基于主成分分析及GA-LM的水产养殖环境溶解氧和氨氮含量预测

Prediction of dissolved oxygen and ammonia nitrogen concentrations in aquaculture environment based on PCA and GA-LM

为了精准预测水产养殖过程中最重要的两个环境参数溶解氧和氨氮,针对预测模型需要解决的有效影响因子确定、预测算法和网络结构优化等问题,将Levenberg-Marquardt(LM)神经网络、遗传算法(genetic algorithm,GA)和主成分分析(PCA)算法相结合,提出一种基于GA-LM-PCA的水产养殖环境溶解氧和氨氮含量预测模型,即采用PCA确定影响因素,实现影响因素的去耦合降维,采用遗传算法对网络结构进行优化,确定合适的隐层节点数目和权值,采用LM训练神经网络,提高神经网络的收敛速度。为了验证GA-LM-PCA的预测效果,将GA-LM-PCA的预测效果与未用PCA方法的GA-LM预测模型进行了试验比较,并探讨了影响因素数量对预测效果的影响。结果表明:用GA-LM-PCA方法预测的溶解氧和氨氮值与实测值吻合较好,平均绝对误差和均方根误差分别为0.0047、1.8727×10^(-4)(溶解氧)和0.0065、9.4287×10^(-4)(氨氮),适用于影响因素数量较多的场合。研究表明,GA-LM-PCA是一种有效的水产养殖环境溶解氧和氨氮预测工具,尤其对于影响因素复杂繁多的非线性系统效果更好。

To accurately predict two of the most important parameters,dissolved oxygen(DO)and ammonia nitrogen concentrations in aquaculture,aiming to solve the problems such as the determination of effective influence factors,prediction algorithm and network structure optimization,a prediction model of DO and ammonia nitrogen concentrations called GA-LM-PCA was proposed by combining the neural network(NN)algorithm of levenberg-marquardt(LM),genetic algorithm(GA)and principal component analysis(PCA).PCA was used to determine the influence factors which were decoupled and reduced in the dimension.The network architecture was optimized by GA for determination of the appropriate number of hidden layer nodes and weight values.LM was applied to train NN to improve the generalization ability and convergence speed.The performance of GA-LM-PCA was compared with that of GA-LM without PCA to verify the forecasting accuracy of the GA-LM-PCA,and the prediction effect of different quantity of influence factors was discussed.The comparison indicated that the predicted DO and ammonia nitrogen values using GA-LM-PCA were in good agreement with the measured data,with the mean absolute errors and the root mean square errors of 0.0047,1.8727×10^(-4)(DO),and 0.0065,9.4287×10^(-4)(ammonia nitrogen),indicating that GA-LM-PCA was more suitable for the occasion with a large number of influence factors.It is proved that the proposed model can be considered as an effective prediction tool for DO and ammonia nitrogen concentrations in aquaculture environment,especially for nonlinear systems with various complex factors.