A novel data driven approach for combating energy theft in urbanized smart grids using artificial intelligence.

Electricity Theft (ET) causes monetary losses for power utilities in the energy sector. It occurs when electricity is consumed without being billed. Several methods are available for automatically detecting ET. Most of these methods evaluate Electricity Consumption (EC) records. However, these methods either have a low Detection Rate (DR) or a high deployment cost with a high False Positive Rate (FPR). Moreover, it is difficult to identify fraudulent consumers based solely on EC records. In addition, owing to data imbalances, such methods prove to be inefficient for classification. To solve the aforementioned problems, we have proposed a combination of various techniques. The first one is the Fastfood Transform, which is used for dimensionality reduction, along with the Time Series Lag Embedded Network (TLENET) neural network, used for classification between honest and dishonest consumers. The second one is the Wavelet Transform used for dimensionality reduction with TLENET, and the third one is the Nyström method used for dimensionality reduction with TLENET. To tackle the risk of high variance that results in overfitting in Deep Learning (DL) models, a Localized Random Affine Shadowsampling (LoRAS) data balancing technique is used. We have employed various data balancing techniques to analyze the performance of our system model. A game theory based approach, SHapley Additive exPlanations (SHAP), is implemented for explaining the output of our deep model. We have used a real-world dataset, referred as the State Grid Corporation of China (SGCC), to perform the simulations. Our model has achieved 94% accuracy, 92% F1-score, 93% Area Under Curve-Receiver Operating Characteristics (AUC-ROC), and 87% Matthews Correlation Coefficient (MCC) with LoRAS, Wavelet Transform, and TLENET. With dimensionality reduction using Fastfood Transform, our model has achieved 93% accuracy, 92% F1-score, 92% AUC-ROC, and 85% MCC. When the Nyström method is employed for dimensionality reduction, our model has achieved 94% accuracy, 92% F1-score, 90% AUC-ROC, and MCC 85%. Extensive experiments indicate that the proposed model outperforms the existing conventional detectors. • Using Fast Food Transform with Time Series Lag Embedded Network (TLENET). • Using Wavelet Transform and Nystroem method with TLENET. • Using LoRAS for tackling the problem of high variance. • Data balancing techniques are applied to solve the class imbalance issue. • A real-world dataset, SGCC, is used to perform simulations. [ABSTRACT FROM AUTHOR]