Integrated Heterogeneous Fusion Transformer Model for Marine Dissolved Oxygen Prediction

Predicting dissolved oxygen (DO) levels in marine environments is crucial for maintaining the health of marine ecosystems and ensuring the effective management of marine ranches. However, DO data often exhibit non-stationary characteristics and significant short-term fluctuations, complicating feature extraction and rendering traditional time series forecasting methods ineffective in addressing these challenges. To this end, this study introduces an integrated heterogeneous fusion long-sequence forecasting model. Initially, the trace of the covariance matrix is utilized as a weighting factor to preprocess the raw DO data, thereby extracting high-value DO sequences and mitigating the influence of noise and outliers. Subsequently, a time-warped attention mechanism combined with a historical time delay penalty term is incorporated to enhance the model's ability to capture contextual correlations within the time series and to improve overall robustness. Experimental validation conducted across twelve marine ranches on the Shandong Peninsula demonstrates that the proposed model outperforms existing methods across various forecasting horizons, achieving predictions up to 48 hours in advance, and exhibits exceptional generalization and stability. This model offers effective technical support for marine ecological monitoring and management, presenting broad prospects for practical applications.