针对天然气管道泄漏风险防控需求,研究构建基于机器学习算法的管道泄漏预测与风险评价模型。采集某区域天然气主干管网2023—2024年多源异构数据,涵盖运行监测、管道属性、环境风险等17项特征因子,经小波去噪、缺失值填补及地理信息系统(Geographic Information System,GIS)空间插值等预处理,构建含3 760条记录的时空数据集。对比极端梯度提升(Extreme Gradient Boosting,XGBoost)、轻量级梯度提升机(Light Gradient Boosting Machine,LightGBM)等5种算法,基于Optuna优化超参数,采用5折交叉验证评估模型性能。结果表明,XGBoost模型性能最优,准确率92.15%、受试者工作特征曲线下面积值0.974,压力梯度、瞬时流量差值、历史泄漏频率等为关键驱动特征。结合特征重要性划分低、较低、较高、高4个风险等级,提出分层防控策略与“数字孪生”系统开发等通用措施。研究结果为管道泄漏预警与风险管控提供了数据驱动的科学方案,可有效提升运维效率与安全水平。

To address the demand for risk prevention and control of natural gas pipeline leakage, this study develops a pipeline leakage prediction and risk assessment model based on machine learning algorithms. Multi-source heterogeneous data from the main natural gas pipeline network in a specific region during 2023—2024 were collected, covering 17 feature factors including operational monitoring, pipeline attributes, and environmental risks. After preprocessing including wavelet denoising, missing value imputation, and Geographic Information System(GIS) spatial interpolation, a spatiotemporal dataset containing 3 760 records was constructed. Five algorithms including Extreme Gradient Boosting(XGBoost) and Light Gradient Boosting Machine(LightGBM) were compared, hyperparameters were optimized using Optuna, and model performance was evaluated through 5-fold cross-validation. Results show that the XGBoost model achieves the best performance, with an accuracy of 92.15% and an area under the cure of 0.974. Pressure gradient, flow difference, and historical leakage frequency are identified as key driving features. Based on feature importance, four risk levels—low, relatively low, relatively high, and high—are classified, and general measures such as hierarchical prevention and control strategies and the development of digital twin systems are proposed. The study result provides a data-driven scientific solution for pipeline leakage early warning and risk management, effectively enhancing operational efficiency and safety levels.