Scholarlink

Data scarcity poses a significant challenge in time series forecasting using Deep Learning (DL), particularly in wastewater treatment, where data are scarce, infrequent, and often operationally constrained. This study focuses on forecasting influent Chemical Oxygen Demand (COD) and evaluating the effect of synthetic Data Augmentation (DA) on model performance. To this end, two techniques, Vector Autoregressive Moving Average (VARMA) and Transformer Time Series GAN (TTS-GAN), were tested across three DL models: Temporal Convolutional Network (TCN), Long- and Short-Term Time-Series Network (LSTNet), and the hybrid ConvRecurrentFusion (CRF). Results show that all models improved with DA, with VARMA consistently outperforming TTS-GAN. The CRF model with VARMA achieved the best performance, with a Root Mean Square Error (RMSE) of 202.0 mg O2/L and a Mean Absolute Error (MAE) of 166.4 mg O2/L.

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

Synthetic Data Augmentation for COD Prediction in WTTPs: A Comparative Study of Deep Learning Models with VARMA and TTS-GAN

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Afonso Bessa,
  • Pedro Oliveira,
  • Millena Santos,
  • S. A. Silva,
  • Paulo Novais

摘要

Data scarcity poses a significant challenge in time series forecasting using Deep Learning (DL), particularly in wastewater treatment, where data are scarce, infrequent, and often operationally constrained. This study focuses on forecasting influent Chemical Oxygen Demand (COD) and evaluating the effect of synthetic Data Augmentation (DA) on model performance. To this end, two techniques, Vector Autoregressive Moving Average (VARMA) and Transformer Time Series GAN (TTS-GAN), were tested across three DL models: Temporal Convolutional Network (TCN), Long- and Short-Term Time-Series Network (LSTNet), and the hybrid ConvRecurrentFusion (CRF). Results show that all models improved with DA, with VARMA consistently outperforming TTS-GAN. The CRF model with VARMA achieved the best performance, with a Root Mean Square Error (RMSE) of 202.0 mg O2/L and a Mean Absolute Error (MAE) of 166.4 mg O2/L.