Water Quality Prediction Of Sewage Pipeline Based On Deep Learning

The urban sewage pipeline network is an important component of the urban drainage system,and it is also an essential basic livelihood project in urban construction,closely related to the daily life of urban residents.The urban sewage conveyed by the pipeline network has the characteristics of a wide range of pollutants and complex water quality,and understanding the changes in water quality in the pipeline is of great significance for improving pipeline maintenance levels and enhancing water treatment efficiency.Traditional sewage pipeline water quality prediction models require determination of the population distribution and quantity of microorganisms in sewage,the types and contents of pollutants in water,and calibration of a large number of parameters,resulting in high modeling complexity and low efficiency.In order to solve this problem,this paper introduces the concept of deep learning and applies it to sewage pipeline water quality prediction.By constructing a deep learning model to achieve water quality prediction for important nodes in the pipeline,the accuracy and efficiency of the prediction are improved.The main work and conclusions of this paper are as follows:(1)A method for collecting and preprocessing sewage pipeline water quality data is proposed.Combining common water quality parameters and the sewage characteristics of the research area,COD concentration and ammonia nitrogen concentration are selected as the indicators for this sewage pipeline water quality prediction.Suitable water quality monitoring sensors are selected and deployed at pipeline nodes to record COD concentration,ammonia nitrogen concentration,pH,temperature,and flow rate.After data collection,data cleaning and wavelet transform are performed to solve the problems of missing data and high-frequency mutation data that cannot be well learned.Using wavelet transform to denoise the water quality monitoring data can effectively remove noise while retaining important signal information.A dataset of 5688 groups of preprocessed data is constructed,and 4266 groups are selected for model training and 1422 groups for validation,with a data ratio of 4:1 between the training and validation groups.(2)Sewage pipeline water quality prediction models based on CNN,LSTM,and Informer networks are established,and corresponding model optimization methods are proposed for different networks.The water quality prediction results of different models are compared,and it is found that the Nash-Sutcliffe efficiency coefficient(NSE)of the Informer model in predicting COD concentration is 0.747,which is 5.21%and 7.33%higher than that of the LSTM model and the CNN model,respectively.In predicting ammonia nitrogen concentration,the NSE indicators of the three models are 0.708,0.718,and 0.751,respectively,with the Informer model showing the highest accuracy,and the LSTM model and CNN model showing relatively poor performance.It should be noted that both the LSTM model and the Informer model are time series models,and in sewage pipeline water quality prediction,the performance of the Informer model is better than that of the LSTM model.(3)A CNN-LSTM coupled model and a CNN-Informer deep learning coupled model were established to improve water quality prediction accuracy.The CNN network effectively extracts water quality monitoring data features through its convolutional layers,while the LSTM network,although better at handling long-term dependencies,has disadvantages in data feature extraction.Therefore,this paper coupled the CNN and LSTM networks to establish a CNN-LSTM coupled model for sewage pipeline water quality prediction.The NSE indicators of this coupled model for predicting COD concentration and ammonia nitrogen concentration were 0.739 and 0.743,respectively,which were 4.08%and 3.48%higher than those of the LSTM model(0.71 and 0.718,respectively).Adding convolutional operations to the input data in the Informer network for feature extraction can reduce the number of parameters and capture the spatial features of water quality information.This paper also established a CNN-Informer coupled model,with an NSE indicator of 0.768 for predicting COD concentration,which was 2.81%higher than the NSE indicator of the Informer model(0.747).The performance in predicting ammonia nitrogen concentration was also improved by 2.93%.Therefore,adding convolutional structures to time series models can process water quality time series features and improve model prediction accuracy.