A dissolved oxygen prediction model integrating ChatGPT expert knowledge-driven attention mechanism and normalization

Dissolved oxygen (DO) is a key indicator of surface water environmental conditions. However, methods for predicting DO concentrations based on historical data lack domain knowledge guidance, resulting in insufficient prediction accuracy. Therefore, this paper proposes a new DO prediction model (ChatGPT-EK-TabNet). First, an improved normalization method incorporating water quality domain knowledge is proposed. The maximum and minimum values of ChatGPT domain expert knowledge are used to replace sample extremes for normalization, mitigating the interference of outliers and inter-basin distribution differences on the scale. Second, a weighted feature constraint matrix based on ChatGPT domain expert knowledge is proposed and incorporated into the Scaled Dot Product Attention Mechanism to suppress unreasonable feature interactions in terms of knowledge. Finally, we improve the TabNet model while maintaining its sparse adaptive feature selection. Experimental results show that under the Miyun Reservoir data in Beijing, the model performance is RMSE=0.3349, MAE=0.2101, and R²=0.9388. Additionally, compared to the DO prediction model without incorporating water quality domain knowledge, the proposed model reduces RMSE and MAE by 0.1222 and 0.2187, respectively, and improves R² by 0.1120. Generalization experiments demonstrate that the proposed model achieves excellent prediction performance on the Xiaoxingkai Lake, Ussuri River, and Muling River datasets. Reliability validation further indicates that ChatGPT-derived domain expert knowledge achieves higher numerical accuracy than data-driven correlation matrices. Its fine-grained representation and cross-domain generalization match or even exceed those of theoretical expert methods. The proposed approach supports sustainable management of surface water environments and can be extended to other water bodies and water-quality indicators.