Surface-water quality and flow Modeling Interest Group
Real-Time Control for Matching Wastewater Discharges to the Assimilative
Capacity of a Complex, Tidally Affected River Basin
by Edwin A. Roehl Jr.1
and Paul A. Conrads2
1Vice President - Systems Development
OptiQuest Technologies, LLC
214 Pelham Davis Circle
Greenville, SC 29615
Internet: info@oqt.com
Phone: (864) 987-0717
FAX: (864) 234-7521 |
2Hydrologist
USGS, Water Resources Division
Stephenson Center, Suite 129
720 Gracern Road
Columbia, SC 29210-7651
Internet: pconrads@usgs.gov
Phone: (803) 750-6140
FAX: (803) 750-6181 |
Editor's note:
This paper was written for the 1999 South Carolina Environmental Conference,
held in Myrtle Beach on March 15-16, 1999.
Citation:
Roehl, E.A. and Conrads, P.A., 1999, Real-time control for matching wastewater
discharges to the assimilative capacity of a complex, tidally affected river
basin, in Proceedings of the 1999 South Carolina Environmental
Conference, March 15-16, 1999, Myrtle Beach, SC.
Contents
A neural network model was applied to simulate the hydrodynamics and water
quality of the Cooper and Wando Rivers in South Carolina. The evaluation of
the model showed that predictions of salinity, water temperature, and
dissolved-oxygen concentration for this complex estuarine system were
accurate. Because neural network models execute without iteration, they are
ideal for integrating with real-time information and control systems. In
this study, the neural network model of the Cooper and Wando Rivers was
coupled with an optimization routine to make maximum use of the assimilative
capacity of the two-river system. Target dissolved-oxygen concentrations, set
at the State water-quality standard, were matched by constraining effluent
discharges. A prototype real-time control system for matching wastewater
discharges to the continuously changing assimilative capacity of the Cooper
and Wando Rivers is presented.
In 1997 the U.S. Geological Survey (USGS) participated in evaluating the use
of neural network models for simulating natural systems, and compared the
approaches and accuracy of neural network models and first-principles-based
finite-difference models that had recently been developed for the Cooper and
Wando Rivers, South Carolina (Conrads and Smith, 1996, 1997) (Figure 1). The
finite-difference models predicted the hydrodynamic, mass transport, and
water-quality phenomena of a system that is particularly difficult to model
because of tidal effects, variable freshwater releases from a hydroelectric
plant, large amounts of poorly defined overbank storage with unmeasureable
flows, and wastewater discharges from municipal and industrial facilities.
To support the modeling effort, a system of gaging stations is operated by
the USGS that collects, via satellite, real-time measurements of water level,
dissolved-oxygen concentration (DO), water temperature, and salinity. These
data were used to develop both sets of models.
Figure 1. Charleston Harbor and its tributaries.
The results of the comparison (Conrads and Roehl, 1999) showed that the
neural network models gave more accurate predictions of the modeled
variables, and that they required about 90% fewer person-hours to develop. An
additional benefit is that trained neural network models can be deployed as
compact programs that execute without iteration, making them suitable for
integrating with real-time information and control systems. This paper
describes a prototype real-time control system for matching wastewater
discharges to the continuously changing assimilative capacity of the Cooper
and Wando Rivers. Simulations, which demonstrate how the control system
would predict assimilative capacity and modulate multiple discharge streams
to avoid violating water-quality standards, and a deployment strategy for the
control system are described.
The literature describes many uses of neural network models to monitor and
control industrial processes. They are most often used as "soft sensors" to
estimate quantities that cannot be measured directly, and as process models
in model-based control schemes. Figure 2 shows an idealized implementation of
a neural network-based process model for controlling DO in water. The input
variables are of two types, state variables and manipulated variables, e.g.,
discharge levels. It is assumed that changes in the inputs will change the
model's outputs, and that these changes are representative of the physical
system's behavior. If the model predicts an undesirable outcome, e.g., DO
below a standard, a set of values for the manipulated variables may exist
such that the model will predict a more desirable outcome.
Figure 2. Control scheme Using a neural network model with an
optimization program
Finding values for the manipulated variables so that an undesirable outcome
can be avoided requires the use of an optimization program. As state
variables change with time, the optimization program will search for
manipulated variable values that allow the model to avoid an undesirable
outcome. The optimization program adheres to specified "constraints" which
place limits on values the manipulated variables can have. If it fails to
obtain an acceptable outcome within a specified number of iterations, the
program will return the values that provides the prediction that least
deviates from a desirable outcome. Effectively, the optimization program
"inverts" the model, so that by automatically adjusting the manipulated
variables a specified outcome can be obtained.
An on-line control system would necessarily be more complicated than the
scheme shown in Figure 2 because of the following constraints:
- Incoming real-time signals have to be conditioned to accommodate
sensor failures. The signal from a failed sensor can be detected and
reconstructed from correlated data by using specialized neural network
models.
- Biochemical reaction kinetics that depend on climatic conditions
require a probabilistic control strategy based on weather predictions.
Neural network models that integrate multiple real-time signals to
accurately predict future conditions are called "soft sensors."
Multiple soft sensors would allow alternative scenarios, graded by
likelihood of occurrence and degree of conservatism, to be evaluated
before control actions are taken.
- A model that is integrated with an optimization program to control a
process, called a "control model," must be tuned to insure that the
functional relationships (gains) between the manipulated and
controlled variables match physical reality; for example, DO falls by
a certain amount as discharged biochemical oxygen demand (BOD)
increases. Because of their different purpose, control models are
generally less accurate than soft sensors.
