Modeling And Prediction Of Wastewater Treatment Process Soft Sensor Based On Deep Neural Network

Real-time measurements of key effluent parameters play a highly crucial role in the wastewater treatment process（WWTPs）.However,there are a large number of variables in the wastewater treatment system that are difficult to measure online,such as biochemical oxygen demand（BOD₅）,which is the standard 5-days off-line delay measured.It makes it inadequate for measurement in real-time and may result in violations of effluent quality.Soft sensor technology addresses these issues effectively.The soft sensor calculates variables that are difficult to measure by correlating them with easily available variables.In soft sensor modeling,conventional machine learning approaches were widely used.However,these methods are regarded as shallow learning methods with one hidden layer of model structures.Shallow learning can be useful for simple processes and can handle problems with the use of a few samples and labeled data that include both input and target values due to time-consuming,cost,or technical limitations.Thus,when faced with highly complex processes,such approaches are often unsuitable for modern applications.To solve the problems of limited labeled samples,nonlinearity,and dynamics in WWTPS,a soft sensor modeling approach based on deep learning is proposed.The main work includes:1)Considering the strong processing capabilities of deep neural network stacked autoencoders（SAEs）in complex nonlinearities and the good performance of the genetic algorithm（GA）in optimization to solve the problem of limited labeled samples and nonlinearity,combining the two SAE+GA soft sensor approach for predictive modeling of BOD₅ parameter for online monitoring is proposed.Firstly,based on experimental data,the secondary variables（easy-to-measure）which have a strong correlation to the BOD₅ are chosen as model inputs,and the original data set was augmented by re-sampling and polynomial interpolation,which improved the completeness of the data.The problem of over-fitting of the model was alleviated.Secondly,unsupervised coarse learning tuning of multi-layer AE is performed,and then supervised learning fine-tuning is done using the stochastic gradient descent（SGD）algorithm.At the same time,a genetic algorithm search strategy is proposed to efficiently determine the number of hidden layers and the number of units in the hidden layer of the model which can automatically search for suitable values.A soft sensor model is studied to predict the BOD₅ in a wastewater treatment plant to evaluate the proposed approach.Nonlinear mapping problem between auxiliary variables and primary variables is solved,and model prediction accuracy is improved.2)Considering the strong processing capabilities of deep neural network long short-term memory（LSTM）in complex dynamics and the good performance of the genetic algorithm in optimization to address the issue of limited labeled samples and dynamic characteristic,a soft sensor approach that combines LSTM with GA for predictive modeling of BOD₅ online monitoring is proposed.Based on actual data,the selection of wrapped features based on the Pearson linear correlation was adopted to determine the prediction model’s optimal input variable set.raise data is used to extend the data set and diminish the over-fitting problem.At the same time,a GA is used to find optimum timestep and the number of units for LSTM model performance optimization in each layer,and adaptive moment estimation（Adam）is selected as the optimizer to minimize the loss function,which solves the dynamic nonlinear problem of primary variable prediction.This approach is appropriate for the dynamic modeling of the soft sensor.Based on the actual data of a wastewater treatment plant,the key effluent parameter BOD₅ is modeled and predicted.The experimental results show that the proposed soft sensor modeling approach based on SAE+GA and LSTM+GA has a better predictive performance of the key parameters of the online monitoring of wastewater than the existing methods.