Optimizing Gas Utility Operations: A Machine Learning Approach to Targeted Inspections and Routing
Optimizing Gas Utility Operations: A Machine Learning Approach to Targeted Inspections and Routing
Mehrdad Ghasemi1
1) Master s Student in Public Administration – Public Policymaking, Faculty of Humanities, Department of Management, Za.c., Islamic Azad University, Zanjan, Iran
IOI Code
:
ACEE-ACAFZH
Publication :
5th International Conference on Researches in Accounting, Economics, Management & Humanities(meaconf.com)
Abstract :
Non-Technical Losses (NTL)—encompassing deliberate gas meter tampering and technical equipment failures—represent a persistent and substantial financial burden for natural gas utilities worldwide. Traditional fraud management relies predominantly on randomized field audits or rudimentary threshold-based alerts, which yield low detection rates, inflate operational expenditure, and inefficiently deploy field personnel. This paper proposes an end-to-end prescriptive analytics framework that couples advanced multi-class anomaly detection with profit-oriented inspector routing. In the first stage, a weather-aware Multi-Class Random Forest classifier is developed using historical consumption records and local meteorological variables. This design explicitly addresses a critical gap in the literature: the absence of models that can simultaneously distinguish between legitimate seasonal consumption drops, deliberate tampering, and technical degradation, thereby reducing weather-induced false alarms. The SHapley Additive exPlanations (SHAP) algorithm is integrated to provide field inspectors with feature-level diagnostic insights, transforming the classifier from a “black box” into a transparent decision-support tool. In the second stage, a Profit-Oriented Vehicle Routing Problem (PO-VRP) is formulated. Unlike conventional routing methods that minimize travel distance or treat all flagged anomalies as equally important, the proposed PO-VRP prioritizes inspections by calculating the expected financial recovery—a function of fraud probability and meter capacity—thereby aligning field operations with corporate revenue-recovery objectives and avoiding the inspection of low-value targets. Experimental evaluation on a simulated dataset of 35,000 consumption records demonstrates that the multi-class diagnostic model achieves an overall accuracy of 87%. While the high precision for the Fraud class (88%) ensures that dispatched inspections are rarely wasted, the moderate recall (36%) indicates that sophisticated theft patterns remain partially undetected. The integrated routing module yields a Return on Investment (ROI) exceeding 16,000% for a single daily route and prevents approximately 96.8 kg of CO₂ emissions by eliminating redundant travel. The framework provides a scalable, data-driven blueprint for utilities to transition from reactive monitoring toward proactive, value-based resource allocation.
Keywords : gas utility operations non technical loss (ntl) explainable ai (shap) fault diagnosis profit oriented vehicle routing (po vrp) prescriptive analytics