Deployment of interpretable machine learning in a water treatment device - feasibility exploration of performance enhancement

Machine learning (ML) provides a promising tool for predictive control and unattended operation of water treatment processes, especially in decentralized water supply due to its high variability of water quality and scarcity of manpower. However, model deployment after ML development is a challenge that hinders their implementation. Herein, this study constructed a novel electrocoagulation membrane cathode reactor (ECMCR) as an example of decentralized water treatment due to its highly integratable. Nine ML models were established to determine the ECMCR operational current density (CD) and the Bayesian technique was employed for adaptive hyperparameters tuning of each ML model. Light gradient boosting machine (LightGBM) with turbidity, humic acid (HA), pH, temperature and time as input performed best (R2 = 0.99, MSE = 0.0004) during training, whereas the random forest (RF) was effective and robust for test data (R2 = 0.98, MSE = 0.0056) with same inputs. The mean decrease impurity (MDI) and Shapley Additive exPlanation (SHAP) method indicated that the RF model made predictions based primarily on influent turbidity, which implied a correct "understanding" of the RF model. Finally, the model was deployed to the ECMCR to explore the regulation capability, and the HA removal and turbidity removal of the ECMCR increased by 23.68 % and 23.83 %, respectively. The results demonstrate that ML technology could assist in achieving the desired performance of drinking water treatment devices without manual involvement.