Research on a dissolved oxygen prediction method for recirculating aquaculture systems based on a convolution neural network

Dissolved oxygen is the most critical parameter to be controlled in Recirculating Aquaculture Systems strictly to maintain healthy conditions for aquatic products. Because of the lag between dissolved oxygen control measures and the regulation effect, changes in the dissolved oxygen must be forecast to maintain stable water quality. Traditional methods, such as back propagation (BP) neural networks and time-series analyses, have poor stability and dynamic responses and thus present difficulties meeting the real-time dynamic regulation needs of industrial aquaculture. Therefore, a simplified reverse understanding convolutional neural network (CNN) prediction model is proposed in this study to solve the dissolved oxygen prediction problem. The model multiplies the input vector by its transpose to format a single depth input matrix. By removing the pooling layer, the characteristics of the relational factors of dissolved oxygen are refined by two successive convolutions of the input matrix. Finally, the data are processed by the full connection layer, which uses the gradient descent algorithm for the reverse update. Real-time data obtained from the Mingbo Experimental Base in Shandong Province are analyzed, and the results show that the reverse understanding CNN is suitable for the prediction of dissolved oxygen. Moreover, its convergence rate during pre-training is faster than that of the BP network under the same conditions, and its prediction stability is superior. The accuracy and stability of the new model results are sufficient to meet actual production demands.