A hybrid autoencoder–transformer‐based soft sensor with temporal embeddings for real‐time and noise‐resistant effluent quality prediction in wastewater treatment plants

Abstract Precise, real‐time forecasting of effluent quality variables is imperative for the stable operation of wastewater treatment plants (WWTPs). However, conventional laboratory‐based measurements of effluent parameters, such as chemical oxygen demand (CODe), total nitrogen (TNe), and total phosphorus (TPe), are expensive and time‐consuming. They, thus, cannot be used for online measurements, feedback control, and optimization. The current models of soft sensors, such as statistical and deep learning approaches, are usually unable to capture nonlinear dynamics, long‐range temporal dependencies, and noise in multivariate time series data, leading to limited robustness and scalability. To address these weaknesses, this paper proposes a hybrid soft sensor, the temporal embedding stacked autoencoder transformer (TESAEFormer) for effluent quality prediction. The method combines a stacked autoencoder to learn denoising and feature compression, and temporal embeddings to encode periodicities (daily and weekly), and a transformer encoder with ProbSparse attention to learn both short‐ and long‐term dependencies effectively. The model has been rigorously tested on Benchmark Simulation Model No. 2 (BSM2) and the real‐world Dongguan WWTP datasets, where it consistently outperformed eight state‐of‐the‐art baselines. TESAEFormer had higher predictive accuracy (R 2 of about 95. 5% on BSM2 and approximately 92% on DWWTP), smaller predictive errors, and greater noise resistance. The results indicate that TESAEFormer can provide a practical, efficient platform for real‐time effluent monitoring, reducing reliance on laboratory results and promoting compliance with environmental standards.