神经网络在养殖水质精准预测方面的研究进展

目前神经网络研究文献成果较多，虽然在水质精准预测方面起到了一定的参考，但由于文献缺少科学分类，使用率不高，导致学者难以找到研究切入点。针对这一问题，本文将神经网络方法在养殖区水质精准预测方面的文献按照海水和淡水两大领域进行分类，主要对每个领域所应用的预测模型从正反馈架构、循环架构和混合架构三个方向对海水时空序列文献进行分类研究和综述，发现混合架构模型的预测性能优于正反馈模型和循环架构模型，有利于提升不同深度水质预测模型的精度。另外，本文对基于神经网络方法的三维水质预测模型进行了初步探讨，发现学者的研究成果更多地集中在水表层和水中层的不同位置水质参数的变化方面，而神经网络方法对水表层水质预测精度比水中层和水深层水质预测精度高。

Research progress in accurate prediction of aquaculture water quality by neural network

China is the world''s largest producer and consumer of aquatic products, with aquaculture production ranking first in the world for more than 20 consecutive years, and the demand for aquatic products provides opportunities for the development of the global aquaculture industry. In aquaculture, the aquaculture water environment provides the living environment, food and oxygen for freshwater or seawater. Due to human activities, environmental pollution, agricultural production and other reasons, it may lead to changes in total phosphorus, dissolved oxygen, pH and other indicators in aquaculture waters, which in turn affect the growth of aquatic organisms. Therefore, real-time monitoring and prediction of water quality parameters is an important part of the aquaculture process and is an important measure to determine the quality of aquatic products. Through the analysis of the collected information, there are more neural network research results, which play an important role in accurate water quality prediction, but the lack of scientific classification in the literature and the low usage rate of the literature have made it difficult for scholars to find the research entry point. To address this issue, this paper classified the literature on neural networks methods for accurate prediction of farmed water quality according to two major fields: seawater and freshwater, and mainly studied and analyzed the neural network models applied in each field for prediction of seawater spatio-temporal sequences from three architectures: positive feedback architecture, recurrent architecture and hybrid architecture, and the analysis results showed that the highest prediction performance in the positive feedback architecture model is the ANN prediction model with 64% accuracy, and in the recurrent architecture model, the highest prediction performance is the convolutional neural network prediction model with 97.1% accuracy, and in the hybrid architecture model, the highest prediction accuracy is the intelligent algorithm-LSTM -RNNs model with an accuracy of 99.72%, which is 35.72% and 2.62% higher than the highest accuracy in the positive feedback architecture model and the recurrent architecture model, respectively. The prediction performance of the hybrid architecture model is better than those of the positive feedback model and the recurrent architecture model, which is conducive to improving the prediction accuracy of the different depth water quality prediction models. In addition, this paper had a preliminary discussion on the three-dimensional water quality prediction model based on the neural network method, and the results showed that the research scholars results are more focused on the changes of water quality parameters in different locations of the water surface layer and water intermediate layer, while for neural network prediction model for water surface layer, water quality prediction accuracy was higher than intermediate and deep water layer quality prediction accuracy.