Abstract: Aiming at the problem of complex intensive aquaculture environment and difficult on-line detection of ammonia nitrogen concentration, a hybrid modeling method based on generalized additive model and support vector regression (SVR) is proposed. First, according to the mechanism analysis, water temperature, dissolved oxygen, pH and conductivity are selected as auxiliary variables. Second, the data is analyzed to determine the distribution of each variable, and  log function is used as the connection function to establish a generalized additive model of anomia nitrogen concentration. In order to improve the modeling accuracy, it is combined with BP neural network, random configuration network and SVR modeling methods, the hybrid soft measurement of ammonia nitrogen concentration in aquaculture water is carried out, and the results are compared. Results show that the generalized additive model can quickly find the relationship between ammonia nitrogen concentration and auxiliary variables and has higher interpretability. The hybrid model based on SVR compensation has higher precision, and the change trend of ammonia nitrogen concentration can better track the changes in laboratory test values. Compared to using the generalized additive model alone, the root mean square error of the hybrid model is reduced by 0.013; compared to using SVR alone, the root mean square error of the hybrid model is reduced by 0.005. The experimental results prove the effectiveness of the hybrid model, which provides a new method for the detection of ammonia nitrogen concentration in aquaculture.

Key words: Aquaculture water, ammonia nitrogen, soft measurement, generalized additive mode, support vector regression

摘要： 针对集约化水产养殖环境复杂、氨氮质量浓度在线检测困难的问题，提出一种基于广义可加模型和支持向量回归的混合建模方法。首先根据机理分析，选择水温、溶氧、pH、电导率作为辅助变量，然后分析数据确定各个变量的分布形式，以对数函数作为连接函数，建立氨氮质量浓度广义可加模型。为提高建模精度，将其与BP神经网络、随机配置网络、支持向量回归建模方法相结合，分别对养殖水质氨氮质量浓度进行混合软测量，并对结果进行比较。结果显示，广义可加模型能较为迅速地找出氨氮质量浓度和辅助变量之间的关系，有更高的可解释性，基于支持向量回归补偿后的混合模型有更高的精度，氨氮质量浓度的变化趋势能更好地跟踪实验室化验值的变化。与单独使用广义可加模型相比，混合模型的均方根误差降低了0.013；与单独使用支持向量回归相比，混合模型的均方根误差降低了0.005。试验结果证明了广义可加模型和支持向量回归混合模型的有效性，为水产养殖氨氮浓度检测提供了一种新的方法。

关键词: 养殖水质, 氨氮, 软测量, 广义可加模型, 支持向量回归