Physics‐Guided Transfer Learning for Predicting Dissolved Oxygen in Smart Aquaculture Farms

Dissolved oxygen (DO) is a vital parameter in aquaculture, directly influencing fish health, growth and overall productivity. However, existing DO prediction methods largely rely on complex physical models that demand numerous parameter measurements and high computational cost. This study introduces a hybrid framework that integrates transfer learning (TL) with physics‐guided neural networks (PGNNs) based on long short‐term memory models (PG‐LSTM‐TL), to accurately predict pond DO levels from only 1 month of sparse and irregularly sampled nighttime data. The TL strategy leverages knowledge from a data‐rich tidal estuary and adapts it to a data‐scarce aquaculture setting, whereas the physics‐guided component embeds two key domain constraints: The temperature‐dependent DO solubility governed by Henry's law and the monotonic nighttime DO depletion trend. Unlike conventional TL approaches that primarily rely on fine‐tuning or statistical feature alignment, we embed physics‐guided constraints directly into the model learning stage. This integration mitigates negative transfer arising from environmental heterogeneity while preserving physically plausible multi‐step DO trajectories. Results show that, for each daily cycle (9:00 p.m.–5:00 a.m.), the model uses just the first four hourly measurements to recursively forecast DO until dawn, achieving an overall mean absolute error (MAE) of 0.3593 mg/L and root mean square error (RMSE) of 0.4813 mg/L. This represents improvements of 35.5% (MAE) and 32.4% (RMSE) over conventional models. By combining physical insights with data efficiency, the proposed approach offers a practical and scalable solution for reliable DO prediction in small‐scale aquaculture systems worldwide.