Improving multi-step dissolved oxygen prediction in aquaculture using adaptive temporal convolution and optimized transformer

Abstract

Accurate dissolved oxygen (DO) prediction is crucial for optimizing aquaculture efficiency. However, existing forecasting methods often struggle to capture periodic patterns, model complex feature dependencies, and maintain high accuracy in multi-step predictions. To address these challenges, this study proposes an enhanced Transformer-based model designed to improve both prediction accuracy and stability. The model first integrates an Adaptive Temporal Convolutional Network (ATCN) to extract periodic patterns and local temporal features from time-series data. Then, a Transformer encoder with linear attention mechanisms and relative position encoding is employed to enhance feature extraction and sequence modeling. To further strengthen temporal dependency learning, the conventional decoder is replaced with a modified Gated Recurrent Unit (GRU), and a linear regression-based error correction mechanism is introduced to refine multi-step forecasting accuracy. Experimental results on Public Dataset 3 demonstrate that the proposed model achieves an average MAE of 1.624 and RMSE of 2.254, reflecting improvements of 45.06% and 40.57%, respectively, compared to the baseline Transformer model. These findings highlight the model’s ability to effectively capture complex temporal dependencies, significantly enhancing DO prediction accuracy and robustness in multi-step forecasting tasks. Additionally, the model demonstrates strong performance in pH prediction, underscoring its potential for multi-parameter water quality forecasting. This work provides valuable insights for improving water quality management and mitigating risks in aquaculture.