Hybrid dual decomposition-machine learning model for forecasting dissolved oxygen in aquaculture

In aquaculture, dissolved oxygen levels are a primary determinant of water quality; leveraging accurate forecasts for this parameter directly contributes to improved yields and higher-quality harvests. Given that current dissolved oxygen prediction models utilize just one decomposition process, they are unable to sufficiently mitigate residual noise interference or properly handle non-stationary features. To enhance prediction accuracy, our model begins by processing the original dissolved oxygen sequence via an improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN), yielding multiple intrinsic mode functions (IMFs). These IMFs are then clustered according to their fuzzy entropy using the K-means algorithm. The component with the maximum fuzzy entropy value is selected for further refinement through a Sequential Variational Mode Decomposition (SVMD). This dual decomposition approach effectively preserves valid information while enhancing detailed feature extraction. For the forecasting stage, an Extreme Learning Machine (ELM) whose parameters are optimized by the rime optimization algorithm is developed to generate forecasts for each sub-component. Ultimately, the predictions of all components are aggregated through a Back-Propagation (BP) neural network to produce the final dissolved oxygen time series. The proposed model is applied to verify dissolved oxygen in the standard pond of Huludao sea cucumber aquaculture base, Liaoning province, China. Evaluation results show that the proposed model exceeds the performance of existing comparative models. With the RMSE of 0.2443, MAE of 0.1782, MAPE of 0.0352, and R² of 0.9835, the model exhibits exceptional predictive accuracy and operational stability, validating that it is well-suited to meet the practical demands of contemporary aquaculture.