Predicting WastewaterBOD Levels with Neural Network Time Series Models

The quality of treated wastewater has always been an important issue, but it becomes even more critical as human populations increase. Unfortunately, current ability to monitor and control effluent quality from a wastewater treatment process is primitive (Wen & Vassiliadis, 1998). Control is difficult because wastewater treatment consists of complex multivariate processes with nonlinear relationships and time varying dynamics. Consequently, there is a critical need for forecasting models that are effective in predicting wastewater effluent quality. Using data from an urban wastewater treatment plant, we tested several linear and nonlinear models, including ARIMA and neural networks. Our results provide evidence that a nonlinear neural network time series model achieves the most accurate forecast of wastewater effluent quality. 701 E. Chocolate Avenue, Suite 200, Hershey PA 17033-1240, USA Tel: 717/533-8845; Fax 717/533-8661; URL-http://www.idea-group.com ITB9531 IDEA GROUP PUBLISHING This chapter appears in the book, Neural Networks in Business Forecasting, edited by G. Peter Zhang. Copyright © 2004, Idea Group Inc. Copying or distributing in print or electronic forms without written permission of Idea Group Inc. is prohibited. Predicting Wastewater BOD Levels 103 Copyright © 2004, Idea Group Inc. Copying or distributing in print or electronic forms without written permission of Idea Group Inc. is prohibited. INTRODUCTION It is a common practice today in most nations to discharge treated wastewater into the ecosystem. Because wastewater treatment is a costly process, wastewater is not typically purified to be safe for human consumption. The aquatic environment is often utilized to help in the “cleansing” process. As populations increase, higher quantities of treated wastewater are discharged into the environment which can overload ecosystems, leading to unsafe conditions for humans and animals. This has actually happened in the Narragansett Bay in Rhode Island, US, which has several wastewater treatment facilities discharging along its shores. The bay has been closed to commercial fishing and recreation on many occasions when residents have become ill from food caught in the bay or from exposure to the water. Unfortunately, it is currently very difficult to predict when a body of water will become unsafe. Thus, public health is under a constant threat from potentially unclean waters (Yeung & Yung, 1999). The present ability to monitor and control the effluent quality from a wastewater treatment process is described by researchers as primitive and notoriously difficult (Wen & Vassiliadis, 1998; Boger, 1997). Wastewater treatment has been characterized as a complex multivariate process with highly variable inputs, nonlinear time varying dynamics, and an autocorrelated time series structure that is subject to large disturbances (Lindberg, 1997). In addition, wastewater measurement systems can be unreliable and require lead times as long as five days to measure biochemical oxygen demand (BOD). The effluent BOD level is an important indication of the quantity of oxygen that will be depleted from the aquatic environment. Recently it has become possible to exercise more effective control of wastewater processes based on (1) the development of advanced forecasting models, (2) more accurate and timely measurement of process variables (Parkinson, 1998), and (3) wireless technology that provides a real time view of the treatment process (Cheek & Wilkes, 1994). In this research we investigate the accuracy of neural network time series models to forecast wastewater effluent BOD from an urban wastewater treatment plant. We also benchmark the performance of an Autoregressive Integrated Moving Average (ARIMA) model which includes intervention terms. Our results provide evidence that the nonlinear neural network time series model achieves a substantially more accurate forecast of wastewater BOD than does the ARIMA model. In the next section we review recent applications of time series neural network forecasting models. We then describe design considerations for short 17 more pages are available in the full version of this document, which may be purchased using the "Add to Cart" button on the publisher's webpage: www.igi-global.com/chapter/predicting-wastewaterbod-levelsneural-network/27246