Data-Driven Dynamic Hybrid Bayesian Network and Random Forest Models for Risk Assessment of the Operational Condition of Water Supply Pipeline Networks

Risk assessment of urban underground pipelines has received much attention in recent years. Unlike traditional expert-based research methods, this study utilises thousands of historical operation and maintenance (O&M) records and pipeline burst accident statistics, and through the application of data mining techniques, identifies seven key disaster-causing factors that directly affect the mechanical performance of pipelines, including temperature, pipeline material, depth of burial, diameter, age, length, and type of joints, which cover almost all the commonly used pipeline types available in today's market. Aiming at the limitation of Bayesian networks in dealing with nonlinear effects, this paper proposes a hybrid Bayesian network and random forest model (HBN-RF), which analyses the features of the historical data by clustering and forms a dynamic hybrid Bayesian network and Random Forest model based on the dominance of the temperature effect in three days' steps. This model has a greater improvement compared with the Naive Bayes Network model, particularly in recall, which is2.84 times higher.This study provides effective decision support for early warning and forecasting of water supply networks.

Keywords: Water supply pipeline networks, Pipeline networkburst risks, Data-driven model, Dynamic risk early warning and forecast, Key event-causing factors, HybridBayesian network and random forest model.