Formatted Title
Cluster Analysis to Support Decision-Making for a Bioremediation Injection Program
Background/Objectives
A bioremediation injection program has been in place at a confidential site that is contaminated with chlorinated solvents, primarily trichloroethene (TCE), and the injection program has been ongoing for approximately 10 years. At irregular intervals, additional bioremediation amendment is injected into areas of the site to ensure that the bioremediation remedy continues to progress toward its goal of reducing TCE concentrations to cleanup levels. However, the injection locations are not the same during each injection event. Rather, the project team evaluates the recent monitoring well data in order to determine which areas of the site require additional bioremediation amendment. Currently, the project team performs this evaluation manually. That is, the team looks at TCE concentrations and several other groundwater parameters from each monitoring well to determine where additional amendment is most likely needed, and conversely where no additional injection of amendment is needed. This process is time consuming and could be more efficient and less prone to human error with the use of data mining tools. The objective of this project was to utilize a data mining method to assist in the evaluation of locations that may require bioremediation injections during the next injection event.
Approach/Activities
Real data from 45 monitoring points were used to perform the analysis. Each monitoring point had four measures: TCE concentration, oxidation-reduction potential, pH, and total organic carbon concentration data. The data were first normalized, and then the k-means clustering algorithm was used to assign each of the 45 monitoring points to a cluster. The k-means algorithm requires setting the desired number of clusters, seeding the initial observations, calculating the average cluster values for each measure, calculating the distance between each observation and the average cluster values, assigning each of the observations to the closest cluster, and repeating the average and distance calculations until the set of cluster assignments does not change from one iteration to the next. The desired number of clusters was set to three based on an initial number of potential outcomes/decisions: 1) high TCE concentrations where additional injections would be necessary, 2) high TCE concentrations where additional injections were not necessary, and 3) low TCE concentrations where additional injections would not be necessary. The initial observation seeds were set to reflect these potential outcomes.
Results/Lessons Learned
The analysis resulted in two clusters with clear recommendations. The first was characterized by high TCE concentrations, low pH, low TOC; this suggests that injections near these monitoring points is recommended and should include a buffering agent to account for the low pH. The second was characterized by low TCE concentrations, suggesting that no injections would be needed near these monitoring points. The third cluster was characterized by high ORP, low pH, and low TOC, but inspection of the data revealed TCE concentrations ranging from non-detect to well-above the cleanup level. Based on the TCE concentrations alone, areas around some of these monitoring points would be recommended for bioremediation injections and additional manual effort is required to make those decisions. Still, the clustering analysis resulted in a 64% reduction in the number of monitoring points that would need to be inspected by the project team. Refinement of the data preprocessing and the clustering algorithm may result in clearer recommendations in future or refined applications.