Dynamic electricity theft behavior analysis based on active learning and incremental learning in new power systems

• Using active learning to extract the most valuable user samples in the new dataset for labeling, thus reducing the labeling cost. • Combine active learning and incremental learning to analyze dynamic abnormal load pattern. • Proposed an adaptive incremental anomaly detection algorithm to solve electricity theft detection with new types of anomalous behaviors. The analysis of energy theft behavior in new power systems is essential for energy sustainable development and maintaining the stable operation of power grids. Traditional data-driven detection models which rely on training on historical data have high cost of training sample labeling. With the progressive development of electricity theft technology, the existing models are limited by the learning of new types of electricity consumption behaviors. Catastrophic forgetting will occur in incremental learning of the model, and a large number of repeated training will increase the training cost of the model. Therefore, this paper combines active learning and incremental learning to analyze the dynamic electricity stealing behavior detection problem in the new power system, and proposes a power theft detection model based on active learning and incremental support vector data description. Firstly, sample filtering strategy for unlabeled data based on gray similarity and kernel function (SFS-GSKF) extracts the most valuable user samples in the new dataset for labeling, so as to reduce the redundancy of user data and reduce the labeling cost. Finally, an adaptive incremental anomaly detection algorithm incorporating active learning (AIAD-AL) is constructed. The model is incrementally updated using the labeled samples of active learning, so as to improve the detection accuracy of new types of anomalous behaviors without forgetting the previous user electricity samples. The simulation results on real electricity consumption datasets show that the algorithm proposed in this paper has excellent sample selection ability, incremental learning capability, and better classification performance compared with existing active learning strategies and incremental detection algorithms.