American Journal of Environmental Engineering

p-ISSN: 2166-4633    e-ISSN: 2166-465X

2018;  8(5): 181-200

doi:10.5923/j.ajee.20180805.03

 

Can Prior Experience Provide a Means to Predict Success of Future Aquifer Storage and Recovery Systems?

Frederick Bloetscher

Florida Atlantic University, Boca Raton, FL, USA

Correspondence to: Frederick Bloetscher, Florida Atlantic University, Boca Raton, FL, USA.

Email:

Copyright © 2018 The Author(s). Published by Scientific & Academic Publishing.

This work is licensed under the Creative Commons Attribution International License (CC BY).
http://creativecommons.org/licenses/by/4.0/

Abstract

This paper is the result of analysis data gathered from a 2013 survey of all 204 Aquifer Storage and Recovery (ASR) sites in the United States. That 2013 ASR site survey included all active, inactive, test and study sites, and collected both operational and construction details. The differences between the operational and inactive sites are of particular interest because the differences are where the most information can often be gleaned as to the potential for success of the test and study sites. The statistical analysis utilized in this analysis focused on the active and inactive sites – all sites in study mode and early stages of development were not included in the initial analysis. The intent was to determine is a predictive model for ASR success could be developed for the test and study ASR sites, as well as potential future sites. The results improve on prior papers by the author related to ASR system success and provides insight on what factors improve the likelihood of successful ASR projects. Using the results of the PCA, a linear regression model was developed for the active and inactive sites, and applied to the test and study sites to predict their likelihood of success. The results provide insight into the potential for success in the 50+ test/study sites that may be years for full development.

Keywords: Groundwater Storage, Recharge, Predicting Success

Cite this paper: Frederick Bloetscher, Can Prior Experience Provide a Means to Predict Success of Future Aquifer Storage and Recovery Systems?, American Journal of Environmental Engineering, Vol. 8 No. 5, 2018, pp. 181-200. doi: 10.5923/j.ajee.20180805.03.

1. Introduction

Water supply challenges exist throughout the world. As a result, in drought or water limited areas, the ability to store water for later use has value for sustainability of the local community. AWWA Manual M21 [1] divides aquifer storage programs into four categories: Artificial Aquifer Creation, Aquifer Recharge, Aquifer Reclamation, and Aquifer Storage and Recovery (ASR). All of these approaches are used as part of the water supply industry to ensure that sustainable water resources are available for agricultural, environmental and urban uses. This paper focusses on the ASR portions only and utilizes the dataset developed in conjunction with AWWA Manual M-63 [2-4]. ASR is touted as a viable concept in the management of both potable and non-potable water supplies. Utilities pursue ASR programs to increase the efficiency of system operations to utilized unused water treatment plant capacity to treat water and pump it into an aquifer for later withdrawal for augmentation of water supplies at a later point of time to avoid the need to construct plants only for peak demands [2-4]. The injection applications include potable water, raw surface and groundwater, and reclaimed wastewater. The storage period can be over multiple months to allow the stored water to meet the next high demand season, an emergency such as a severe drought or during an interruption of water withdrawal due to equipment breakdown.
The concept of ASR has only been applied in the United States since the late 1960s and little development occurred until the 1990s (see Figure 1). As a result, until recently, the number of sites has been limited, and the fact that it may take 10 years to develop an operational ASR system, means that truly acquiring data has only recently become available to a number of sites. Hence, the first complete survey of ASR sites was completed in 2013, and little has changed since that time [3, 4]. Dataset was the first comprehensive analysis of the 204 sites in the US. U.S. EPA and environmental agencies in each state with ASR wells were contacted by phone or email to whether the state had such programs in place or not, and where they might be located. The list of ASR sites identified by the regulatory agencies was a critical component of the project because while prior inventories were prepared by regulatory agencies and consultants, none were complete and most excluded projects that were no longer active [5-9]. In each of these documents, the goal was to provide information on successful ASR sites as case studies and were relatively limited to a few sites as opposed to a nationwide survey (for example, AWWA [5] included only 4 sites in Florida, as opposed to 54). Hence, while AWWA [5] and Bloetscher et al [6] provided more extensive summaries that the texts by other authors, these reports were also very limited in scope. No analysis of the data was conducted to identify trends, success and challenges for ASR projects. The first to analyze the successes and challenges encountered by ASR projects were Bloetscher et al [3, 4].
Figure 1. Cumulative ASR sites by Decade
Bloetscher et al [3, 4] outlined the basic findings of the survey, the lessons learned from the ASR survey and summarized the salient commonalities. Variables of interest were identified to account for operations, physical construction, and localized differences [2-4]. One of the issues that arose as a part of the survey was that nearly a third of these systems were not active and another third were in some phase of the testing mode. The inactive sites include a total of 220 wells that are not in use at this time. A statistical attempt was made in Bloetscher [3] to identify why ASR projects were active (or not), but no attempt was made to predict the likelihood of success of the wells in test mode. While much can be learned from successful projects, comparing the successful and inactive systems can provide insights into the criteria and process associated with the development of an ASR program and perhaps highlight factors that will suggest a high rate of success for those test and study projects.

