Multivariate multi-horizon fish growth forecasting using temporal convolutional networks in IoT-enabled aquaculture

Abstract

• A multitask TCN framework forecasts fish length and weight from multivariate IoT water-quality data. • Multi-resolution resampling (1–30 min) captures short-, medium- and long-term dynamics. • TCN outperforms LSTM/GRU by 30–50% across all ponds, horizons, and temporal scales. • Plateau-aware loss enhances smooth, biologically consistent multi-step growth trajectories. • The end-to-end IoT–TCN pipeline supports real-time, data-driven decision making for smart aquaponics. Aquaculture is increasingly dependent on intelligent monitoring systems capable of capturing dynamic environmental–biological interactions to support sustainable, data-driven fish production. Existing fish growth prediction approaches, however, treat length and weight as isolated static regression targets and largely overlook the temporal dependencies inherent in real aquaculture environments. This paper presents a novel multivariate, multitask, multi-horizon deep learning framework for real-time forecasting of fish length and weight using continuous sensor data. We introduce the first Temporal Convolutional Network (TCN)-based architecture for aquaculture forecasting and benchmark it extensively against state-of-the-art recurrent models, including Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), across multiple temporal resolutions (1-minute, 15-minute, and 30-minute) and forecasting horizons (short-term, medium-term, and long-term). The proposed model incorporates a multi-resolution temporal resampling pipeline, a sliding-window supervised formulation, and a shared TCN encoder with task-specific heads, enabling simultaneous prediction of multiple biological traits across future time steps. A plateau-aware joint loss function is proposed to enforce temporal smoothness and produce biologically consistent growth trajectories, while Bayesian hyperparameter optimization is used to optimize models. Experiments conducted on real IoT datasets collected from multiple independent fish ponds demonstrate that the TCN consistently outperforms LSTM and GRU by 30–50% in Root Mean Square Error and Mean Absolute Error, with superior stability across all resampling scales and forecasting windows. The model also exhibits strong cross-pond generalization, confirming its robustness to variable environmental conditions. These results demonstrate that TCN-based multitask forecasting effectively models complex, multiscale relationships between water quality and biological growth.