Scholarlink

<p>Dissolved oxygen (DO) is a crucial indicator for assessing the health of aquatic ecosystems. Accurate prediction of DO can help monitor water quality, prevent ecological degradation, and support sustainable water resource management. In this study, due to the limited improvement of bidirectional long short-term memory (BiLSTM) over long short-term memory (LSTM) in DO prediction, the pronounced seasonal fluctuations of DO were analyzed firstly. It is revealed that the multi-factor–driven seasonal nonlinear variations largely explain the limited effectiveness of BiLSTM in optimizing DO predictions. Differencing transformation (DT) preprocessing was found to effectively mitigate the impact of DO’s seasonal variations on the stationarity of long-time series inputs. Based on this, a new model framework by combining DT with BiLSTM was proposed to optimize the prediction results. The proposed framework significantly improved DO prediction accuracy, achieving reductions of 23.92%, 25.12%, and 48.92% in RMSE, MAE, and MASE, respectively, while increasing the coefficient of determination (R²) by 64.76% compared with the LSTM model. It also outperformed all the other models considered in this study. Furthermore, the error distribution density analysis showed that DT-BiLSTM can more effectively correct the systematic bias of prediction results and reduce the risk of underprediction, thereby enhancing the reliability of early DO pollution warnings. This study elucidates the influence mechanism of seasonal variability on DO prediction and offers an accurate early-warning tool for the study area.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An improved DT-BiLSTM framework to eliminate seasonal interference in DO prediction of Chengdu, China

  • 期刊论文

  • 发表日期：

    2026年7月2日

查看全文
  • Mei Li,
  • Kexing Chen,
  • Zengyao Li,
  • Yilin He,
  • Rui Xu,
  • Zhaoli Wang

摘要

Dissolved oxygen (DO) is a crucial indicator for assessing the health of aquatic ecosystems. Accurate prediction of DO can help monitor water quality, prevent ecological degradation, and support sustainable water resource management. In this study, due to the limited improvement of bidirectional long short-term memory (BiLSTM) over long short-term memory (LSTM) in DO prediction, the pronounced seasonal fluctuations of DO were analyzed firstly. It is revealed that the multi-factor–driven seasonal nonlinear variations largely explain the limited effectiveness of BiLSTM in optimizing DO predictions. Differencing transformation (DT) preprocessing was found to effectively mitigate the impact of DO’s seasonal variations on the stationarity of long-time series inputs. Based on this, a new model framework by combining DT with BiLSTM was proposed to optimize the prediction results. The proposed framework significantly improved DO prediction accuracy, achieving reductions of 23.92%, 25.12%, and 48.92% in RMSE, MAE, and MASE, respectively, while increasing the coefficient of determination (R²) by 64.76% compared with the LSTM model. It also outperformed all the other models considered in this study. Furthermore, the error distribution density analysis showed that DT-BiLSTM can more effectively correct the systematic bias of prediction results and reduce the risk of underprediction, thereby enhancing the reliability of early DO pollution warnings. This study elucidates the influence mechanism of seasonal variability on DO prediction and offers an accurate early-warning tool for the study area.