Owing to the consumption of considerable resources in developing physical pipe prediction models and the fact that the statistical models cannot fit the failure records perfectly, the purpose of this paper is to use data mining method to analyze and predict the risks of water pipe failure via considering attributes and location of pipes in historical failure records. One of the Automatized Machine Learning (AutoML) methods, tree-based pipeline optimization technique (TPOT) was used as the key data mining technique in this research.
By considering pipeline attributes, environmental factors and historical pipeline broke/breaks records, a water pipeline failure prediction method is proposed in this research. Regression analysis, genetic algorithm, machine learning, data mining approaches are used to analyze and predict the probability of pipeline failure. TPOT was used as the key data mining technique. A case study was carried out in a specific area in China to investigate the relationships between pipeline broke/breaks and relevant parameters, such as pipeline age, materials, diameter, pipeline density and so on.
By integrating the prediction models for individual pipelines and small research regions, a prediction model is developed to describe the probability of water pipe failures and validated by real data. A high fitting degree is achieved, which means a good potential of using the proposed method in reality as a guideline for identifying areas with high risks and taking proactive measures and optimizing the resources allocation for water supply companies.
Different models are developed to have better prediction on regional or individual pipeline. A comparison between the predicted values with real records has shown that a preliminary model has a good potential in predicting the future failure risks.
