Abstract
To achieve long-term accurate predictions of dissolved oxygen in aquaculture environments. In this paper, we propose the Time varying filter based empirical mode decomposition (TVFEMD) - Permutation Entropy (PE) -Temporal Convolutional Networks (TCN) - bidirectional gated recurrent units (BiGRU)-Improved Sparrow Search Algorithm (ISSA) (TPBI) model. First, dissolved oxygen data are applied to the Time-Variant Filtered Empirical Mode Decomposition (TVFEMD) model to remove noise factors in the data. Permutation Entropy (PE) is then applied to reconstruct the data and reduce its complexity. Additionally, Temporal Convolutional Networks (TCN) and Bidirectional Gated Recurrent Units (BiGRU) are combined to extract features from the denoised data, improving the model’s learning efficiency and prediction accuracy. Based on this, the Dynamic Opposite Learning Strategy Improved Sparrow Search Algorithm (ISSA) is introduced to optimize hyperparameters such as batch size and the number of hidden layer units.
The framework was applied to predict dissolved oxygen data from an aquaculture farm in Guangdong Province, with future 1-step, 3-step, and 6-step prediction experiments. The experimental results show that the proposed model outperforms the comparison models in predicting dissolved oxygen, particularly excelling in long-term predictions (6 steps). In terms of Mean Absolute Error (MAE), compared to models such as RNN, BiGRU, CEEMDAN-PE-TCN-BiGRU-ISSA, and TVFEMD-PE-TCN-BiGRU-SSA, the proposed model improved dissolved oxygen prediction by 50%, 55.2%, 50%, and 27.7%, respectively.
Ablation experiments were conducted to verify the effectiveness of all components. In terms of Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), dissolved oxygen prediction improved by 41.3% and 21.0%, respectively. The outstanding performance of this framework in long-term predictions of dissolved oxygen provides effective support for precise environmental control and early warning in aquaculture.
