WEFTEC®.09

Predictive Source Control: Application of Advanced Numerical

Methods and Surrogate Monitoring for Microconstituents in Water

Reclamation

Christopher Stacklin, P.E.

Orange County Sanitation District, 10844 Ellis Avenue, Fountain Valley, CA 92708-7018

*Email: cstacklin@ocsd.com

ABSTRACT

Having the water reclamation capacity of 70 million gallons per day (MGD) and expandable up

to 130 MGD, the Ground Water Replenishment (GWR) System is the largest water purification

and reuse project of its kind in the world. With the advent of groundwater replenishment and

reuse, a new paradigm of source control has started at the Orange County Sanitation District

(OCSD). Traditionally, Source Control was primarily based on enforcement and compliance of

the EPA’s National Pretreatment and NPDES Programs. At OCSD, Source Control has

expanded its role to monitoring and controlling microconstituents from both point and nonpoint

sources using two innovative methods: 1) Predictive Modeling and 2) Real-Time Modeling.

As part of its expanded program to assure that the water produced by the GWR System is of the

highest quality, Source Control has developed key programs for both predictive and real-time

monitoring of microconstituents in source water received by the facilities. Information from

predictive and real-time monitoring can be used to alert facility operations of abnormalities such

as microconstituent concentration spikes, growth trends, or reductions. The information is also

used as triggers for implementation of various point and nonpoint source control measures to

reduce or smooth concentration spikes.

Predictive Modeling

Predictive modeling is based on a stochastic, time-series approach using Kalman filters and

autoregressive integrated moving average methods such as Box-Jenkins. These models are

applied to microconstituent analytical data of 1) the source stream recognizing that there is a

high potential of outliers and noise due to matrix interference and signal suppression; and 2)

intermediate streams where noise due to matrix interference and signal suppression are

diminished but facility dynamics come into play.

Unlike conventional constituents where levels of concentrations are established or exhibit

discernable patterns, microconstituents may follow consumer trends that may be influenced by

competing new products, product replacement and substitution, health advisories, or consumer

news. Historical (past) data can potentially bias the results if it is not discarded after new

consumer trends emerge. Therefore, a recursive approach is necessary where the output is used

to condition the estimate and past data is discarded.

Results of the stochastic time series model are compared with economic indicators such as

pharmaceutical sales estimates or usage statistics, and heuristics based on expert opinion and

WEFTEC®.09

product literature. In this manner, the predictive model is tempered with real-life judgment.

Real-Time Modeling

Real-time modeling uses surrogates and analytical equipment to identify the magnitude of

concentration and upstream locations of point and nonpoint sources or in-plant. Statistical

methods are employed that test for correlation of the surrogate to speciated analytical data. If the

surrogate concentration increases, it may indicate an increase in concentration of a family of

microconstituents. The use of surrogates, in effect, reduces the time and cost of tracking

individual microconstituents. Control and action levels for surrogates are set by a mass balance

around the facilities using regulatory limits, standards, or guidelines and removal efficiencies

derived analytical values of microconstituents and stream flow rates.

Predictive and real-time models for microconstituents will be developed based on analytical

results and flows from the GWR System and Reclamation Plant No. 1. The results will be

discussed relative to practical (day-to-day) implementation and use.

KEYWORDS: Microconstituents, Groundwater Replenishment, Surrogates, Predictive

Modeling, Real-Time Modeling, Kalman Filer, Robust Regression on Order, Dynamics.

THE CHALLENGE – THE NEED FOR REAL-TIME DETECTION

Microconstituent Monitoring Logistics and Frequency

Microconstituents, also known as emerging pollutants of concern, or micropollutants, etc. occur

at relatively low concentrations, tend to be both ubiquitous and persistent in the environment,

and have known to unknown toxicity effects.

At water reclamation facilities, there is an emphasis on characterizing microconstituents in the

influent streams received by the facilities for the purpose of meeting operational objectives and

implementing effective source control of microconstituents being discharged into the

wastewater.

Based on today’s analytical technology, monitoring frequency of microconstituents is a balance

between regulatory requirements, logistics, and economics. Sampling for microconstituents is a

batchwise versus continuous process and consists of collecting a representative sample, then

analyzing it, and reporting results.

Due to practicality and economics, microconstituent sampling is typically taken infrequently

whether it is on a monthly, quarterly, or annual basis. This batchwise mode leaves data gaps in

the time in-between sampling when there are no samples taken and therefore no analytical results

reflecting microconstituent concentrations entering the treatment facilities.

