Multi-scale decomposition and adaptive feature selection coupled heterogeneous model fusion for reliable dissolved oxygen prediction in complex aquatic environments

Accurate dissolved oxygen (DO) prediction is essential for water quality management and ecological risk assessment. However, DO series are characterized by strong nonlinearity, nonstationarity, and heterogeneous temporal dynamics, which pose significant challenges to accurate prediction. To address these issues, this study proposes a novel forecasting framework termed SAXSG, integrating Seasonal-Trend decomposition using Loess (STL), Adaptive Multi-Correlation Feature Selection (AMCFS), and heterogeneous model fusion. STL is first employed to decompose the original DO series into multiple temporal components. AMCFS is then used to identify informative external variables and optimal lag structures through adaptive weighting of multiple correlation measures. Subsequently, guided by component-wise dynamic-characteristic analysis and empirically verified through component-level comparison, XGBoost, support vector regression (SVR), and GRU-MLP are assigned to the trend/seasonal, middle-frequency, and high-frequency components, respectively; the final prediction is then obtained through component reconstruction. The proposed framework was evaluated using long-term water quality and meteorological observations from the Taihu Lake watershed and compared with eight benchmark models. Results show that SAXSG achieved the best overall performance, yielding the highest R 2 of 0.905 and the lowest RMSE of 0.535 on the testing dataset. Compared with RF, TCN, and CNN-BiLSTM-AM, the proposed framework reduced RMSE by 29.0%, 8.4%, and 16.1%, respectively. Furthermore, the proposed framework demonstrated superior performance in multi-step forecasting, extreme-event prediction, uncertainty quantification, and robustness analysis. These results indicate that SAXSG provides an effective and reliable solution for DO forecasting in complex aquatic environments.