Coastal water quality monitoring is critical for ecosystem health and regulatory compliance under Vietnam’s QCVN 10:2023/BTMT standards. Thus, this study addressed the limitations of traditional predicting models by proposing a hybrid framework that integrate Hidden Markov Models (HMM, dual-branch Long Short-Term Memory (LSTM) networks, and attention mechanisms for multi class water quality classification. The model was developed using 1,305 daily observations from 2020-2023 of five water parameters – TSS, pH, TPH, Total Coliform, and DO – at the Sao Den-Ben Dinh station positioned in the east coast of Ho Chi Minh City. The hybrid architecture achieved a moderate accuracy of 74.87% and F1-score of 64.11%, representing a statistically significant improvement over a standalone HMM (44%), LSTM (68.2%), and ARIMA (70,77%). Attention weights identified a one-to-three-day lead time critical for predicting pollutions events. However, a significant limitation remains in a minority class detection, where the “Poor” category reached only 14% recall due to a severe 7:5:1 data imbalance. Generally, this framework provides an interpretable, regulatory-aligned decision-support tool for early-warning systems, offering a theoretically principled approach to capturing multi-scale temporal dynamics in coastal environments.