State Water Resources Research Institute Program
Project ID: 2011DC128B
Title: Urban Stormwater Runoff Prediction Using Computational Intelligence Methods
Project Type: Research
Start Date: 3/01/2011
End Date: 2/28/2012
Congressional District: DC
Focus Categories: Non Point Pollution, Hydrology, Water Quality
Keywords: Nonpoint Source Pollution, Urban Runoff, Stormwater Management, Water Quality, Neural Network, Prediction
Principal Investigators: Zhang, Nian (University of the District of Columbia); Behera, Pradeep K.
Federal Funds: $ 10,800
Non-Federal Matching Funds: $ 4,968
Abstract: It has been recognized that urban stormwater pollution can be a large contributor to the water quality problems of many receiving waters. Depending upon the type of sewer system the stormwater runoff transports a wide spectrum of pollutants to local receiving waters through combined sewer overflows (CSOs) and/or stormwater discharges. Stormwater pollution is one of most important issues the District of Columbia faces. The downtown core of the District is serviced by combined sewer system. The development of the District over the years has increased its impervious area significantly which combines with inadequate drainage capacity of the sewer system results in CSOs and stormwater discharges to the Anacostia River, Potomac River and Rock Creek. To address this stormwater problem the DC Water (previously known as DC WASA) has developed a Long Term Control Plan (LTCP) which would cost several billion dollars. In order to support LTCP a continuous monitoring and modeling of the system is necessary not only to provide technical assessment but also to develop a cost-effective solution. Moreover, evaluations of runoff quantity and quality are necessary to assess the problem and to assess the performance of proposed best management practices. Forecasting of runoff quantity and quality benefit substantially from the progress of computational intelligence techniques, particularly neural networks. Computational intelligence relies on heuristic algorithms such as in fuzzy systems, neural networks, and evolutionary computation . In addition, it also embraces techniques that use swarm intelligence, chaos theory, artificial immune systems, and wavelets. Comparatively, various runoff forecast models based on neural networks perform much better in accuracy than many conventional prediction models -. However, a fact could not be neglected that most of such existing neural networks based models have not yet satisfied researchers and engineers in forecast precision so far, and the generalization capability of these networks needs further improving. For example, most publications used the feedforward neural networks with Backpropagation algorithms. However, a critical "drawback" of the Backpropagation algorithm is the local minima problem caused by neuron saturation in the hidden layer . Because of this, the algorithm cannot converge to the minimum error, and thus it cannot get accurate prediction results. To overcome the above challenges, it is extremely important to investigate new models with the potential for higher rates of prediction. According to the time series prediction competition results in the 2006, 2008, and 2010 Artificial Neural Network & Computational Intelligence Forecasting Competitions , recurrent neural networks, wavelet neural networks, particle swarm optimization methods, and fuzzy neural networks etc. have been widely recognized as the best models for time series prediction -. Because time series prediction is a generalized form of runoff quality prediction, we can expect these models will also work the best for the specific runoff quality prediction. However, these prospective methods have never been used for the runoff quantity and quality prediction problems. Therefore, we believe that it is imperative to investigate these state-of-the-art computational intelligence methods, or the combination of these methods on the application of runoff quality prediction. On the other hand, this effort can in turn promote the progress of computational intelligence technology. Moreover, the generalization capability of these methods can be further improved by applying them to other kinds of water quality prediction, such as water quality parameter prediction (i.e. total dissolved solids, electrical conductivity, turbidity, dissolved oxygen, plumbum, and water temperature etc.) , or to the assessment of class of water quality . The proposed research is intended to fill this gap by seeking broader computational intelligence solutions to the modeling and simulation of the runoff quantity and quality prediction. We will focus on the best models for time series prediction including recurrent neural networks, wavelet neural networks, particle swarm optimization, fuzzy neural networks, or the combination of these methods.
Progress/Completion Report, 2011, PDF