Time-aware attention network for multi-step future prediction in wastewater treatment

Modern wastewater treatment plants (WWTPs) face challenges from increasing influent loads, strict regulations, and operational volatility. Existing machine learning and deep learning models often rely on single-target point predictions and fail to capture multivariate, temporally coupled, multi-step dynamics, while also underutilizing heterogeneous sensor data and multiscale temporal patterns. To overcome these limitations, we developed a deep learning framework, the Time-Aware Attention Net (TAANet), designed to predict parameters at future time points in the treatment process. TAANet introduces three key advances: (1) time-ahead forecasting that simultaneously predicts multiple indicators at the trajectory level; (2) an attention-based fusion module that dynamically captures time-varying interdependencies among influent and process variables, enabling robust and adaptive prediction; and (3) multi-component temporal embedding that explicitly encodes periodic and calendrical structures to represent nonstationary system behavior. Using datasets derived from long-term experiments of anaerobic biochemical treatment applied to real municipal wastewater, TAANet demonstrated state-of-the-art accuracy and resilience, even under extreme and nonstationary conditions. Compared with the baseline models, including Informer, PatchTST, iTransformer, and Long Short-Term Memory, TAANet achieved substantial improvements in predictive performance. For example, in chemical oxygen demand removal efficiency prediction, the R 2 value improved from baseline ranges of 0.19–0.30 to 0.46, while the mean squared error was reduced by up to 32.8%. These results demonstrate strong capability for multi-step forecasting in wastewater treatment processes. In practical applications, the proposed framework has the potential to support future-state awareness and operational decision-making in WWTPs, thereby contributing to more proactive process management and energy-efficient system operation.