The time from sample collection to reporting results can also be significant. At OCSD, the

average turnaround time for analytical data is 36 days from the time that a sample is collected to

the time that analytical data is disseminated, although there are ways to expedite turnaround

WEFTEC®.09

times.

Implications of System Dynamics

In consideration of the sampling frequency and turnaround time, if a significantly high

microconstituent concentration is detected in a sample, what is the available response time to

adjust operations or implement source control measures?

The response time can be defined as the time that a microconstituent would take from its point of

discharge, through the collection system, then to the start of treatment. As an example, a

microconstituent would flow through the: 1) regional collection system, through 2) Reclamation

Plant No. 1 treatment processes, then through 3) microfiltration before it reaches reverse osmosis

and advanced oxidation where it would be removed or destroyed.

Hydraulic detention time in the regional collection system within Orange County can range from

instantaneous to about 8 hours, averaging 4 hours. Next, wastewater from the collection system

is received by Reclamation Plant No. 1. At Reclamation Plant No. 1, the hydraulic detention

time ranges from 4 hours to 8 hours, depending on the diurnal flow received by the plant.

Diurnal flow is the daily variation in flow received by the treatment facilities. The change in

flow rate received by the treatment facilities is influenced by the water use habits of 2.5 million

residents in Orange County. During the day, people work and tend to use more water than

during the night when they are sleeping. Also, heavy industrial water users operate in the day.

The flow variation at OCSD typically ranges from a 116 MGD peak at about 8:00 am to a 40

MGD trough at about 4:00am. Finally, hydraulic detention time across the GWR System

upstream of reverse osmosis is 40 minutes based on an estimated 2 million gallon detention

capacity at the current rated capacity of 70 MGD.

Thus, the average time it takes for a chemical constituent to reach the treatment facilities is 8½

hours versus the analytical processing time of 36 days. If sampling is done on a daily or weekly

basis, then the time lapse that a potential problem can be detected ranges from 15½ hours to up

to 7 days.

A Logistical Conundrum

Say that 1,4-dioxine is accidentally being discharged into the collection system due from a small

leak which is undetected at the source. If 15.5 hours up to seven days elapse before the

concentration spike is detected, and assuming that it is at the action level of 3 parts per billion,

then the potential additional mass reaching treatment is 1 pound to 7.3 pounds, respectively.

If the treatment system is operating at its design capacity, then there is risk of breakthrough.

Although 7.3 pounds is very small compared with a reservoir which may contain 200,000 acre

feet of water, drinking water quality regulations typically are concentration -based rather than

mass -based.

WEFTEC®.09

SHIFTING FROM REACTIVE TO REAL-TIME, PREDICTIVE SOURCE CONTROL

Traditional sampling is batchwise and requires turnaround time from sample collection to

analysis that typically exceeds the hydraulic detention time of the water treatment technology

and collection system. Recent technology reduces turnaround time from days and weeks to

almost instantaneous and continuous measurement.

Analysis in Real-Time

New analytical equipment is available which allows on-line, in stream and continuous sampling

for parameters including total organic carbon (TOC), conductivity, pH and temperature. TOC is

an indicator of the presence of organics, although it lacks speciated data to define exactly what

constituent was detected which would narrow down the sources of discharge. Conductivity is an

indicator of the presence of salts such as metals. pH indicates presence of acids or bases.

Temperature measurement is required for pH correction. These real-time and continuous

parameters help to quickly identify new trends such as an increase in concentration or narrow

down the source of the discharge.

Conceptually, these parameters which detect in real-time function as surrogate indicators for

classes of chemical compounds. A surrogate is something to take the place for another thing, in

this case, TOC will take the place of speciated analytical data for organic compounds. An

indicator is a measure of difference between a reference or standard and the observation, such as

an organic chemical grouping versus TOC which serves as a standard.

A challenge to this technology is that it is less effective in a wastewater matrix. A wastewater

matrix is both corrosive and fouling, requiring constant maintenance. Also, the wastewater

matrix tends to cause analytical interference and signal suppression.

Analytical equipment can be deployed in strategies areas of the treatment technology unit

processes and collection system to reduce the time for detection. Further, analyzing key

parameters which may serve as surrogates will also reduce the resources and cost required to

maintain such programs. For example, TOC analysis can be routinely used to detect significant

changes in organics concentration instead of running a detailed characterization for volatiles and

semi volatiles, and base/neutral/acid organic compounds. If an excursion is observed in the TOC

analysis, then a detailed characterization can be performed to identify significant compounds.

