An AISSA-Optimized Transformer Model for Sediment Concentration Prediction in Multi-Sediment Reservoirs

Accurately predicting reservoir sediment concentration under abrupt fluctuations and frequent extreme peaks remains a major challenge. To address this, we develop an adaptive Improved Sparrow Search Algorithm–Transformer (AISSA-Transformer) for multi-sediment prediction. A distribution-sensitive adaptive loss function reweights errors based on the sample distribution, strengthening model responsiveness to extreme sediment events. Meanwhile, an enhanced sparrow search algorithm—integrating chaotic initialization, dynamic weighting, and multi-distribution perturbation—performs global hyperparameter optimization and stabilizes convergence. Using long-term observations (1998–2024) from the Xiaolangdi Reservoir, the proposed model reduces MAE and RMSE by approximately 42% and 34%, and improves NSE and R by 23% and 33%, respectively. During the 2024 water–sediment regulation period, it accurately captures peak magnitudes and rapid transitions, reducing peak errors by about 42%. These results demonstrate that the adaptive loss effectively enhances extreme-event learning, while the improved optimization strategy mitigates local convergence traps and strengthens robustness. The AISSA-Transformer provides a reliable tool for real-time sediment monitoring and decision-making in reservoir operations.