Fraud detection using event logs with LSTM and gradient boosting.

Abstract

Automatic non-technical power loss detection methods have advanced significantly as data volume has increased with smart meter installation. Recently, academic works have mainly focused on the impact of the high resolution of the energy consumption time series, leaving aside the integration of event logs within machine learning solutions. Due to the variety of alarms and depending on electrical installation health, millions of alarm events can be generated requiring an automatic analysis of them. In this work, we propose a method that considers the sequential nature of alarm log information using a recurrent neural network and evaluate two strategies for including this information within an existing state-of-the-art NTL classifier. The experiments are reported in actual smart meter data provided by the Uruguayan utility, showing that it is possible to double the precision for on-field applicable operating thresholds.