Characterization is paramount to determining the root cause of the excursion. When analytical

concentration is combined with flow data, the mass flux of a constituent entering the system can

be assessed. Also, by deploying the system in strategic areas and utilizing a geographic

information system (GIS), the root cause may be quickly isolated to a local area.

Forecasting Micropollutant Concentrations

Since real-time monitoring can detect general changes in parameters instantaneously, recalling

that the dynamics of the regional collection system ranges from instantaneous near the treatment

WEFTEC®.09

facilities to an average of 4 hours, real-time modeling is still more of a reactive element versus

pro active. Therefore, to have a pro-active program, making forecasts from the analytical

concentration data is necessary.

Advanced numerical methods have been used as a basis for forecasting in the past. Challenges

with respect to predicting concentrations of microconstituents are: 1) the presence of interference

and noise in the data; 2) the impact of left-censored data on the prediction; 3) outliers and 4) the

impact of past data that is no longer relevant. Following is the method development for the

predictive source control model deployed at OCSD.

METHODOLOGY AND DESIGN BASIS DEVELOPMENT

Numerical methods used to model microconstituent concentrations were carefully chosen based

on defining a detailed design basis.

Describing the Water Reclamation Domain Space

Why model the influent concentration to a wastewater treatment plant (WWTP) as a stochastic

time series process?

Univariate Time Series

By definition, a time series is an ordered sequence of observations (analytical measurements)

made over a continuing length of time. In the case of WWTP’s, influent concentrations of

microconstituents such as 1,4-dioxane are periodically and continuously measured. If the

sequence of observations is comprised of a single set of numbers, then the time series is

univariate. If the sequence of observations is n-dimensional vectors, then it is multivariate, and n

is the dimensionality of the time series. For example, consider the time series dataset of a single

constituent, ammonia in Table 1. The data comprises a univariate time series where { ,

,...,  } = {30.1, 26.4, ... , 30.6}.





 

Table 1. Ammonia average daily concentration, mg/L.

2/18/2009 2/19/2009 2/20/2009 2/21/2009 2/22/2009 2/23/2009 2/24/2009

30.1 26.4 30.8 30.8 30.5 28.6 30.6

Data is from Phase IV sampling data.

An example of a multivariate time series would be one comprising not just ammonia

concentrations, but nitrates, and nitrites concentrations as well. This would be a 3-dimensional

time series where for i number of microconstituents, a datapoint observation would be

represented as,  . This paper will develop a univariate time series model, for simplification.

,

Discrete Stochastic Process

Observations that generated a time series, such as microconstituent concentrations will continue

into the future. Future microconstituent concentration values are of interest, and are treated as

random. Therefore, to model these values, a model called a stochastic process based upon the

time series is used. The word stochastic means random.

WEFTEC®.09

In probability theory, a stochastic process is the counterpart to a deterministic process. A

deterministic process deals with only one possible outcome at a future time based on an initial

condition. In a stochastic process, an outcome may be one of many possible outcomes which can

be described by a probability distribution. This means that even if the initial condition such as a

past observation of microconstituent concentration is known, there are many possibilities that the

future concentration may be at, but some paths are more probable and others less.

A stochastic process is a set of random variables or random vectors ordered with respect to time

t. If t takes on integer values, the process is a discrete-time or discrete process. If it takes on real

values, it is a continuous-time or continuous process. Since microconstituent sampling is a

batchwise versus continuous process, the domain is therefore a discrete time-series stochastic

process.

Time series analysis is the fitting of stochastic processes to the time series. This typically

involves statistical analyses, but it is not a straightforward application of statistics. Examples of

processes modeled as stochastic time series include stock market and exchange rates, predicting

weather, GPS navigation and autopilot, robotic vision, audio and video signals, etc.

Stationarity

Another criterion is if microconstituent concentration values represent a stationary process or

non stationary process. A stationary process is a stochastic process whose joint probability

distribution does not change when shifted in time or space. As a result, parameters such as the

mean and variance, if they exist, also do not change over time or position. For example,

analytical matrix interference is stationary, but the sound of an echo is not because it diminished

over a period of time.

Examples of discrete-time stationary processes with continuous sample space include

autoregressive and moving average processes which are both subsets of the autoregressive

moving average model, Markov chains, and the Kalman filter.

To sum up the water reclamation domain space, water quality or more specifically

microconstituent concentration can be described as a univariate time series, discrete stochastic

process with stationarity. Therefore, a Kalman filter is proposed as a predictive model.

