Scholarlink

<p>High-precision prediction of estuarine water quality parameters is of great significance for ecosystem protection and water environment management. However, the superposition of multi-source driving mechanisms, pronounced high-frequency disturbances, and complex nonlinear coupling among water quality indicators in estuarine environments make it difficult for conventional approaches to achieve a satisfactory balance between predictive accuracy and stability. To address these challenges, this study proposes a multi-indicator water quality prediction framework, termed RimeSG-ATGNet, which integrates adaptive denoising, residual learning, and joint hyperparameter optimization. Specifically, Savitzky–Golay (SG) filtering is first employed to suppress high-frequency noise, while the rime optimization algorithm (RIME) is introduced to construct a unified hyperparameter search strategy that jointly optimizes denoising parameters, model architecture parameters, and training-related hyperparameters. On this basis, an Autoregressive Integrated Moving Average (ARIMA) model is used to characterize the linear trend components of water quality time series, and a TimesNet-conditioned GRU network with period-aware contextual information is adopted to model the nonlinear residuals, enabling joint prediction of dissolved oxygen (DO), pH, and turbidity (Turb). Experimental results based on datasets from three estuarine monitoring stations with distinct hydrological and environmental characteristics demonstrate that the proposed method outperforms multiple benchmark models across a range of evaluation metrics and exhibits statistically significant advantages under the Diebold–Mariano test. These results indicate that the proposed framework can effectively capture the multivariate temporal evolution of estuarine water quality and exhibits strong generalization ability and stable predictive performance under different estuarine monitoring conditions.</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

A multi-feature water quality prediction method based on adaptive denoising and periodicity-aware residual learning

  • 期刊论文

  • 发表日期：

    2026年6月16日

查看全文
  • Rong Ma,
  • Yulong Bai,
  • Xiaoxin Yue,
  • Xianbao Tan,
  • Wenjie Bao

摘要

High-precision prediction of estuarine water quality parameters is of great significance for ecosystem protection and water environment management. However, the superposition of multi-source driving mechanisms, pronounced high-frequency disturbances, and complex nonlinear coupling among water quality indicators in estuarine environments make it difficult for conventional approaches to achieve a satisfactory balance between predictive accuracy and stability. To address these challenges, this study proposes a multi-indicator water quality prediction framework, termed RimeSG-ATGNet, which integrates adaptive denoising, residual learning, and joint hyperparameter optimization. Specifically, Savitzky–Golay (SG) filtering is first employed to suppress high-frequency noise, while the rime optimization algorithm (RIME) is introduced to construct a unified hyperparameter search strategy that jointly optimizes denoising parameters, model architecture parameters, and training-related hyperparameters. On this basis, an Autoregressive Integrated Moving Average (ARIMA) model is used to characterize the linear trend components of water quality time series, and a TimesNet-conditioned GRU network with period-aware contextual information is adopted to model the nonlinear residuals, enabling joint prediction of dissolved oxygen (DO), pH, and turbidity (Turb). Experimental results based on datasets from three estuarine monitoring stations with distinct hydrological and environmental characteristics demonstrate that the proposed method outperforms multiple benchmark models across a range of evaluation metrics and exhibits statistically significant advantages under the Diebold–Mariano test. These results indicate that the proposed framework can effectively capture the multivariate temporal evolution of estuarine water quality and exhibits strong generalization ability and stable predictive performance under different estuarine monitoring conditions.