Prediction of Waste Sludge Production in Municipal Wastewater Treatment Plants by Deep-Learning Algorithms with Antioverfitting Strategies

Accurate prediction of waste sludge production in municipal wastewater treatment plants (WWTPs) is critical for aligning it with treatment capacity. However, traditional statistical methods typically fail to capture the complex relationships inherent in wastewater treatment, thus leading to predictions with significant deviations. Machine learning methods (MLMs) enhance prediction accuracy but are challenged by extensive manual data labeling, limited adaptability, and low interpretability. Herein, we developed several deep-learning algorithms (DLAs), including bidirectional long-short-term memory (BiLSTM), BiGRU (bidirectional gated-recurrent unit) with attention, Transformer, and BiLSTM combined with convolutional neural networks (CNN), for predicting waste sludge production by automatically labeling data sets from a WWTP in Beijing (north China). These models with better interpretability were optimized with antioverfitting techniques, such as layer-wise neuron dropout, batch normalization, and ridge regularization, to enhance generalization across diverse data sets. The Transformer and BiLSTM + CNN, demonstrated the high prediction accuracy with mean square errors (MSE) of 0.014–0.027 and the correlation coefficients of 0.53–0.72 between validation outcomes and actual data. When tested on data sets from two WWTPs in Guangzhou (south China), with different wastewater qualities and treatment scales, these DLAs showed prediction errors as low as 3.15 ± 1.81 to 3.78 ± 1.75%, confirming their high adaptability. In contrast, MLMs, including random forest, eXtreme gradient boosting, and support vector regression, yielded the higher prediction errors of 30–65% and MSE of 0.026–0.14 as well as the lower correlation coefficients of 0.25–0.48, even using antioverfitting strategies. Shapley-additive-explanations analysis revealed that temperature variation was a key factor influencing sludge production. This study highlights the great potential of DLAs in the accurate prediction of waste sludge production.