基于模糊-MPC控制的集约化养殖精准投饲方法

Precision Feeding Method for Intensive Aquaculture Based on Fuzzy-MPC Control

摘要:

为提高集约化鱼类养殖中投饲的精准性与智能化水平,解决因投饲不精准导致的生长缓慢、鱼病频发和饲料浪费问题,本研究提出一种融合模糊控制与模型预测控制(Model predictive control, MPC)的双层智能投饲策略。 系统通过水下摄像头与多参数水质传感器实时采集鱼类行为图像及水质数据,输入基于多任务学习与注意力机制的多任务学习长短时序预测软注意力机制 (Multi-task learning-short and long-term sequence prediction-soft attention mechanism,MTL-LSTM-SAT)模型,预测未来 7 d 鱼体生长速度与患病比例。 下层模糊控制器根据当前状态快速响应,输出初步投饲建议;上层 MPC 控制器以饲料利用最大化和疾病风险最小化为目标,进行多目标滚动优化,输出最终投饲指令。 仿真结果表明,该策略在生长速度、患病控制与饲料转化效率 3 方面均衡能力优于比例积分微分控制 (Proportional-integral-derivative, PID)、 模糊控制、 MPC 及线性二次型调节器 (Linear quadratic regulator,LQR)方法。 实际养殖试验中,该策略使饲料系数 (Feed conversion rate, FCR)降至 1. 23, 特定生长率 (Specific growth rate,SGR)达 1. 11% / d,存活率提升至 94. 21% ,显著优于人工经验组与定时自动投饲组,验证了其在实际生产中的有效性与优越性。

Abstract:

Aiming to enhance the precision and intelligence of feeding in intensive fish farming systems while addressing feed waste and frequent fish diseases caused by improper feeding practices, a dual-layer intelligent feeding strategy integrating fuzzy control and model predictive control (MPC)was proposed. The system collected real-time fish behavior data and water quality parameters through underwater cameras and multi-parameter sensors, inputting this information into a multi-task learning and attention mechanism-based MTL-LSTM-SAT model to predict fish growth rates and disease incidence over the next seven days. The lower-level fuzzy controller provided initial feeding recommendations based on real- time conditions, while the upper-level MPC controller performs multi-objective rolling optimization to maximize feed conversion efficiency and minimize disease risks, ultimately issuing final feeding instructions. Simulation results demonstrated that this strategy outperformed PID control, fuzzy control, MPC, and linear quadratic regulator (LQR)methods in balancing growth rate, disease control, and feed conversion efficiency. Field trials showed the strategy reduced feed conversion rate (FCR)to 1. 23, achieved a specific growth rate (SGR)of 1. 11% per day, and increased survival rates to 94. 21% , significantly outperforming manual experience-based groups and scheduled feeding groups, validating its effectiveness and superiority in practical applications.