Fusion of transformer-based deep learning and Monte-Carlo fish growth simulation for aquaculture smart transformation

作者：Hsun-Yu Lan, Naomi Ubiña, Kai-Xiang Zhanga, Shyi-Chyi Cheng, Shih-Yu Li · 发表于：Engineering computations · 年份：2025 · DOI：10.1108/ec-07-2024-0599 · 被引用次数：2

Develop a deep learning-based fish growth model to improve the accuracy of fish growth predictions and optimize feeding strategies in open-sea aquaculture cages. We employed the Monte Carlo approach to generate big data for training transformer-based deep learning models to predict fish growth trajectories during cultivation. In generating big data, each key factor of the fish growth model is modeled with a probability distribution parametrized by real-world fish growth data from IoT-based monitoring systems and open weather datasets. Minimal prediction errors from 2.02 to 3.01% for weight; growth rate errors consistently below 2.5% and feeding amount and the meat conversion rate exhibit slightly higher but still acceptable error margins (∼7.4–8.0 and ∼5.1–5.4%, respectively). Dependency on accurate initial parametrization of the probability distributions and the reliability of data collected from IoT systems or other dataset sources. The model should be robust against varying environmental conditions, and it has limitations in application to different types of aquaculture environments. First, the techniques can enhance precision in aquaculture by providing accurate fish growth predictions and optimized feeding strategies. Second, reducing feed consumption not only lowers production costs but also minimizes the environmental impact of excessive feed waste. Lastly, the use of IoT-based monitoring systems and smart feeding machines can streamline...