Research On Water Quality Prediction For Cost Reduction In Wastewater Treatment

Wastewater treatment is crucial to solve the problem of water resources.With the national attention and a large amount of investment,domestic wastewater treatment industry is developing rapidly,but there are still two problems restricting its sustainable development.On the one hand,wastewater treatment plants’ management is difficult,and there is a lack of reference for their expansions and reconstructions.On the other hand,the fact that their management relies on manual practical experience further leads to the high operating costs,the extensive management mode and the low degree of automation.In order to solve these problems,we can start from three aspects,including performing long sequence prediction of long-term influent water quality by discovering its changing law,which provides auxiliary decision-making suggestions for the daily operation management,the expansion and reconstruction of wastewater treatment plants;modeling the wastewater treatment process,predicting the future effluent water quality according to the current influent water quality,so as to ensure the effluent water quality under the limits;automatically optimizing chemical consumption based on the accurate prediction of effluent water quality,which provides a breakthrough for the fine management and cost reduction of wastewater treatment.Related work for these three research fields still has shortcomings,such as the lack of researches that studied a stable and accurate long sequence prediction model for the influent water quality series with large volatility,strong randomness and aperiodicity;the effluent water quality prediction accuracy of existing models decrease when the input sequences are too long;in the task of automatic chemical consumption optimization based on effluent water quality prediction,the variable length of input sequences and the reliability of prediction models were not studied.In view of the above shortcomings,a variety of deep learning methods were attempted to solve the problem of accurate prediction of the influent and effluent water quality.Through the analysis of the relationship between the input and output features,the reliability of the effluent water quality prediction was studied.Specifically,this thesis includes the following three aspects:1.In order to solve the problem of steadily and accurately predicting long influent water quality sequences with large fluctuation,strong randomness and aperiodicity,a model based on long short term memory network was proposed to model the series,and capture the changing law of the sequences through the autoregressive prediction method.The experiments on a real wastewater treatment dataset shown that,compared with other methods based on deep learning,traditional machine learning and statistics-based approach,this method achieved better prediction accuracy and stability.2.Aiming at the problem that the accuracy of the existing effluent water quality prediction methods declines when the length of the input sequences are too long,an convolution neural network was used by arranging the input sequence into a twodimensional matrix and regarding it as an ”image”,which models the wastewater treatment process and accurately predicts the effluent quality.The experiments on real wastewater treatment datasets shown that,this method achieves the accuracy and efficiency of prediction required by practical applications.3.Aiming at the problem that in the automatic chemical consumption optimization task the input sequences with variable length will cause the worse effluent water prediction and the reliability of the prediction model is uncertain,a deep learning model based on attention mechanism was proposed to model the input sequences of any length and to eliminate the influence of the irrelevant part in the input sequences.The reliability of the prediction model was analyzed by the partial dependence plot of the input and output features.Experiments on a real wastewater treatment dataset verified the accuracy and reliability of the proposed model.