Kalman Filter

What is a Kalman filter? A Kalman filter is a recursive filter that estimates the state of a linear

dynamic system from a series of noisy measurements (Welch Greg, et. Al., 2006). Signal noise,

for example, is fluctuation and external factors added to a datastream signal received by a

detector, such as an analyzer. The receiving device can also be source of more signal noise.

Data is corrupted with this noise. Kalman filters are a means to filter noise from data. A good

filtering algorithm can filter the noise from a datastream while still retaining useful information.

Rudolph E. Kalman is credited in publishing the algorithm and approach in 1960 in his landmark

paper, “A New Approach to Linear Filtering and Prediction Problems” (Kalman, R.E. 1960).

WEFTEC®.09

Kalman filters were initially used largely for data compression algorithms in weather satellite

communications in the early 1960’s. Since then, Kalman filters are commonly used as predictors

for the stock market and econometrics, robotic navigation and visualization and computer

animation, to name a few.

The Kalman filter is ideally suited for stochastic, time series systems or domains, which is why it

can be applied to filtering the noise in analytical data of microconstituents (Stacklin, Christopher,

2008)(Stacklin, Christopher and Evangelista, Jerry, 2008). Contributors to noise in the analytical

data are variations in analytical methodology, matrix interference and signal suppression.

External factors are changes in influent flow rate, accidental or illegal discharges of chemicals,

and behavioral factors of dischargers. A potential stumbling block to the application of Kalman

filter or any other time series analysis is that the analytical data of microconstituents is left-

censored.

Non Detect and Data Analysis

Analytical data frequently are left-censored due to detection limits of laboratory methods. Left-

censored means that some of the observations are known only to fall below a censoring point

commonly known as a method detection limit or reporting limit, where the concentration is

reported as a non detect. A non detect value simply means that the concentration of a

microconstituent may be at or below the method detection or reporting limit. It can even be zero.

As an example, a non detect value of <2.0 mg/L means that the concentration is at or between

zero and 2.0 mg/L. This presents difficulties in statistical analysis of the data.

In the past, people have been dealing with non detect values by simply substituting the non

detect with one half or the full value of the method detection or reporting limit (Helsel, Dennis

R., 2006b). Since microconstituents by definition are detected at relatively low concentrations,

substitution can bias the results significantly (Helsel, Dennis R., 2005c). Fortunately, there are

several methods available for dealing with non detect values that will minimize biasing the data.

What are among the best methods for handling non detect data for microconstituent datasets?

Estimating descriptive statistical methods are Kaplan-Meier (KM), Robust Regression on Order

Statistics (ROS), and Maximum Likelihood Estimate (MLE) (Helsel, Dennis R., 2005a)(Helsel,

Dennis R., 2005d) (Singh, Anita and Nocerino, John, 2006). Note that descriptive statistics are

distinguished from inductive statistics in that they aim to quantitatively summarize a data set,

rather than being used to support statements about the population that the data are thought to

represent. Key attributes of these methods are: 1) the KM is a non parametric method and does

not yield values for non detects; 2) Robust ROS and MLE are parametric methods; 3) MLE

requires larger datasets then KM and Robust ROS for accuracy.

For time series algorithms such as Kalman filters to work properly, non detects must be

converted to values, therefore parametric methods are required. Therefore, KM cannot be used,

leaving Robust ROS and MLE. Since Kalman filters use small datasets to reinitialize or restart

WEFTEC®.09

due to trend changes and microconstituent datasets by nature will tend to be smaller, the method

selected for data preprocessing for the Kalman filter is Robust ROS.

Robust Regression on Order Statistics

An ordinary least squares (OLS) regression estimate assumes that the error term has a constant

variance or that observations are drawn from identical distributions. If the error term varies with

each observation as is typically the case with time-series measurements, then this reflects that the

random variables have different variances. Outliers can have the same influence. Therefore the

domain is heteroskedastic.

Heteroskedasticity does not cause OLS coefficient estimates to be biased nor inconsistent.

However, the variance (and, thus, standard errors) of the coefficients tends to be underestimated,

inflating t-scores and sometimes making insignificant variables appear to be statistically

significant. To circumvent underestimation, robust regression is required (Helsel, Dennis R.,

2005a).

Implementation of Robust ROS

Details of Robust ROS calculations are described in, Nondetects and Data Analysis: Statistics for

Censored Environmental Data (Helsel, Dennis R., 2005a). The basic steps to implement Robust

ROS are to rank, test, and convert the data for lognormal, normal, or root mean square

distribution (Shumway, et. Al. 2002); determine the probability of non detects and regress their

values from the distribution (Helsel, Dennis R., 2005a). When the values of the nondetects are

defined, the dataset can be then processed by the Kalman filter.

