Prediction Of Key Process Parameters In Drinking Water Treatment Plants Based On Machine Learning

Water resource is an important part of the ecological environment.Constructing a healthy water circulation system to manage the water resources as a whole is an important development goal.As the core of the urban water circulation system,water treatment plant is the source of urban water.However,the process of water treatment often involves many complex technologies.A large number of physical and chemical reactions within the water treatment process and the complex interaction between different stages become the obstacle to the overall control and optimization of water treatment.With the rapid development of machine learning,a new technical path for efficiency prediction and optimization of the drinking water treatment process is provided.However,a systematic and comprehensive model evaluation method has still lacked.In order to solve this question,this paper takes the logic layer as the core,establishes the connection between the data layer and the interaction layer,and constructs a multi-dimensional evaluation method that include model-,feature-,sustainability-level evaluation.Among them,the prediction performance of multiple candidate models will be evaluated in the model-level evaluation stage.Through this stage,the high accuracy of the final model will be ensured.Then,in the feature-level evaluation stage,we will estimate whether the model logic fits the domain knowledge from the perspective of features to ensure the reliability of the final model.Finally,the forward-looking sustainability-level evaluation stage will be used to evaluate the environmental impact of the model building and evaluation processes.Through this stage,the sustainability of final model will be ensured.In the experiments,the practical application potential of the above methods will be explored through case studies.During the construction of the operating parameter predicted model,the model-level evaluation results showed that among the five candidate models,the random forest(~2=0.837)and Light GBM(~2=0.878)have the best predictive performance.Subsequently,in the feature-level evaluation stages,the reliability of the above models is verified through Recursive feature elimination(RFE)and SHAP(SHapley Additive ex Planations)from the perspective of feature importance and model interpretability.Finally,the sustainability-level evaluation results show that the energy and time consumption of LGB is much smaller than that of RF.Therefore,the LGB model is selected as the final model.The results imply that the models constructed by our evaluation method have high accuracy,strong reliability,and good sustainability.On the basis of the above research,the effluent parameter prediction model is constructed through our evaluation method.Subsequently,by integrating above models,A drinking water treatment process parameters prediction and analysis system is constructed and the application domain is preliminarily discussed.The results show that our system can provide a reliable and powerful analysis technology for the water treatment process to cope with future environmental changes.