Bad Data Detection for State Estimation via a Dual- Stream Multimodal Convolutional Neural Network

Bad data detection (BDD) is fundamental to power system state estimation and online security analysis. Although traditional model-driven methods based on residual tests offer a certain degree of interpretability, they struggle to meet the real-time and robustness requirements of modern power grids in scenarios involving complex nonlinear disturbances, topology errors, or severe noise contamination. This paper proposes a multimodal fusion convolutional neural network for bad data detection in power systems. Specifically, a dual-stream feature learning architecture is designed to enable end-to-end modeling of heterogeneous data, and a class-sensitive gated fusion mechanism is introduced to enhance the identification of critical anomaly classes. Experimental results demonstrate that, compared with baseline models such as CNN, the proposed method achieves a favorable balance between accuracy and inference efficiency in multiclass anomaly recognition, while exhibiting superior robustness and improved detection performance on critical classes. These results indicate that the proposed method provides a lightweight and deployable solution for bad data detection and multimodal decision-making in power systems.