A Hybrid Machine Learning and Deep Learning Framework for Effective Electricity Theft Detection in Smart Grids

Electricity theft is a significant issue for modern smart grids, necessitating the use for advanced detection technologies to ensure reliable and efficient power delivery. In this research, a hybrid machine learning and deep learning technique, feature engineering, and data preparation are used to provide a framework for identifying energy theft (ETD). The sigma outlier detection consist of three detectors and adaptive sampling techniques (ADASYN and SMOTE) ensure a clean and balanced dataset. The extraction of temporal and behavioral data, such as sliding window analysis and Fourier transform provides significant indicators of irregular consumption habits. The utilization of cross-validation metrics like precision, recall, F1-score and AUC a stacked ensemble method that combines ensemble models with a temporal convolutional network was developed to increase detection accuracy. The model was evaluated and demonstrated good performance in a variety of scenarios. The result validates the effectiveness of the proposed ETD system in the theft detection and showcase how adaptable and scalable it is in dynamic grid systems.