evaluation (November 2019–December 2023; n = 400), CEEMDAN-DHEF achieved R²/NSE = 0.3448, RMSE = 0.658 mg/L, and MAE = 0.481 mg/L, providing a conservative assessment of temporal generalization. A separate retrospective later-period evaluation yielded R²/NSE = 0.8345, indicating stronger predictive agreement under that evaluation setting. Complementary leakage-controlled analyses showed that lagged predictors accounted for 42.1% of the total SHAP importance, followed by observed-history variables (20.2%), rolling predictors (10.4%), and CEEMDAN-derived multiscale features (9.5%). Prediction-interval analysis further quantified forecast uncertainty, with empirical coverages of 92.25% and 98.50% for the nominal 90% and 95% intervals, respectively. These findings demonstrate that decomposition-assisted ensemble learning can capture multi-scale TN dynamics while providing interpretable and uncertainty-aware evidence for reservoir water-quality forecasting.