Enhanced Insights into Effluent Prediction in Wastewater Treatment Plants: Comprehensive Deep Learning Model Explanation Based on SHAP

Models are increasingly being utilized to improve the understanding and operation of wastewater treatment plants (WWTPs) in the face of escalating water resource challenges. Abundant operational data provide extensive opportunities for the development of machine learning (ML) and deep learning (DL) models. However, the coupling and time lag among the features exacerbate the black-box nature of such models, hindering their application in WWTPs. In this study, we construct a DL model using a long short-term memory (LSTM) algorithm capable of accurately predicting the effluent quality in a full-scale WWTP with finely tuned hyperparameters and rationally chosen input features. Comprehensive model explanation based on Shapley additive explanations (SHAP) is implemented to clarify the contributions of multivariate time series (MTS) inputs to the predicted results in terms of feature and time dimensions. The LSTM models exhibit excellent accuracy (R2 of 0.96, 0.95, and 0.76 and MAPE of 5.49, 7.17, and 13.37%, respectively) in predicting effluent chemical oxygen demand (COD), total phosphorus (TP), and total nitrogen (TN) better than other baseline ML models. The SHAP results quantify what input features are most important when they exert influence and how they impact results. The analysis from the temporal dimension further explains the time lag characteristics of the wastewater treatment process and justifies the introduction of MTS. Compared to correlation analysis and without feature engineering, the feature selection method by SHAP significantly enhances the predictive accuracy. The combinations of input features are adjusted based on the Shapley values, and features with strong interactions and significant contributions to the model output are identified. This is a novel attempt to construct a WWTP model based on LSTM with both excellent accuracy and explainability and to clarify the influence of MTS inputs on prediction results. This work shows the potential of applying DL to model WWTPs and enhances their performance.