DecompSTNet: A Decomposition-Based Deep Learning Network for Multi-Scale Dissolved Oxygen Prediction

Accurate prediction of dissolved oxygen (DO) concentration is crucial for water quality management and aquatic ecosystem health. However, achieving precise multi-scale DO predictions remains challenging due to the multiple influencing factors. This study proposes DecompSTNet, a novel decomposition-prediction-aggregation framework that effectively extracts and predicts both seasonal and trend characteristics of DO. The model incorporates two key components: a Periodic Trend Predictor (PTP) designed to capture long-term evolution patterns, and a Transformer-based seasonal prediction module specialized in handling short-term fluctuations. Experiments demonstrate that DecompSTNet significantly outperforms five mainstream deep learning models across different prediction horizons (6-step, 12-step,18 steps, 24 steps and 30-step). Compared to the baseline Transformer model, it reduces MAE by 3.6%-27%, exhibiting superior prediction accuracy.