Application of PatchTST Model in Time Series Prediction of Dissolved Oxygen in Stratified Coastal Waters of Hong Kong

Abstract Dissolved oxygen (DO) is a critical water quality indicator for assessing the health of marine ecosystems, and its accurate prediction is of great significance for environmental protection and risk early-warning in coastal waters. Based on 20 years of monitoring data from Hong Kong waters, this study constructs a DO time series prediction model using PatchTST and, for the first time, proposes an independent modeling strategy stratified by water depth (middle layer / bottom layer) to enhance predictive adaptability and model generalization capability. The results show that at the representative station SM3, the PatchTST model achieves R² = 0.77 and RMSE = 0.85 in middle-layer prediction, and R² = 0.80 in bottom-layer prediction. Overall performance improves by approximately 20%–40% compared to traditional machine learning models, with greater stability observed under the stratified modeling framework. Subsequently, model interpretation via Permutation Importance reveals that water temperature, salinity, historical DO, and seasonal features are the core drivers of DO dynamics, with bottom-layer water showing significantly stronger dependence on historical DO. This study validates the cross-spatial prediction capability of PatchTST across multiple water layers and sites, and proposes a DO prediction framework integrating stratified modeling and feature interpretability, providing technical reference for intelligent monitoring and ecological early-warning in coastal waters.