Hybrid KNN–LSTM Framework for Electricity Theft Detection in Smart Grids Using SGCC Smart-Meter Data
Stephanie Ness
Abstract
The incorporation and integration of advanced metering infrastructure (AMI) and real-time monitoring have improved energy distribution through smart grid, being introduced at fast speeds. Electricity theft is, however, a key issue as it causes the world to lose more than 96 billion dollars annually, with an estimated 40 percent recorded in developing countries occasioned by the illegal meter messing and open connections. The continuously changing fraud tactics have increasingly rendered traditional rule-based and manually driven detection methods ineffective hence the need to implement the intelligent learning, evolving systems. The proposed framework is a hybrid deep learning solution in which several machine learning algorithms can be combined with Long Short-Term Memory (LSTM) networks to identify unusual patterns of energy consumption. The models were trained on time-series manifestations of consumer actions by using data about electricity consumption in January, 2014 through October, 2016. The best accuracy was produced by KNN+LSTM architecture 81.32, whereas the RF+LSTM one had a rather low accuracy level of 50.34. Both XGB+LSTM and LR+LSTM gave similar outcomes with the accuracy of 73.16% and 73.78%, respectively. These results prove that optimized feature extraction along with temporal modelling is able to improve fraud detection. The suggested hybrid system is a flexible and auto mechanism of resolving issues on grid protection and minimizing energy losses in superior power distribution systems.