Implementation of the Kalman Filter Model

Note that the Kalman filter algorithm herein includes zeroth, first, and second order filters. A

zeroth order filter is good for smoothing the data, e.g., removing the noise to discern general

trends. The second order filter is better a predicting values. A second order Kalman filter with

restart will be even better at predicting values, but will be described in another paper.

The measurement of the variable, 



represents a measured concentration value of a

microconstituent. The Kalman filter for sequential least squares estimating (Sorenson, H.W.,

1970) starts with the definition of the minimum mean square estimator,



|



|





󰇛









|

󰇜



|



|













(1)

Where

= Minimum mean square estimate

= Predicted estimate

= Gain matrix

= Measurement of the variable

= Observation matrix

k = Measurement number

WEFTEC®.09

The gain matrix is defined by,



|



|













|









|







|









|



|



|



|













|









|











|





|



|



|



|



|



|

|

󰇟

󰇠

󰇣

󰇤



110

011

001





󰇟

󰇠

󰇟

󰇠

󰇟

100

󰇠

 











∑󰇛









󰇜









(2)

Where

= Covariance of error in the predicted estimate

= Transpose of the observation matrix

= Variance of values from the estimate

Covariance of the estimator error or matrix inversion lemma,

(3)

Where

= Covariance of error in the estimate

Q = Model variance error

The predicted estimate can be found using the transition matrix such that,

(4)

Where

= Transition matrix

= Previous best estimate

The transition matrices,  for the zeroth, first and second order filters are,

(5)

(6)

(7)

Corresponding observation matrices,  for the zeroth, first and second order filters are,

(8)

(9)

(10)

To initialize the estimator,  is the median of the variances (bias) of n joint observations. 

will remain constant for any k. The sample variance (bias) of n measurements is,

(11)

Where

WEFTEC®.09



∑











(12)

he previous cova

(13)

Covariance of erro

(14)

The predicted estimate

(15)

After initialization, the fu

ression on Order

Figure 1 shows the results of the application of Robust ROS to the Estriol dataset. Note that

rmone therapy. Out of the 35 datapoints, 13 are non detects

T riance of error in the estimate



|







′



r of the predicted estimate is,



|



|





is,



|



|



|

ll equations may be used.

RESULTS

Robust Reg

Estriol is used for estrogenic ho

which represents 13% of the data. The red line shows the regressed non detect values.

Figure 1. Robust ROS results for Estriol.

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.5

0.45

Concentration,μg/L

Estriol NDRegress

WEFTEC®.09

Summary statistics are presented in Table 2. Note that substitution of ½ of the method detection

limit and maximum likelihood estimate indicate poor results. In this case, the maximum

likelihood estimate is off due to the small number of observations. The Kaplan-Meier statistics

look good and is useful for small datasets. Unfortunately, Kaplan-Meier is non parametric, e.g.,

non detect values cannot be derived from the method.

Table 2. Summary statistics for Estriol.

Method Mean Std. Deviation UCL 95%

Detected Values 0.298 0.099 - - -

Substitution ½ Detection Limit 0.191 0.162 0.237

Kaplan-Meier 0.228 0.119 0.263

Robust ROS 0.232 0.118 0.264

Maximum Likelihood Estimate 0.140 0.231 0.206

Figure 2 shows the results of the application of Robust ROS to the Bisphenol A dataset. Note

that Bisphenol A is a chemical building block that is used primarily to make polycarbonate

plastic and epoxy resins. Out of the 35 datapoints, 6 are non detects which represents 17% of the

data. The dotted red line shows the regressed non detect values.

igure 2. Robust ROS results for Bisphenol A.

Concentration,μg/L

BisphenolA NDRegress

WEFTEC®.09

Summary statistics are presented in Table 3. Again, substitution of ½ of the method detection

limit and maximum likelihood estimate indicate poor results. The Kaplan-Meier is better,

followed by Robust ROS statistics.

Table 3. Summary statistics for Bisphenol A.

Method Mean Std. Deviation UCL 95%

Detected Values 1.283 1.248 - - -

Substitution ½ Detection Limit 1.065 1.232 1.417

Kaplan-Meier 1.113 1.177 1.455

Robust ROS 1.093 1.209 1.437

Maximum Likelihood Estimate 0.921 1.400 1.321

Kalman Filter

Kalman filter results are presented below. The Kalman filter results presented are based on a

second order filter with no reset. Microconstituent concentrations (observed values) are

indicated by a diamond, . Predicted values are indicated by a red line, .