2. Methodology

The data utilized for this analysis are noted in Table 1, which were variables extracted from the 2013 ASR site inventory [2], and then converted to numerical variables as required for the statistical methods employed (see Tables 2 and 3). Also, information was updated to reflect known changes in the ASR wells. Am0ong the issues noted was that complete information was not available for all sites and decisions needed to be made to determine is those ASR sites would be retained or the variables deleted. For example, the salinity of the injection zone is relevant when injecting fresh water into a brackish zone. Freshwater will float based on the principles of differential density, creating a challenge for recovery of the injected water. However, the dataset denoted that the majority of sites were injecting into freshwater (total dissolved solids under 1000 mg/L) except in south Florida [2]. As a result this variable was deleted as opposed to deleting several dozen sites that did not report the salinity. The decisions were important because those sites with incomplete data, or those variables that were incomplete, cannot be used in principal component analysis which would reduce the available data considerably. Likewise, the casing material was commonly not reported and the confined layer material was not well defined. These variables were also deleted to permit as many sites to remain as possible.
Table 1. Variables used in ASR Analysis
Table 2. Descriptive statistics of continuous variables related to the ASR sites in the United States
Table 3. Descriptive statistics of categorical variables per ASR program status in the United States

2.1. PCA and FA Analysis

The factor analysis method dates from Spearman [10] and continues to develop. Today, there are two main types of factor analysis: Exploratory factor analysis (or EFA) and Confirmatory factor analysis (or CFA). EFA is used by XLStat® to reveal the possible existence of underlying factors which give an overview of the information contained in a very large number of measured variables. For EFA, the structure linking the variables is initially unknown, but the number of factors is assumed. CFA uses a method identical to EFA but the structure linking underlying factors to measured variables is assumed to be known [11].
Principal Component Analysis (PCA) is popular multivariate technical mainly used to reduce the dimensionality of p multi-attributes to two or three dimensions [11-13]. PCA is a special case of factor analysis (where k, the number of factors, equals p, the number of variables). While FA assumes a number of factors, PCA is used to reduce the number of variables to factor sets, while maximizing the unchanged variability in order to obtain independent (non-correlated) factors [14]. The mathematics of PCA uses an orthogonal transformation convert observations of possibly correlated variables into a set of values of linearly uncorrelated variables called principal components [13-16]. PCA uses a multivariate statistical parameter called an eigenvalue, which is a measure of the amount of variation explained by each principal component. PCA summarizes the variation in a correlated multi-attribute to a set of uncorrelated components, each of which is a particular linear combination of the original variables [17]. PCA is the simplest of the true eigenvector-based multivariate analyses. A Scree Plot is a simple line segment plot that shows the fraction of total variance in the data as explained or represented by each component [18].
There are several uses for PCA, including [11]:
• The study and visualization of the correlations between variables to hopefully be able to limit the number of variables to be measured afterwards;
• Obtaining non-correlated factors which are linear combinations of the initial variables so as to use these factors in modeling methods such as linear regression, logistic regression or discriminant analysis.
• Visualizing observations in a 2- or 3-dimensional space in order to identify uniform or atypical groups of observations.
Two methods are commonly used for determining the number of factors to be used for interpreting the results: the Scree test [19] is based on the decreasing curve of eigenvalues. The number of factors to be kept corresponds to the first turning point found on the curve. However, these representations are only reliable if the sum of the variability percentages associated with the axes of the representation space are sufficiently high. If this percentage is high (for example 80%), the representation can be considered as reliable. If the percentage is reliable, it is recommended to produce representations on several axis pairs in order to validate the interpretation made on the first two factor axes.
The correlation biplot interprets the angles between the variables as these are directly linked to the correlations between the variables. The position of two observations projected onto a variable vector can be used to determine their relative level for this variable [11]. The Kaiser-Guttman rule suggests that only those factors with associated eigenvalues which are strictly greater than 1 should be kept [11]. The number of factors to be kept corresponds to the first turning point found on the curve. Crossed validation methods have been suggested to achieve this aim.

