Smart grid electricity theft prediction using cascaded R-CNN and hybrid metaheuristic optimization

The theft of electricity is regarded as a global problem which creates negative impacts for both electricity users and utility companies. The economic development of utility companies gets destabilized which further leads to electric hazards, thereby increasing the energy cost. Numerous methods are utilized for substantial detection of electricity theft, but these approaches consume more time and are inefficient and expensive. Electricity theft detection also uses artificial intelligence techniques like deep learning and machine learning. Despite innovative and remarkable characteristics of these approaches, their performance is unsatisfactory. Taking these aforementioned issues into consideration, a cascaded region-based convolutional neural network with a cascade of specialized regressors is proposed in this work for efficient detection of electricity theft. The proposed classifier determines the close false positives for adjacent stage training enabling the generation of high quality detection of electricity theft. Initially, pre-processing which combines data interpolation and data normalization is carried out for the process of recovering missing values. An adaptive synthetic technique is utilized to address class imbalance issue owing to unbalanced data. In order to extract relevant features, a hybrid whale optimized chicken swarm algorithm is used which selects the accurate features thus performing the effective modelling of obtained electrical parameters. In comparison with existing approaches, the proposed work generates optimized results for performance metrics values with an accuracy of 94.3%, F1-Score of 94.58%, and precision of 94%.