Long-term water quality model in aquaculture based on dilated convolution feature fusion and interactive learning

Abstract

The prediction of water quality parameters is of great significance for maintaining water stability and controlling marine environments. However, due to the temporal nature of water quality parameters and the influence of surrounding changeable environments, water quality is characterized by complex nonlinearity and instability. In current prediction research, there are many training parameters and poor generalization ability, and its prediction accuracy cannot be guaranteed. Thus, this paper proposes a network model called Feature Fusion Interactive learning network (FFINet) based on dilated convolution feature fusion and interactive learning. FFINet is composed of the SCFINet module and TCINet module in parallel, and the two modules accept the same original sequence. The SCFINet module captures the local features of the sequence through feature fusion and proposes interactive learning to extract the trend, periodicity and irregularity information of the subsequence. The TCINet module uses expansion convolution to capture the long-term dependencies of sequences and uses interactive learning to compensate for information loss during separation, can process sequences simultaneously in parallel mode, and can retain irregular information of sequences. The model adds a dual residual connection to weaken the problem of model gradient vanishing, and each layer outputs sequences of the same length to be added as input to the lower layer. Dissolved oxygen concentration experiments are conducted at multiple marine ranches, and the results of the proposed model are compared with those of other deep learning models, showing that FFINet has better prediction performance than existing approaches. This paper provides important theoretical guidance for sustainable aquaculture development.