Classification and characterization methods of non-tchnical losses on smart grid scenarios

Abstract

Nowadays, grid resilience as a feature has become non-negotiable, significantly when power interruptions can impact the economy and society. Smart Grids (SGs) widespread popularity enables an immense amount of fine-grained e lectricity consumption data to be collected. However, risks can still exist in the Smart Grid (SG), since SG systems exchange valuable data, the distribution system loses substantial electrical energy. We divide this loss into two categories: technical and non-technical loss. A substantial amount of electrical energy is lost throughout the distribution system, and these losses are divided into two types: technical and non-technical. Non-technical losses (NTL) are any electrical energy consumed that is not invoiced. They may occur due to illegal connections, fraudulent activities, issues with energy meters such as delay in the installation or reading errors, contaminated, defective, or non-adapted measuring equipment, very low valid consumption estimates, faulty connections, and disregarded customers. Non-technical losses are the primary cause of revenue loss in the SG. Annually, electrical utilities incur billions in losses due to non-technical reasons. This thesis presents two detection methods of NTL: classification a nd c haracterization. We c reate a n ensemble predictor-based time series classifier t o c lassify N TL d etection. T his p redictor u ses the user’s energy consumption as a data input for classification, f rom s plitting t he d ata to executing the classifier. A lso, i t a ssumes t he t emporal a spects o f e nergy consumption data during the pre-processing, training, testing, and validation stages. The classification method has the advantage of classifying heterogeneous features in data. The characterization method proposes a study based on Information Theory Quantifiers (ITQ) to mitigate this challenge. First, we use a sliding window to convert the user’s energy consumption time series into a Bandt-Pompe (BP) probability distribution function. Then, we extract the used ITQ. Finally, we apply each metric to the Probability Density Function (PDF) and map the layers to characterize their behavior. The characterization method is advantageous to be used when we have big data. Overall, our best results have been recorded in the fraud detection-based time series classifiers (TSC) model, improving the empirical performance metrics by 10% or more over the other developed models. Our results show that users with normal and abnormal energy consumption can be distinguished using only Information Theory Quantifiers by considering the range of values for each metric.