Multimodal selective state space model-based time series classification for electricity theft detection

Abstract

In smart grids, electricity theft detection is essential for ensuring power system security and minimizing revenue losses, causing global annual losses of approximately USD 89.3 billion in the utility sector. While deep learning-based approaches have demonstrated effectiveness in utilizing users’ electricity consumption records, formulated as time series, challenges persist in balancing performance and time complexity when processing long time series. Meanwhile, the extraction of long-range temporal dependencies, especially with non-uniform sequence lengths, requires improvement. We propose a novel Multimodal Mamba Model-based Time Series Classification approach (MMM4TSC), which integrates the selective state space of Mamba with the Relative Position Matrix (RPM). This innovation transforms non-stationary time series data into two-dimensional images, thereby enhancing spatial feature and long-range temporal feature extraction. The proposed Multimodal Mamba Layer effectively extracts features from both the original time series and its 2D image representation through sub-channel splitting, while enhancing long-range temporal feature learning by incorporating channel dependencies and multi-scale context. Comprehensive evaluations are conducted on 128 public UCR datasets and a real-world electricity consumption dataset. Experiments demonstrate that MMM4TSC exhibits strong adaptability in handling time series classification tasks of varying lengths, achieving an accuracy of 96.9%, along with an AUC of 0.994 and an F1-score of 0.963 in electricity theft detection, outperforming state-of-the-art time series classification methods and existing electricity detection approaches. Furthermore, MMM4TSC strikes an excellent balance between classification accuracy and computational efficiency, with an 84.2% reduction in model parameters.