Scholarlink

<p>Access to clean water is one of the most critical challenges facing modern society, especially as rapid urban growth, industrial expansion, and population pressures continue to strain water resources. In light of these growing concerns, Sustainable Development Goal 6 (SDG 6) represents a global pledge to ensure clean water and sanitation for all. Central to achieving this goal are Wastewater Treatment Plants (WWTPs), which require reliable tools to monitor and predict water quality parameters for better management and compliance. This study presents an innovative deep learning framework built on Transformer-based architectures to address this challenge. First, introduce TransGAN, a Transformer-driven Generative Adversarial Network designed to produce high-fidelity synthetic tabular data, helping to overcome the common issue of limited training data in WWTP systems. Next, propose TransAuto, a dual-purpose Transformer Autoencoder capable of detecting anomalies and identifying key features in complex multivariate time series data—two crucial tasks for ensuring data quality and interpretability. To enhance predictive performance, multiple Transformer-based architectures were explored. The Time Series Transformer (TST), a streamlined variant of the Temporal Fusion Transformer (TFT) that integrates Variable Selection Networks (VSN) and Gated Linear Units (GLU), demonstrated superior accuracy among single-model approaches. In addition, the pretrained TimeGPT model was evaluated for its ability to capture complex temporal dynamics, offering valuable performance benchmarks. For further performance gains, an ensemble learning model that stacks three state-of-the-art Transformers: Informer, Autoformer, and FEDformer. Empirical results demonstrate strong performance: TST achieved MSE of 0.0028 and R<sup>2</sup> of 0.9643, while the ensemble model reached an MSE of 0.0036, RMSE of 0.0582, MAE of 0.0438, and R<sup>2</sup> of 0.9646. These outcomes highlight the effectiveness of Transformer-based models in accurately forecasting wastewater quality and support their use in real-world water management applications, aligning strongly with the goals of SDG 6.</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

Data driven multi-stage transformer based framework for intelligent water quality monitoring

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月24日

查看全文
  • Ramya S,
  • S. Srinath,
  • Pushpa Tuppad

摘要

Access to clean water is one of the most critical challenges facing modern society, especially as rapid urban growth, industrial expansion, and population pressures continue to strain water resources. In light of these growing concerns, Sustainable Development Goal 6 (SDG 6) represents a global pledge to ensure clean water and sanitation for all. Central to achieving this goal are Wastewater Treatment Plants (WWTPs), which require reliable tools to monitor and predict water quality parameters for better management and compliance. This study presents an innovative deep learning framework built on Transformer-based architectures to address this challenge. First, introduce TransGAN, a Transformer-driven Generative Adversarial Network designed to produce high-fidelity synthetic tabular data, helping to overcome the common issue of limited training data in WWTP systems. Next, propose TransAuto, a dual-purpose Transformer Autoencoder capable of detecting anomalies and identifying key features in complex multivariate time series data—two crucial tasks for ensuring data quality and interpretability. To enhance predictive performance, multiple Transformer-based architectures were explored. The Time Series Transformer (TST), a streamlined variant of the Temporal Fusion Transformer (TFT) that integrates Variable Selection Networks (VSN) and Gated Linear Units (GLU), demonstrated superior accuracy among single-model approaches. In addition, the pretrained TimeGPT model was evaluated for its ability to capture complex temporal dynamics, offering valuable performance benchmarks. For further performance gains, an ensemble learning model that stacks three state-of-the-art Transformers: Informer, Autoformer, and FEDformer. Empirical results demonstrate strong performance: TST achieved MSE of 0.0028 and R2 of 0.9643, while the ensemble model reached an MSE of 0.0036, RMSE of 0.0582, MAE of 0.0438, and R2 of 0.9646. These outcomes highlight the effectiveness of Transformer-based models in accurately forecasting wastewater quality and support their use in real-world water management applications, aligning strongly with the goals of SDG 6.