2.2. Linear Regression

Ultimately the goal is to determine if the condition has a consequence – i.e. the potential for failure. If so, one needs to know what that consequence is – in this case operation or inactive. The values were assigned for operational (1) or inactive (0) of aquifer storage units in the United States, as the dichotomous dependent variable. The impact of these factors can be developed via a linear regression model [12]. The model would be developed as follows [11, 20-21]:
SSI = w1C1+w2C2+w3C3+w4C4+…wiCi
where:
SSI = Site success index (consequence)
w = weighting factor
C is condition factor
If one knows the consequence, the weights can be found:
where the values of cn are real numbers and
are the factors which are a compilation of the original carriable to maximize variance. It assumes these constraints and linear variables in the matrices are non-negative. If there are negative values, they must be made positive as follows [11]:
the linear regression model provides a mechanism to model the data to determine if differences between the active and inactive projects exists. For the existing sites, if the site was active, the consequence value was assigned a value of 1. If not, 0. As a result the hypothesis was that those sites likely to be successful if the SSI would tend toward a value of 1, and those that likely would not pan out, would trend toward 0. Note that because certain factors may have no value at present (example depth of an undrilled well), it is possible that the regression equation provides an SSI result that is greater than 1 or less than 0.

2.3. Further Data Manipulation

Because the test and study site have incomplete data, the linear regression model was re-run to include only that data that would apply to the test and study sites. For example, if no well was drilled, the casing and well depths could not be known. The revised linear regression model was used to model the test and study sites to predict the likelihood of success.

