A CPFL Energia Fraud Detection Model Based on Geographic Census Sectors Analysis

Non-technical losses and irregular energy consumption severely jeopardize distribution utilities, and customers in general. Therefore, reducing them and pursuing revenue recovery stand for a crucial mechanism to secure companies financial health and quality in provided services. In such context, this work incorporates socioeconomic variables from demographic census into fraud detection models to improve the existing algorithms and enhance their performance. The inclusion of geographically sectioned explanatory characteristics reduces the existing models bias, preventing that only vulnerable areas are addressed in inspections, whereas others are unwisely left out. In order to achieve so, demographic census data are combined with historical inspection results and myriad predictive modeling tools to evaluate sectors fraud scores–i.e., the likeliness that an inspection will lead to the identification of a fraud within a given sector–and point out areas to which more inspections should be directed. The final proposed methodology applies Linear Regression, Spatial Auto-regressive Models, Geographically Weighted Regression and Logistic Regression techniques and the inclusion of sectors fraud scores has increased model accuracy to 72% (a 3 percentage points growth), improving success rates for fraud detection.