A novel time-aware attention-enhanced transformer to prediction harmful algal Bloom's occurrence

Harmful Algal Blooms (HABs) present growing threats to marine ecosystems and coastal economies worldwide, yet accurate prediction remains challenging due to complex multi-scale temporal dynamics and severe class imbalance in monitoring data. Current models often fail to capture both short-term fluctuations and long-term seasonal patterns while maintaining performance under realistic data constraints. To address these limitations, we propose HRED-TIDE (Hierarchical Recurrent Encoder-Decoder with Time-Informed Dynamic Embeddings), a novel Transformer-based architecture that integrates three specialized attention mechanisms comprising temporal decay, periodicity-aware, and relative-time attention for modeling key environmental patterns across different timescales. Evaluated on monitoring datasets from multiple major HAB-prone regions in China and Florida's Caloosahatchee River, our model achieved 95.17% accuracy in predicting bloom occurrence within 24 h and maintained a high F1-score of 0.7498 despite extreme class imbalance (only 6.4% positive samples), significantly outperforming all baseline methods. The model's attention mechanisms provide interpretable insights into key environmental drivers, offering a robust, scalable solution for operational HAB forecasting systems that balances predictive performance with computational efficiency.