集约化水产养殖关键参数稳定学习方法研究

The concentration of water temperature and dissolved oxygen are the most concerned parameters in the process of intensive aquaculture, they affect the normal growth of fish directly. According to the heat balance principle, a temperature control mechanism model is established. It is used to resolve the current situation that a simple estimation or rule to achieve a rough control. Considering the uncertainties and difficulties in modeling, the hybrid modeling technology is carried out in intensive aquaculture process. It combines knowledge and data to establish the dissolved oxygen, pond aeration and feeding mechanism model and error compensation model to solve a single neural network prediction model which can’t be explained and have poor precision. For the practical application of the process, the neural network learning algorithm will generate long-running divergence and instability. In order to slove this problem, a model correction mechanism is proposed which is called ellipsoid bounded algorithm. The ellipsoid bounded stable learning parameter correction algorithm is based on ellipsoid intersection principle. It can be used to update the model parameters and ensure the boundedness of modeling errors. The launching of this project is not only to lay the foundation for the control and optimization of the aquaculture water environment, thereby the stable learning model parameters can be extended to other areas, with a theoretical value.

水体温度和溶解氧浓度是集约化水产养殖过程最关注的两个参数，直接影响鱼类的正常生长。根据热量平衡原理，建立水温控制机理模型，解决目前仅凭经验和简单估计实现粗略控制的现状。针对养殖过程影响因素较多，模型较难建立的现状，引入混合建模技术，结合机理知识和采集到的数据建立溶解氧浓度和增氧及饲料投喂机理模型和误差补偿模型，从而解决单一神经网络预报模型不易解释，精度不高的现状。由于实际应用过程中，神经网络学习算法长时间运行产生发散、不稳定，开展椭球定界稳定学习算法的研究，此算法依据椭球交集原理，不仅可以对模型参数进行更新，还能保证建模误差稳定有界。本项目的开展不仅可以为控制和优化水产养殖水质环境奠定基础，而且模型参数稳定学习可以推广应用于其他领域，具有一定的理论价值。

随着水产养殖工厂化和精养化程度的提高，及时掌握水质的动态变化，提前预测水质情况是工厂化养殖亟待解决的重要问题。本项目针对集约化水产养殖过程中关键参数难以在线检测，无法实现实时控制和优化的问题，采用混合建模技术和基于稳定学习的模型参数更新算法，解决了单一神经网络预报模型不易解释，精度不高的现状，为控制水温、增氧和饲料投喂奠定了基础，结合水温控制和增氧及饲料投喂控制策略，达到了精准控制养殖生产过程的目的。在溶解氧浓度预测方面，提出了基于偏最小二乘和神经网络的建模方法；在氨氮浓度软测量方面，提出了基于遗传算法优化的支持向量回归模型；模型参数更新方面，结合机理知识，开展了随机配置网络稳定学习算法研究，利用椭球交集理论，重新构建随机配置网络的模型形式，构造递推辨识算法，找到最优的模型参数。将提出的方法在实验室集约化水产养殖系统中进行了实验验证。本项目的研究不仅能够推动软测量技术在水产养殖学科的发展，而且具有广阔的工业实际应用前景。