Abstract
The primary objective of this research is to classify reservoirs based on their physical characteristics—such as dam height, width, age, and repair history—to improve maintenance prioritization. A dataset of 13,976 reservoirs was analyzed using k-means clustering and Self-Organizing Maps (SOMs), resulting in the identification of four distinct clusters. A decision tree was then generated to extract the critical factors and values defining these groups. The resulting classification rules are validated by survivor functions, which demonstrate unique reliability patterns for each reservoir group, providing a scientific basis for life cycle management.
Introduction: Addressing the Aging Infrastructure Crisis
South Korea manages approximately 18,000 rural reservoirs, a significant portion of which were constructed in the early 20th century. As these structures age, determining the urgency of repairs becomes a complex logistical and financial burden. Traditional classification methods often rely on oversimplified metrics like dam height, failing to account for environmental stressors such as drought frequency or economic factors like funding sources. This study proposes a data-driven framework to categorize reservoirs more accurately, enabling the use of "survivor functions" to predict remaining service life and optimize maintenance strategies.
Methodology: A Multi-Stage Data Mining Approach
The researchers implemented a multi-stage data mining pipeline to transform raw infrastructure data into actionable management rules.
1. Clustering via SOM and K-Means
Initially, Self-Organizing Maps (SOMs) were used to project high-dimensional reservoir data into a two-dimensional space. This visualization helped determine the optimal number of clusters ($k=4$). Subsequently, the k-means algorithm was applied to group the reservoirs, using an inter/intra cluster dissimilarity ratio to confirm stability.
2. Classification via Decision Trees
To interpret the clusters and derive human-readable rules, a decision tree (C4.5 algorithm) was trained on the clustered data. This step identifies the most influential variables that distinguish one reservoir class from another, turning abstract clusters into "if-then" logic.
3. Survivor Function Analysis
Finally, Weibull functions were applied to each class to model reliability over time. This statistical approach confirms whether the identified classes exhibit different aging behaviors, which is crucial for long-term infrastructure planning.
Dataset and Key Variables
The study utilized a database from the Korea Agricultural and Rural Infrastructure Corporation (KARICO) from the year 2000. From an initial pool of 13,976 records, a subset of 3,358 reservoirs with complete data across 21 variables was analyzed. Key variables included construction year, irrigated area, drought frequency, drainage area, dam height, structural type, and financial resource type (COSFUND).
Key Insights and Results
The data mining approach yielded several critical findings:
- Significant Predictors: The decision tree revealed that Drought Frequency, Source of Funding (COSFUND), and Structural Type were the most significant factors for classification, suggesting that environmental and economic context is as vital as physical dimensions.
- Model Robustness: The decision tree achieved a validation accuracy of 99.81%, indicating that the derived rules are highly reliable for categorizing new data.
- Divergent Aging Patterns: Survivor functions proved that different classes age at different rates. For instance, reservoirs in "Leaf 1" of the decision tree showed higher initial reliability but deteriorated faster after 35 years compared to those in "Leaf 3."
Conclusion
By combining unsupervised clustering with supervised classification, this research provides a robust framework for managing aging water infrastructure. The identification of four distinct reservoir classes allows for tailored maintenance schedules, moving away from inefficient "one-size-fits-all" approaches and toward a more predictive, data-driven lifecycle management system.
Summary updated