Transformer-Based Modeling on Multi-Target Water Quality Forecasting in Aquaponic Fish Ponds

Aquaponic systems require reliable forecasting of water quality parameters to support proactive monitoring and maintain suitable conditions for aquatic organisms. This study investigates Transformer-based deep learning for multi-step, multi-target water quality forecasting using multivariate sensor measurements from aquaponic fish ponds. A leakage-safe time-series framework is developed in which historical observations are used to simultaneously forecast pH, temperature, and dissolved oxygen over multiple future time steps. The Transformer is systematically compared with CNN, LSTM, GRU, and their attention-enhanced variants under identical experimental conditions. Among the learned models, the Transformer achieves the best overall performance, attaining an MAE of 1.3908, RMSE of 2.1103, R2 of 0.7534, and sMAPE of 21.09%. Statistical analysis further demonstrates significant improvements over several competing architectures, while horizon-based evaluation investigates forecasting behavior as the prediction range increases. The Transformer remains computationally lightweight, requiring approximately 0.91 MB and achieving a mean inference latency of 14.55 ms, supporting real-time deployment. Cross-pond validation and controlled noise experiments further assess generalization and robustness under challenging conditions. The findings demonstrate the potential of Transformer-based temporal modeling for efficient and proactive water quality forecasting in intelligent aquaponic monitoring systems.