A Transformer-Mamba framework for cross-regional long-term forecasting of key water quality in aquaculture

High-precision multistep prediction of water-quality parameters is crucial for aquaculture. However, real water-quality sequences generally exhibit strong nonstationarity and noise, multiscale process aliasing, sparse abrupt events, and time-varying intervariate coupling, making it difficult for long-step multivariate-to-multivariate predictions to balance long-term structure preservation, sensitivity to key segments, and coherence. To address this, this study developed an end-to-end long-term multivariate water-quality prediction framework: First, a dual-branch bidirectional multiscale state-space encoder was designed to build a Mamba-based bidirectional multiscale SSM backbone, thereby capturing long-range dependencies and preserving multiscale structural information efficiently. Second, block-token two-stage sparse selection attention was introduced to enhance attention allocation to sparse abrupt changes and inflection point segments, suppressing noise propagation. Finally, a correlation-prior guided dynamic variable graph fusion head was constructed to introduce controllable intervariate dynamic gating and prior constraints, mitigating coupling drift and reducing error propagation. On seven real-world water-quality monitoring datasets, the proposed method was evaluated and compared with seven representative baseline models using mean absolute error (MAE), root mean square error (RMSE), coefficient of determination (R²), and Kling-Gupta Efficiency (KGE). The proposed method demonstrated stability across multiple objective variables. Compared with those of the overall baseline level, the MAE/RMSE decreased by an average of 44.24%/43.53%, respectively, and the R²/KGE increased by an average of 14.29%/7.67%, respectively. These results indicated that the framework could improve prediction accuracy and consistency under complex, dynamic, and time-varying coupled conditions, providing reliable support for water-quality early warning and refined management.