3. Results and Discussion

The states with the most ASR programs are Florida (54), followed by California, New Jersey, Arizona and Oregon (see Figure 2). However, the presence of ASR sites is not necessarily an indicator for success of ASR projects. For example, in Florida, over half the sites are not active or have wells that are no longer used. With the elimination of inactive and test sites, there are only 22 active ASR sites (as compared to 54 ASR sites) in Florida.
Figure 2. Scree Plot showing that 11 factors are needed to get 70 percent of variance
Table 2 outlines the descriptive statistics for descriptive statistics for all sites for the full 2013 database. Table 3 includes the categorical variables from the 2013 dataset. From the 204 sites in Tables 1 and 2, 74 were removed as a part of the process because they were in study or test mode and therefore lacked certain data that is helpful in understanding the potential for success. Removal of these sites led to Table 4 which summarizes the remaining variables. Note because PCA and FA require no missing information, the number of complete datasets was reduced to 111. Also note that the wells were grouped into regions of the country to determine if there were commonalities across different regions. The regions were similar to those proposed in Bloetscher et al [3, 4], except that the Florida wells were removed.
Table 4. Summary Statistics for Retained Variables
Table 5 is a correlation analysis between variables. Significant correlations exist for:
• Sand/sandstone formations in the east
• Unconfined alluvial formations in the west/southwest
• Confined limestone formations in Florida, and
• Reclaimed water being stored for irrigation
Table 5. Correlation Analysis of retained Variables
Table 5. Continue
The Scree plot [19] showed that the factors created by the eigenvalues required 11 factors to obtain 70 percent of the variance, which is a lot of factors and suggests that there is much scatter in the variables – one reason the locations were developed as a means to attempt to compare commonalities among regions and create greater degrees of correlation. The factor loadings revealed the factor loading and therefore the factor correlation with the original variables:
• F1 – relates to the location – Southwest and west locations were correlated with unconfinement and alluvial formations,
• F2 – relates to formation and water source – limestone and the use of groundwater, correlated with Florida
• F3 – relates to the number of active wells on a site (more increasing likelihood of success), depth of the well and depth of the casing
• F4 – relates to raw water as a source for the ASR wells
• F5 – relates to the number of wells (more increasing likelihood of success)
• F6 – relates to northwest wells (and basalt formations)
• F7 – relates to the number of inactive wells
All other actors had very limited factorial combinations as demonstrated by the relationships in Table 6. Each of these factors also contributed significantly to the factor loading (see Table 7).
Table 6. Factor Correlations (All 11)
Table 7. Percent Contribution to the Factor
PCA permits the use of a varimax rotation to improve correlations to explain variability. However, the varimax rotation does not significantly help to reduce the number of variables for the project but does reinforce several things:
• D1 – relates to the location – Southwest and west locations were correlated with unconfinement and alluvial formations, and not limestone
• D2 – relates to formation and water source – limestone and the use of groundwater, correlated with Florida, and differentiated from sand in the remaining southeast
• D3 –depth of the well and depth of the casing are related, perhaps weakly to recovery
• D4 – raw water and potable use from the ASR well are related.
• D5 – relates to the number of wells and number of active wells (more increasing likelihood of success)
• D6 – reinforces the relationship between northwest wells and basalt formations
• D7 – relates to active status of wells
• D8 – notes that ground and surface water system are inversely related
• D9 – relates to injection horizon and metals recovery and
• D10 – relates to withdrawal and injection capacity beings strongly correlated.
Table 8 shows that the varimax variables and their contribution to variance (see Figure 3).
Table 8. Factors after Varimax Rotation
Figure 3. Varimax Plots of Factors from PCA Analysis
The next step was to run a linear regression model in an attempt to understand if an equation could be developed to predict success. A linear regression model was run in XLSTAT®. Several variables were deleted from the original data set because they do not appear in the data for the test or study projects. The remaining variables are shown in Table 9. Table 10 shows the weight components applied to each variable (see also Figure 4). Figure 5 shows the results – predicted results for active versus non-active sites based on the predicted means (0.687 vs 0.35), and standard deviations (0.280 vs 0.180) for the active and inactive sites, respectively (note there were 58 active sites and 53 inactive sites). The model predicts these relatively well. Note that potable water use and the number of inactive wells did not contribute to the variance.
Table 9. Linear Regression Model Parameters (all compete data for Active and inactive sites only)
Table 10. Variables for the Active and Inactive Sites that also exist for the Test and Study Sites used in the revised Linear Regression model
Figure 4. Perspective on Linear Regression Model Parameters
Figure 5. Correlation between predicted and actual successful wells
These same factors were applied to the data on the test/study wells. Table 11 outlines the variables for these test/study wells. They are similar to those of the active and inactive wells. Using the components in Table 11, Table 12 outlines the results applied to the sites (listed by state only). The factors have a range from just under zero to just over 1, as does the analysis of active and inactive wells. It appears that this model may provide useful information for likely success. Figure 6 shows that 11 of the sites have values under 0.5 (including 5 under 0.35), which means their likelihood of success is low. Twenty-three of the sties have a value greater than 0.687, which suggests that these sites are likely to have success.
Table 11. Linear Regression Weights for Use in the Predictive Model
Table 12. Summary of prediction for Success for study or Test wells
Figure 6. Prediction on success of Test wells

4. Conclusions

The goal of this paper was to use this data and apply the results from active or inactive wells to those currently in the test of study phase in an effort to determine if there was a means to predict their likely success. This paper builds on the 2013 a nationwide survey of ASR systems as discussed in Bloetscher, et al [3,4] and AWWA [2]. The data from the 2013 was analyzed using factor and principal component analysis to determine correlations and variance combinations on the data. The goal was to determine which factors correlated best as a means to determine if a useful analysis could be developed to predict success of ASR systems currently in the test phase based on the success of active ASR sites.
The results indicate that the use of PCA and linear regression can be used to project the potential for the test and study sites. Two thirds of the current 204 ASR sites are either active or inactive, and, once the data was sorted, 111 of those sites were able to be used to project future status. While the actual results may not be known for many years, the results shed light on the over 50 sites in this stage and their likelihood for success. The data suggests that about 1/3 of the wells have low likelihood for success and perhaps should not be pursued further.
Several caveats exist for this analysis. First, the regional locations ignore that geological differences can be very different between nearby sites. Some effort was made to address this issue – for example Florida (mostly limestone) was separated from the rest of the southeast that was not. Information on salinity in the injection formation would be useful as a number of people, including the author, believe this is a major barrier to success. However, the results also suggest that more complete information would be useful for further analysis. Many sites lack full information, especially those in the study phase or are prior to 1990.