Estriol is shown in Figure 3 with the residual sum of squares (RSS) value at 0.06. The RSS is a

measure of the discrepancy between the data and an estimation model. A small RSS indicates a

tight fit of the model to the data. The Kalman filter results are also shows for Bisphenol A with

the RSS of 4.75.

Figure 3. Kalman filter results for Estriol.

123456789101112131415161718192021222324252672

0.05

0.1

0.3

0.4

0.5

2 8293031323334

0.15

nce

0.2

rat

0.25

0.35

0.45

Co nt io µg/L

PredictedValues ObservedValues

WEFTEC®.09

Figure 4. Kalman filter results for Bisphenol A.

Kalman filter versus Moving Average

Figures 5 and 6 compare the Kalman filter results versus two other discrete time series methods,

simple moving average (simple MA) and exponential moving average (exponential MA), cousin

methods to ARIMA Box-Jenkins. Both methods are very fast and easy to implement in contrast

to the Kalman filter. However, the results in Table 4 show that the Kalman filter is a much better

fit in both cases.

Table 4. Comparison of RSS values of time series methods.

Microconstituent Simple MA Exponential MA Kalman Filter

12345678910111213141516171819202122232425262728293031323334

Concentration, µg/L

PredictedValues ObservedValues

Estriol 0.33 0.20 0.05

Bisphenol A 12.17 12.37 1.58

WEFTEC®.09

Figure 5. Kalman filter results for Estriol.

Figure 6. Kalman filter results for Bisphenol A.

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

1 2 3 4 5 6 7 8 9 101112131415161718192021222324

Concentration, µg/L

2526

ObservedValues SimpleMovingAverage

ExponentialMovingAverage KalmanFilter

0.5

2.5

1 2 3 4 5 6 7 8 9 1011121314151617181920212223242526

Concentration, µg/L

1.5

ObservedValues SimpleMovingAverage

ExponentialMovingAverage KalmanFilter

WEFTEC®.09

TOC as a Surrogate Indicator

It was originally proposed that TOC be used as a surrogate indicator for microconstituents.

However, the concentration of TOC at the influent to Reclamation Plant No. 1 averages 80 mg/L.

The high TOC concentration relative to the concentration of a microconstituent which may be

several orders of magnitude lower than TOC (in the µg/L to ng/L range) would make discerning

any relevance very difficult. For example, if the average microconstituent concentration was 50

µg/L, then a microconstituent would constitute about six onehundreths of a percent of the TOC.

Figure 7. TOC at Reclamation Plant No. 1 versus GWR System influent.

In contrast, the TOC level of the secondary effluent to the GWR System is much lower and

averages below 20 mg/L. In this case, an average microconstituent would constitute about one

quarter of a percent of the TOC. The wastewater matrix is also much cleaner. Therefore, it may

be possible to correlate microconstituent values with TOC based on downstream locations such

as immediately upstream of reverse osmosis.

CONCLUSIONS AND DISCUSSIONS

Real-Time Monitoring

over batchwise sampling, the advantage

100

120

1234567891011121314151617181920

Concentration, mg/L

ReclamationPlantNo.1Influent GWRSystemInfluent

While real-time monitoring has a distinct advantage

WEFTEC®.09

gained is limited to the dynamics of the collection system and the treatment system which may

tive

location allows for a cleaner matrix to analyze and the TOC value is proportionally closer to the

sum of the microconstituent values.

TOC values can be collected in real-time. By converting the TOC continuous telemetry data to

discrete data, a Kalman filter can be used to predict future concentrations. If the TOC

concentration is predicted to rise significantly, then the TOC is required at the influent to the

plant to determine a mass-based removal efficiency. If the removal efficiency is negative, then

the TOC is being added or created inside of the treatment processes. If the removal efficiency is

positive, then the TOC is being added from the collection system.

TOC analyzers coupled with Kalman filters can be located on the regional trunk lines and pump

stations in strategic locations allowing geographic localization of the potential cause of the

fluctuation. If the TOC concentration is uniformly increased, then the source is ubiquitous.

Therefore a general outreach program, or certification program, or product ban may be

considered to achieve a reduction. If the source is localized, then a targeted outreach program,

certification, or permitting of a point source or commercial market sector may be considered.

The TOC results can be integrated with a Geographic Information System and Chemical

Inventory Program to localize the source quickly.

