Dual purpose of Shapley Additive Explanation (SHAP) in model explanation and feature selection for artificial intelligence-based digital twin of wastewater treatment plant

Abstract

In this study, artificial intelligence (AI) black-box models called feedforward neural network (FFNN) as shallow learning and long short-term memory (LSTM) as deep learning were used to evaluate the biological oxygen demand of effluent (BOD_eff) and chemical oxygen demand (COD_eff) of the Tabriz wastewater treatment plant (WWTP). Daily data of the treatment plant from 2015 to 2021 were utilized for this modeling. Given the importance of selecting effective input parameters for modeling, four scenarios were employed for optimal input selection. The first scenario was based on the correlation coefficient (CC) method, the second on the mutual information (MI) method, the third utilized the Shapley additive explanation (SHAP) method for ranking the parameters, and the fourth scenario applied newly proposed hybrid MI-SHAP method for a two-step selection of input parameters. In MI-SHAP method, the number of inputs was first reduced using the MI method, and then the remaining parameters were ranked by the SHAP algorithm and used in the modeling process. This procedure significantly reduced the SHAP runtime. The explainable artificial intelligence (XAI) algorithm named SHAP was also used to visualize and illustrate how parameters influence the modeling results. To evaluate the results provided by the SHAP, another algorithm of the XAI, called accumulated local effects (ALE), was used. Through the use of XAI to illustrate the contribution of each input to the results, it was determined that BOD_eff and COD_eff with a one-day lag (BOD_eff(t-1) and COD_eff(t-1)) were the most influential features in the FFNN and LSTM. These parameters contribute approximately 54 % and 65 % of the total impact on the modeling outcomes in FFNN, and 56 % and 60 % in LSTM modeling, respectively.