References

[1]  American Water Works Association (2014). M21 Groundwater, AWWA, Denver, CO.
[2]  American Water Works Association (2015). M63 Aquifer Storage and Recovery, AWWA, Denver, CO.
[3]  Bloetscher, F.; Sham, Chi Ho, Ratick, S. and Danko III, J.J. 2015. Status of Aquifer Storage and Recovery In The United States – 2013, British Journal of Science 70 April 2015, Vol. 12 (2).
[4]  Bloetscher, F.; Sham, C.H.; Danko, J.J.; and Ratick, S. (2014). Lessons Learned from Aquifer Storage and Recovery (ASR) Systems in the United States, Journal of Water Research, 2014, 6, 1603-1629. http://www.scirp.org/journal/jwarp http://dx.doi.org/10.4236/jwarp.2014.617146.
[5]  American Water Works Association (2002). Survey and Analysis of Aquifer Storage and Recovery (ASR) Systems and Associated Regulatory Programs in the United States, AWWA, Denver, CO.
[6]  Bloetscher, F.; Muniz, A; and Gonzalez, A.M. (2002). Summary of AWWA Survey of ASR Practices in the United States, Florida Section AWWA Annual Conference Proceedings. Hazen and Sawyer, Hollywood, FL.
[7]  Missimer, T. M. and Maliva, R. (2010). Aquifer Storage and Recovery and Managed Aquifer Recharge Using Wells: Planning, Hydrogeology, Design, and Operation, Schlumberger, Houston, TX.
[8]  Pyne, R. D. G (1995). Groundwater Recharge and Wells: A Guide to Aquifer Storage Recovery, 1st Ed. , CRC Press, Boca Raton, FL.
[9]  Pyne, R. D. G (2005). Groundwater Recharge and Wells: A Guide to Aquifer Storage Recovery, 2nd Ed, CRC Press, Gainesville, FL.
[10]  Spearman, C. (1904). The Proof and Measurement of Association between Two Rings. The American Journal of Psychology, 15, 72-101. http://dx.doi.org/10.2307/1412159.
[11]  Bloetscher, F. (2018). Risk and Economic Development in the Provision of Public Infrastructure, Journal of Environmental Protection, Vol.09 No.09(2018), Article ID:86955,18 pages. 10.4236/jep.2018.99061.
[12]  Johnson, R.A. and Wichern, D.W., (1992). Applied Multivariate Statistical Analysis, Prentice-Hall, Englewood Cliffs, NJ.
[13]  Pearson, K. (1901). On Lines and Planes of Closest Fit to Systems of Points in Space. Philosophical Magazine. 2 (11): 559–572. doi:10.1080/14786440109462720.
[14]  Hotelling, H. (1933). Analysis of a complex of statistical variables into principal components. Journal of Educational Psychology, 24, 417–441, and 498–520.
[15]  Hotelling, H (1936). Relations between two sets of variates. Biometrika. 28 (3/4): 321–377. doi:10.2307/2333955.
[16]  Jolliffe I.T. Principal Component Analysis, Series: Springer Series in Statistics, 2nd ed., Springer, NY, 2002, XXIX, 487 p. 28 illus. ISBN 978-0-387-95442-4.
[17]  Pleitez Herrera, Fernando J. (2012). Predicting Removal Efficiency of Reverse Osmosis with Respect to endocrine disruptors and pharmaceuticals and personal care products. A Thesis Submitted to the Faculty of College of Engineering and Computer Science in Partial Fulfillment of the Requirements for the Degree of Master of Science, Florida Atlantic University, Boca Raton, FL.
[18]  IOS (2012). Creating a Cree Plot. Improved Outcomes Software. http://www.improvedoutcomes.com/docs/WebSiteDocs/PCA/Creating_a_Scree_Plot.htm.
[19]  Cattell R. B. (1966). The Scree test for the number of factors. Multivariate Behavioral Research, 1, 245–276. http://dx.doi.org/10.1207/s15327906mbr0102_10.
[20]  David A. Freedman (2009). Statistical Models: Theory and Practice. Cambridge University Press. p. 26.
[21]  Rencher, Alvin C.; Christensen, William F. (2012), "Chapter 10, Multivariate regression – Section 10.1, Introduction", Methods of Multivariate Analysis, Wiley Series in Probability and Statistics, 709 (3rd ed.), John Wiley & Sons, p. 19, ISBN 9781118391679.