An efficient approach for electricity theft detection based on transfer learning in temporal sequence

Investigating non-technical losses (NTLs) is a challenging issue around the globe and results in a huge revenue loss. The Literature proposes many solutions, however, there are still many unanswered questions. This paper introduces a novel concept of transfer learning (TL) in temporal sequence data to investigate NTLs. The concept of learning target domains through target learners is highlighted using transformed knowledge. This concept reduces the requirement of excessive data of target domains, which minimizes execution time and resources. A novel approach based on TL is introduced to identify electricity theft. Temporal sequence data of State Grid Corporation of China (SGCC) are utilized to identify anomalous and benign consumers using pre-trained approach-based hybrid models of CNNLSTM and BiGRU-BiLSTM. Data utilized for pre-trained and target domains have similar symmetry and are identical. Initially, the hybrid models are trained and evaluated in their original forms, and later on, their customized TL-based versions are evaluated using the same data statistics, to classify the categorical labels. It is observed that TL-based variants reduce trainable parameters, and require less data and expert knowledge. The models are customized and are introduced as novel versions of the already available hybrid models (BiGRU-BiLSTM and CNNLSTM). Moreover, to evaluate the performance of the model AUC curve, precision-recall curve, AUC-score, F1-score, recall, precision, and accuracy are used. For the TL-based CNN-LSTM model, scores of 0.76, 0.80, 0.78, and 0.84, 0.85 are reported for precision, recall, Fl-score, and AUC and accuracy, respectively. However, precision, recall, Fl-score, and AUC, the accuracy achieved for the TL-based BIGRU-BILSTM model are 0.80, 0.77, 0.78, 0.89, and 0.86, respectively. • It is observed that the computational complexity of the DL models is a serious issue and takes time for their execution. The problem needs serious dimensionality reduction and less computational parameters, which is tackled in this study by adopting a TLbased mechanism. • High FPR is an expensive parameter and results in an excessive revenue loss. The study highlights the aforementioned aspect and is tackled through novel introduced hybrid models. • Furthermore, the study proposes two novel hybrid models as novel classifiers to classify the temporal sequence data based on the TL mechanism. • Moreover, the study highlights data augmentation through data manipulating techniques to overcome the issue of the nonavailability of fraudulent class data.