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

Area of Science:

• Environmental Science
• Water Quality Monitoring
• Artificial Intelligence in Environmental Management

Background:

• Dissolved oxygen (DO) is crucial for assessing surface water health.
• Existing DO prediction models lack domain knowledge, leading to suboptimal accuracy.
• Outliers and distribution differences in historical data hinder reliable predictions.

Purpose of the Study:

• To develop an accurate DO prediction model by incorporating water quality domain knowledge.
• To mitigate the impact of outliers and inter-basin differences on DO concentration predictions.
• To enhance feature interaction analysis for more reliable water quality modeling.

Main Methods:

• Proposed an improved normalization method using domain expert knowledge to handle outliers.
• Developed a weighted feature constraint matrix with domain knowledge for attention mechanisms.
• Enhanced the TabNet model for sparse adaptive feature selection in DO prediction.

Main Results:

• The ChatGPT-EK-TabNet model achieved high accuracy (RMSE=0.3349, MAE=0.2101, R²=0.9388) on Miyun Reservoir data.
• The proposed model outperformed traditional methods, reducing RMSE by 0.1222 and MAE by 0.2187.
• Demonstrated excellent generalization across multiple datasets (Xiaoxingkai Lake, Ussuri River, Muling River).

Conclusions:

• ChatGPT-derived domain knowledge enhances DO prediction accuracy beyond data-driven approaches.
• The model offers fine-grained representation and cross-domain generalization capabilities.
• This approach supports sustainable surface water management and can be applied to other water quality indicators.