Dissolved oxygen prediction method for recirculating aquaculture system, based on a timing attenuation matrix and a convolutional neural network

摘要

Dissolved oxygen (DO) plays a crucial role in recirculating aquaculture systems. The DO concentration determines the biological growth of aquatic organisms and the final economic benefits. Considering the hysteresis effect of DO increments in recirculating aquaculture, as well as the complex relationships among influencing parameters, this paper describes a method of DO prediction, based on a longitudinal timing attenuation matrix and convolution to extract the horizontal interactions of parameters. A longitudinal relationship matrix of DO and DO related parameters is obtained using the concept of time series attenuation, followed by the refinement of the matrix, which describes the parameter's potential relationships, through two successive convolutions. The relationship factors are inputted into a fully connected network for processing. Compared to the traditional back-propagation (BP) full-connection network and to the matrix convolution neural network, experimental results show that the method, described herein, has better accuracy and model stability, while requires fewer training epochs under the same requirements.