Prediction of Wastewater Treatment Plant Performance using Artificial Neural Network Model

Abstract

환경 규제가 점차적으로 강화됨에 따라, 하수처리장의 안정적이고 경제적인 운전을 위한 ICA (Instrumentation Control and Automation) 기술들은 빠르게 발전해 왔다. 개발된 대부분의 다른 모델들은 시간에 변화하는 많은 변수들을 가짐으로써 보정하기 어럽다는 단점을 가지는 결정론적 모델들이었다. 본 연구에서는 유출수 예측 모델의 필요성에 따라 공정으로부터 확보된 측정 데이터를 기반으로 주어진 입력변수와 목표 변수 간의 변화 패턴만을 고려하여 생성되는 Black-box 모델인 인공신경망을 사용하여, 실제 하수처리장의 1차 침전지 유출수질 뿐만 아니라 2차 침전지 유출수질을 예측 할 수 있었다. 인공신경망은 그 구조에 따라 Single Layer Feed-forward Network Multi-layer, Feed-back network recurrent network, self-organized network으로 분류된다. 사용목적에 따라 적절한 형태의 신경망 구조를 선택하게 된다. 유출수 예측이 목적이기 때문에 모델링과 예측 분야에서 가장 일반적으로 사용되고 있는 feed-forward back-propagation network을 사용하였다. 본 연구에서는 신호전달을 위한 전달함수로써 tan-sigmoid function을 사용하였다. 그리고 역전파 알고리즘의 단점을 극복하고 더 빠른 학습이 가능한 Levenberg-Marquart (LM) 알고리즘을 학습 알고리즘으로 사용하였다. 대상 하수처리장에서 실제 측정된 2005년 유입 수질 데이터들과 다른 데이터들을 수집하였고 전처리를 거친 200개의 데이터을 예측 모델 개발을 위해 사용하였다. 만들어진 신경망의 예측 성능을 검증하기 위해 나머지 164개의 데이터를 사용하였다. 1차 침전지 유출수 예측의 경우에는 인공 신경망뿐만 아니라 결합성 신경망 모델을 하였다. 2차 침전지 유출수 예측에서는 IWA task group에 의해 제안된 활성 슬러지 모델 데이터와 포기조 안에 데이터들에 의해 만들어진 신경망을 사용하였다. 따라서, 시행착오법 (Trial and Error Method)에 의해 적절한 신경망 의 구조와 epoch수를 바탕으로 만들어진 신경망을 일반화하기에 충분하게 가중치 값들을 결정하였으며 상용소프트웨어인 MATLAB에서 제공되는 neural network toolbox를 이용해서 모든 시뮬레이션 작업을 수행하였다.

Proper operation and control of municipal wastewater treatment plants is important in producing an effluent which meets quality requirements of regulatory agencies and in minimizing detrimental effects on the environment. Predicting the plant water quality parameters using conventional experimental techniques is also a time consuming step and is an obstacle in the way of efficient control of such processes. For control and automation of the plant treatment processes, lack of reliable on-line sensors to measure water quality parameters is one of the most important problems to overcome. And the accuracy of existing hardware sensors is also not sufficient. This paper deals with the development of software sensor techniques that estimate the target water quality parameter from other water quality parameters. Here an artificial neural network (ANN) and a hybrid ANN model, combining with principal component analysis (PCA), both of them were applied to predict the wastewater effluent quality parameters, biological oxygen demand (BOD), chemical oxygen demand (COD), suspended solids (SS), total nitrogen (TN), total phosphorous (TP) of the primary settlement tank based on past information. The PCA was used to synthesize the input water quality parameters in order to reduce the dimension of the inputs. And the back-propagation feed-forward neural network (FBNN) was chosen to model the wastewater treatment plant through this study, which is the South Wastewater Treatment Plant (WWTP) at Busan City, Korea. The tan-sigmoid function was used as activation function to transfer signal at the neural network. And the Levenberg-Marquart algorithm was used as learning algorithm to train neural network. All the 364 data sets, which were collected from the plant during 2005, 200 data sets and other 164 data sets, were used for training and validation，respectively. The hybrid ANN&PCA models for prediction of water quality parameters was also used in the primary settlement tank (PST) effluent, comparing with the prediction results by ANN. Following the prediction of first physical and chemical process in the wastewater treatment, it is the biological wastewater treatment process, which is commonly used to treat municipal and industrial wastewaters and so important process in the treatment. Special attention has been paid to biological processes modeling, both for wastewater treatment and sludge stabilization processes. In the prediction of secondary settlement tank (SST) effluent presented here is another hybrid ANN model, which is using some data from the activated sludge model (ASM) simulator in order to strength and advance ANN model. The hybrid ANN techniques show an enhancement of prediction capability and reduce the over-fitting problem of neural networks. The results showed that the hybrid ANN technique can be used to extract information from noise data and can provide more accurate predictions of the primary and secondary settlement tank effluent stream and then further to describe the nonlinearity of complex wastewater treatment.