A dual-stage deep learning model for dissolved oxygen forecasting in aquaculture using wavelet decomposition and graph convolution

Dissolved oxygen (DO) sensors often collect data that exhibit instability and latency, seriously impeding accurate monitoring and timely responses. Given the critical importance of DO in aquaculture—directly influencing the metabolism, growth, and survival of aquatic organisms—optimizing sensors or enhancing data processing techniques becomes essential. However, existing prediction methods often face limitations such as instability, low accuracy, and poor real-time performance, especially in multi-step predictions involving non-stationary features, complex dependencies, and long-term trends. To address the limitations of existing DO prediction models—particularly in capturing non-stationary patterns, modeling complex dependencies, and achieving high accuracy over long forecasting horizons—this study introduces a novel forecasting framework inspired by the Transformer paradigm. Rather than adopting the standard architecture, the proposed model is designed from the ground up, integrating domain-tailored mechanisms such as wavelet-based decomposition, graph-based spatial modeling, and temporal encoding enhancements. These design choices aim to improve the model’s adaptability to dynamic aquaculture environments and its robustness in multi-step forecasting tasks. Validation on the Data1 dataset demonstrates that the proposed model achieves a minimum MAE of 0.037 and RMSE of 0.048, representing improvements of 48.61% and 44.19%, respectively, compared to the baseline Transformer model. The model effectively captures dynamic variations and long-range dependencies in DO time series, significantly enhancing the accuracy and stability of predictions. These improvements are of great significance for advancing intelligent aeration control and environmental management in aquaculture.