Data-Driven Predictive Maintenance for Gas Distribution Networks

Abstract

A generic data-driven approach is presented that employs machine learning to predict the future reliability of components in utility networks. The proposed approach enables utilities to implement a predictive maintenance strategy that optimizes life-cycle costs without compromising safety or creating environmental issues. Any machine learning technique that qualifies as a probabilistic classifier can be employed within the proposed approach. To identify the data-driven model that performs best, a practical metric to assess the performance of the competing models is proposed. This metric is specifically designed to quantify the forecasting performance with respect to maintenance planning. Additionally, a data-driven sensitivity analysis approach is discussed that allows for an assessment of the influence of the different features on the model prediction. Through an application example, it is demonstrated how the proposed approach can be applied to predict future defect rates of pipe sections for maintenance planning in a large gas distribution network.