A Context-Aware Temporal Convolutional Network for Water Replacement Prediction in Catfish Biofloc Ponds With Imbalanced Event Handling

Abstract

Water replacement is a biologically critical yet under-automated decision in biofloc-based aquaculture systems. Mistimed actions can destabilize microbial ecosystems, elevate fish mortality, and compromise sustainability through excessive water usage. Traditional rule-based heuristics often fail to account for the nonlinear and multiscale dynamics of pond environments. To address this, we propose a Context-Aware Temporal Convolutional Network (CA-TCN), a rare-event classification framework that combines deep temporal modeling with aquaculture-specific logic. The CA-TCN combines a dilated Temporal Convolutional Network with biologically guided SMOTE+Tomek resampling, Focal Loss for imbalance-sensitive learning, ROC-based threshold calibration, and a rule-based override system for decision assurance. Trained on 213 real-world multivariate time-series samples, each consisting of 23 features across 5 sequential timesteps, representing sensor data for water quality (total dissolved solids, pH, dissolved oxygen, electrical conductivity), feeding events, and fish mortality. The proposed model achieved 98.68% accuracy, 1.0000 precision, 0.9737 recall, an F1-score of 0.9867, and a ROC-AUC of 1.0000 on the held-out test set, demonstrating its ability to identify rare yet operationally critical water replacement events with high precision. Ablation studies reveal the cumulative contributions of each component: +2.6% F1-score improvement from context-aware sampling, +1.3% gain from Focal Loss, a 2.56% reduction in false positives via threshold calibration, and a 0.9% recall increase due to rule-based override. Compared to state-of-the-art baselines, CA-TCN outperforms SMOTE only (F1 = 0.9600), SMOTE+ENN (F1 = 0.9610), and Tomek only (F1 = 0.2963), offering up to +69.04% F1-score improvement and eliminating all false negatives, a critical requirement in early warning systems for aquaculture risk mitigation. This work contributes a validated, domain-informed artificial intelligence pipeline that advances sustainable aquaculture management.