DO is the water-quality variable of primary interest to many water-resource
managers, and is affected by environmental conditions such as temperature,
rainfall, river flow, and tidal action. DO is also affected by the BOD of
organic material from rainfall runoff, tidal overbank storage, and wastewater
discharge streams. The largest municipal and industrial wastewater
dischargers lie on the west bank of the Cooper River in the vicinity of
station 021720675; therefore, the controlled variable chosen for the
simulation was the DO at this station.
The prototype control system consisted of two neural network models, a soft
sensor to give an accurate prediction of DO, and a control model tuned to
have appropriate gains between the wastewater BOD and the modeled DO. The
output of the control model was adjusted (biased) by the soft sensor prior to
running the optimization program at each time step. Both models were
configured to predict the DO at noon one day ahead of time. An on-line system
could use a series of predictions with staggered time horizons to provide
trend information that could be used in concert with a margin of safety to
maintain the minimum standard. The inputs to both models consisted of the
seven largest BOD and NH3 discharge streams and a number of gage
variables. The soft sensor predictions are compared to the actual DO in
Figure 3.
Figure 3. Actual data and soft sensor predictions (SS Pred) of DO at
station 021720675.
The control system simulation covered a period corresponding to the summer of
1994 when the DO fell below 5.0 mg/L. The control scheme was configured so
that the historical discharges would not be modified unless the soft sensor
predicted a DO below a threshold of 5.5 mg/L. In an on-line system, the
constraints under which the optimization program operates would reflect the
permitted wasteload allocations of the individual dischargers. The
optimization constraints used for this simulation were:
- a discharge could not be set to a value less than its historical
minimum, and
- a discharge could not exceed the average of its historical maximum and
its optimized value for the preceding day.
The latter allowed for a gradual redistribution among dischargers of the BOD
reductions.
Figure 4 shows the simulation during a 30-day period when the control scheme
was engaged. Two sets of predictions, computed using the same soft sensor,
but with different input discharge BODs, are shown. "SS Pred" uses the actual
BODs and "SS Op. Pred" uses BODs computed by the optimization program. The
control system approaches its target of 5.5 mg/L until day 290, when a severe
drop in the actual DO to as low as 3.2 mg/L could not be fully corrected.
However, SS Op. Pred predicts that the DO can be elevated to about 4.8 mg/L
through reductions in BOD discharges. The drop in the actual DO was due to a
combination of environmental conditions and discharge levels that have yet to
be studied in detail.
Figure 4. Simulation results at station 021720675 for days 280 to
310.
ABOVE: Actual (historical) and predictions of DO with control system
engaged.
BELOW: Actual (historical) and optimized discharges for dischargers
D1 and D2.
Figure 4 also shows the modulated discharges for the two largest wastewater
streams, D1 and D2. D1 is the most affected, being held at its historical
minimum of 2,197 pounds per day from day 301 onward when the actual DO
reached the lowest value in the entire time series shown in Figure 3. The
optimization program obtained the balance of the needed discharge reductions
from the five remaining discharge streams (not shown).
The control scheme could be implemented in a real-time system as shown in
Figure 5. The essential elements consist of the USGS gaging network, a
computer system that operates the control software, and an Internet web site
where real-time data, model predictions, and discharge recommendations are
made accessible to stakeholders in the protection and use of the Cooper and
Wando Rivers. System operation is as follows:
- Real-time data are collected by the USGS gaging network, processed,
and archived.
- Data are sent to a computer that runs the control system, predicts
water-quality trends, and recommends wastewater loadings.
- The real-time data and control system recommendations are posted to
the web site for review and use by regulators, dischargers, and other
organizations.
Figure 5. Schematic of Real-Time Control Scheme.
Previous work has shown that neural network models can be very effective for
modeling the complex water quality of the Cooper and Wando Rivers. The work
described in this paper shows how these models can be deployed in a real-time
control scheme that has the potential to match permitted wastewater
discharges to the changing assimilative capacity and hydrologic conditions of
the environment, thus avoiding or minimizing violations of the State
water-quality standard. An additional benefit for all concerned with the
protection and use of the Cooper and Wando Rivers would be derived from
providing real-time open access to the data, models, and control
recommendations via an online application.
Conrads, P.A. and Smith, P.A., 1996, Simulation of water level, streamflow,
and mass transport for the Cooper and Wando Rivers near Charleston, South
Carolina, 1992-95: U.S. Geological Survey Water-Resources Investigations
Report 96-4237, 51 p.
--- 1997, Simulation of temperature, nutrients, biochemical oxygen demand,
and dissolved oxygen in the Cooper and Wando Rivers near Charleston, South
Carolina, 1992-95: U.S. Geological Survey Water-Resources Investigations
Report 96-4151, 58 p.
Conrads, P.A. and Roehl, E.A., 1999, Comparing physics-based and neural
network models for predicting salinity, temperature, and dissolved-oxygen in
a complex, tidally affected river basin: 1999 South Carolina Environmental
Conference, Myrtle Beach, March 15-16, 1999.
Back to the SMIG Features Page
Home |
Mailing List |
Features |
Conferences |
Classes |
Reading |
Model Archives |
Feedback
Stewart Rounds, SMIG coordinator
<sarounds@usgs.gov>
U.S. Geological Survey
http://smig.usgs.gov/SMIG/features_0399/nnm2.html
Last modified Wednesday, 17-Dec-2003 14:06:58 EST
Privacy Statement ·
Disclaimer ·
FOIA ·
Accessibility
Abstract