A hybrid enhanced optimization architecture-based model for long-term water temperature prediction in aquaculture

Abstract

Addressing issues such as inadequate prediction accuracy, significant noise interference, and difficulties in capturing long-term dependencies in aquaculture environment water temperature (WT) forecasting, this paper proposes a long-term prediction model based on a hybrid enhanced optimization architecture, namely the HEOA-BiTCN-MLSTM model. This model utilizes the human evolutionary optimization algorithm (HEOA) to optimize model parameters, thereby improving prediction accuracy. It employs a bidirectional temporal convolutional network (BiTCN) to deeply extract latent features from data, enhance noise resistance, and capture long-term trends and complex nonlinear variations in WT data. Finally, it integrates a multi-head long short-term memory (MLSTM) network, which combines the long short-term memory (LSTM) network with the multi-head self-attention mechanism (MHSA), to improve the model’s ability to capture long-term dependencies and model global information when processing long sequences. This model significantly improves WT prediction performance through the synergistic effects of optimization, hybridization, and enhancement while effectively addressing noise interference and data irregularity in complex environments. Experimental results demonstrate that this model significantly outperforms traditional benchmark models such as the gated recurrent unit (GRU), backpropagation neural network (BPNN), and recurrent neural network (RNN) in long-term WT prediction. Specifically, when accurately predicting WT 4 h ahead, the model achieves R², MSE, MAE, and wMAPE values of 0.912, 0.352, 0.191, and 1.390%, respectively. This research provides robust support for intelligent monitoring and regulation in aquaculture environments.