In wastewater treatment processes, the efficient modeling of key water quality parameters is crucial for achieving process control optimization, anomaly detection, and decision support. However, the process data generally exhibits characteristics such as temporal dependence, multivariable coupling, and non-stationarity under varying operating conditions, posing significant challenges to accurate modeling. To address these issues, this paper proposes a Lightweight Multivariate Time Series Prediction Method for Wastewater Treatment based on the Stationary Wavelet Transform (SWT) and Collaborative Attention (CA) mechanism. This model first performed multi-scale decomposition on wastewater data and used the stationary wavelet transform to extract data features from sequences at different scales. Subsequently, a collaborative attention mechanism based on geometric attention and sparse attention was constructed to effectively capture the complex coupling relationships and temporal features among key water quality parameters. Finally, the features reconstructed via inverse wavelet transform were mapped to the final prediction results through a dual-prediction layer. The model was trained and validated on a measured dataset from a wastewater treatment plant in Dongguan, with multi-step prediction tasks and partial data visualization analyses conducted. Experimental results show that, in the 24-step multi-output prediction tasks, the proposed model achieves a reduction of 9.15% to 37.70% in multi-output root mean square deviation (RMSSD) compared to benchmark models. In other prediction tasks, its accuracy ranks second only to TimesNet, which has a significantly larger parameter scale. These results demonstrate an effective balance between lightweight design and high accuracy, thereby validating the efficacy of the proposed model for time-series prediction in wastewater treatment.

Activated sludge process is the most commonly used sewage treatment process in China. The occurrence of sludge bulking is an unavoidable and urgent problem for the stable and reliable operation of activated sludge process. To solve this problem, this paper proposed a new full life-cycle fault diagnosis method to monitor sludge bulking and provide reasonable decision support after accurate fault warning. In order to fully mine the hidden information of sludge bulking data, this paper used the canonical correlation analysis (CCA) and absolute average amplitude value (AMAV) to extract the relevant features and apply them to fault detection. The contribution plots were improved by rearranging historical observation samples and applied to fault isolation. According to the results of fault warning, a fault propagation location method based on feature extraction of AMAV and multivariate Granger causality (MVGC) analysis was proposed. The field data collected in a sewage plant were used for experiments. The results show that the proposed method can detect, separate and analyze the occurrence of sludge bulking timely and effectively.

Wastewater treatment system is a dynamic system with complex nonlinearity and large time delay. Due to the complexity of the process, the incompleteness of the testing equipment and the constraint of economic cost, some important effluent indicators cannot be detected accurately. To solve this problem, this paper proposes a soft-sensor method based on an ensemble Kalman filter-Elman neural network. The traditional dynamic neural network has the dynamic memory ability to process time-delay data, so it can be used in data-driven soft sensing modeling. However, the conventional training method is easy to trap in a local minimum, resulting in poor prediction performance. This paper introduces the ensemble Kalman filter and the dual finite-size ensemble Kalman filter, and, together with the Elman neural network for gradient-free training, to construct two soft sensor models, which not only improve the prediction performance of Elman neural network but also provide a simple and gradient-free training method for neural network. The two models are then applied to a dataset of the University of California, Irvine (UCI data). The results show that the proposed method based on ensemble Kalman filter-Elman neural network possesses good prediction performance, and that the ensemble Kalman filter can be used as an alternative gradient-free method to train neural networks.