Predictive Kalman Filter

Based on the domain space definition, a Kalman filter was proposed and demonstrated as a

model to predict future microconstituent concentrations. It is important to note that the predicted

concentration must be converted to a mass value using flow rate in order to have real relevance.

entration of a

icroconstituent which is approaching an action level, say 20 ng/L equates to 12 pounds per day

a 70 MGD treatment facility. Finding a 12 pound per day source within the treatment facility

puts

forecast by rationalization. This can be done by

nds, marketing growth cycle, publicity and most of all, common sense.

cs.

average only 8 ½ hours. When real-time modeling is coupled with Kalman filters, the predic

model can extend upward to a week or more, depending on discharge habits.

It may be possible to use TOC as a surrogate indicator, e.g., correlate TOC to microconstituent

values, but immediately upstream of reverse osmosis and advanced oxidation processes and

downstream of primary and secondary treatment. Placing in-stream TOC monitoring at this

A mass value establishes the magnitude of the condition. For example, a conc

or being discharged into the collection system can be very challenging. Hence the mass rate

things into perspective.

The Kalman filter predictor should only be used for near-term prediction initially until a

thorough understanding of the behavior of the dynamics of the system is gained.

Tempering Prediction Using Heuristics

It is very important to temper any mathematic

considering economic tre

These are experience-based techniques that help in problem solving and are known as heuristi

WEFTEC®.09

Complex rules can be chained together using a Bayesian decision model to form an assessment

or decision (Winkler, Peter, 2000).

Figure 8 is based on a simple, quick and dirty econometric forecast for the discharge of Estriol in

tion

pact of the FDA’s position on the benefits of hormone

main Space

Orange County. The econometric model considers economic criteria to determine Estriol

demand. The waste mass load of Estriol received at OCSD was allocated on a per capita basi

based on the female population in Orange County at the time of sampling. The female popula

growth forecast in Orange County is factored in as well as the gross domestic product growth (or

decrease) and an assessment of the im

treatment theory drugs.

Figure 8. Econometric forecast of Estriol in wastewater discharged in Orange County.

CONCLUSIONS

Water Reclamation Do

14.00

14.10

Estriol

13.40

13.50

08Q1 08Q2 08Q3 08Q4 09Q1 09Q2 09Q3 09Q4 10Q1 10Q

13.60

13.70

13.80

13.90

Pounds Discharged Per Quarter

Water quality or more specifically microconstituent concentration can be described as a

univariate time series, discrete stochastic process with stationarity.

Real-time Modeling

Real-time modeling of water reclamation treatment processes and is possible with recent

technology innovations, but the benefits do not allow enough lead time for pro-active operational

and source control programs. Real-time modeling can provide a finite amount of lead time

WEFTEC®.09

which could range from hours to as much as a day depending on system dynamics.

Predictive Modeling

It was demonstrated that a Kalman filter can be used to provide near-term prediction of

microconstituent concentrations ranging from days to weeks depending on the behaviors o

dischargers.

Robust ROS, an estimating descriptive statistical method, was used to regress values for non

detects in the analytical datasets while minimizing any biasing of the descriptive statistics.

Results from the Kalman

filter predictive model can be tempered using an econometric, heuristic

model that describes local demographic trends and subjective impacts, e.g., consumer news,

regulations, etc. Figure 9 shows the predictive model.

oconstituent concentration and source control.

owing people. Without their support, this paper would not be possible:

Figure 9. Predictive model for micr

ACKNOWLEDGMENTS

I wish to thank the foll

Jerry Evangelista, OCSD Source Control Supervisor; Mahin Talebi, OCSD Source Contro

Manager; Steve Fitzsimmons, OCWD Lab Director; and my wife, Lauren Viscardi.

Proactive

Action Plan

Predict Mass

Flux Trends

Regress Non

Detects

Robust ROS

Kalman

Filter

Flow Data

Mitigative

Action

Operations

Source

Control

Analytical

Data

Econometric

Model

WEFTEC®.09

REFERENCES

Al. (1998). Control of sewer systems and wastewater treatment plants

–93.

em,

Water, 1(1), 71–78.

Bechmann, Henrik (1999). Modeling of Wastewater Systems, ATV Erhvervsforskerprojekt EF

John Wiley & Sons, 1968; 2nd Edition, AMS Chelsea Publ., 2005.

isk From Pharmaceuticals in Drinking

, C-07-021, WA-B-02, Task 6, U.S.

Environmental Protection Agency: Washington, D.C.

