Multi-parameter prediction of seawater quality based on dynamic spatio-temporal relationship network

Seawater quality parameters serve as vital indicators reflecting the health of marine environments. Their dynamic monitoring and accurate prediction are crucial for supporting the sustainable development of marine ecosystems. However, traditional prediction models struggle to effectively capture spatio-temporal information across multiple buoys, exhibiting significant limitations in extracting and utilizing unstructured spatio-temporal relationship data. To address these challenges, this paper proposes a multi-parameter seawater quality prediction model based on a Dynamic Spatio-temporal Relationship Network (DSTRN). Building upon graph neural network concepts, this model employs an adaptive adjacency matrix to dynamically aggregate information from neighboring nodes. This enables refined modeling of spatio-temporal correlations at each time step, thereby better reflecting the influence of seawater quality parameters across varying temporal and spatial scales. Experimental results using multi-buoy monitoring data from the Beibu Gulf demonstrate that the adopted multi-buoy collaborative prediction strategy outperforms single-buoy prediction strategies across all four buoy stations and prediction time horizons. Compared with other advanced models, the DSTRN model also achieved the best prediction accuracy. Taking buoy 14 as an example, its average R² value across all prediction durations reached 0.978, with a maximum deviation of only 1.7 %, further confirming the model's robust performance across different prediction time horizons. In summary, this study confirms that the DSTRN model effectively captures and utilizes the spatio-temporal correlation information of seawater quality parameters, demonstrating strong spatio-temporal collaborative predictive capabilities.