Dearden, John C., et. Al. (1989). The Development and Validation of Expert Systems for

Predicting Toxicity, The Report and Recommendations of an ECVAM/ECB Workshop

(ECVAM Workshop 24), ECVAM, TP 580, JRC Environment Institute, 21020 Ispra

(VA), Italy.

Grewal, Mohinder S. and Andrews, Angus P. (2008). Kalman Filtering: Theory and Practice

Using MATLAB, John Wiley and Sons; 3rd ed., Hoboken, New Jersey, 592p.

Hajek, Bruce (2009). An Exploration of Random Processes for Engineers, Notes for ECE 534,

University of Illinois at Urbana-Champaign, Champaign, Illinois, 373p.

Helsel, Dennis R. (2005a). Nondetects and Data Analysis: Statistics for Censored Environmental

Data. John Wiley and Sons, New York, 250p.

Helsel, Dennis R. (2006b). Fabricating Data: How substituting values for nondetects can ruin

results, and what can be done about it. Chemosphere 65 (11), 2434-2439.

Helsel, Dennis R. (2005c). Insider Censoring: distortion of data with nondetects. Human and

Ecological Risk Assessment 11, pp. 1127-1137.

Helsel, Dennis R. (2005d). More Than Obvious: Better methods for interpreting nondetect data.

Environmental Science and Technol. 39 (20), p. 419A–423A.

Houseman, E. Andres (2004). A Robust Regression Model for a First-Order Autoregressive

Time Series with Unequal Spacing: Technical Report, Harvard School of Public Health,

ahousema@hsph.harvard.edu, 18p. Retrieved on 2008-04-23.

Kalman, R.E. (1960). "A new approach to linear filtering and prediction problems". Journal of

Basic Engineering 82 (1): 35–45. Retrieved on 2008-05-03.

Shumway, R.H., Azari, R.S., and Kayhanian, M., (2002). Statistical Approaches to Estimating

ce and

ohn (2006). Robust Estimation of Mean and Variance Using

ets with Below Detection Limit Observations,

tacklin, Christopher (2008). Pollutant Prioritization Project for Water Reuse, 81st Annual

Bechmann, Henrik, et.

using pollutant concentration profiles. Water Science and Technology, 37(12), 87

Bechmann, Henrik, et. Al. (1999). Grey-box Modeling of Pollutant Loads from a Sewer Syst

Urban

623 IMM-PHD-1999-69, Technical University of Denmark, and Krüger A/S.

Bucy, R.S. and Joseph, P.D., Filtering for Stochastic Processes with Applications to Guidanc

Chui, Charles K. and Chen, Guanrong (2009). Kalman Filtering: with Real-Time Applications,

Springer-Verlag, 4th ed., Belin Heidelburg, Germany, 229p.

Conerly, Octavia (2008). Approaches to Screening for R

Water and Prioritization for Further Evaluation

Mean Water Quality Concentrations With Detection Limits, Environmental Scien

Technology 36, 3345-3353.

Singh, Anita and Nocerino, J

Environmental Data S

Sorenson, H.W. (1970). Least Squares Estimation: From Gauss to Kalman, IEEE Spectrum, vo

7, July, 1970, pp. 63-68.

Sorenson, H. W. (1980). Parameter Estimation, Principles and Problems, Mercel Dekker, Ne

York, pp. 56-60.

WEFTEC®.09

ironment Federation Technical Exhibition and Conference (WEFTEC 2008),

Session 15, McCormick Place, Chicago, Illinois, October 20, 2008.

08.

6, pp. 1-16.

s, Canberra,

Winkle lications of Threshold

Zarcha

American Institute of Aeronautics

Water Env

Stacklin, Christopher and Evangelista, Jerry (2008). Technical Risk Assessment of Water Reus

in Consideration of Emerging Pollutants, presented at the 2008 International Water

Conference, Crowne Plaza - San Antonio Riverwalk, San Antonio, Texas, paper no.

IWC-08-69, October 29, 20

Welch, Greg and Bishop, Gary (2006). An Introduction to the Kalman Filter, TR 95-041,

Department of Computer Science, UNC-Chapel Hill, TR 95-041, July 24, 200

Whitehead, P. G. (Unknown). Water Quality Models for Waste Water Management, Australian

National University, Centre for Resource and Environmental Studie

Australia. 421-430

r, Peter (2000). Optimization Heuristics in Econometrics: App

Accepting, John Wiley and Sons, West Sussex, England, 1st ed., 320p.

n, Paul and Musoff, Howard (2005). Fundamentals of Kalman Filtering: A Practical

Approach (Progress in Astronautics and Aeronautics),

& Astronomy; 2nd ed., 764p.