Misoperation of power grids is one of the main factors leading to equipment damage and system failure incidents. To enhance the grid′s ability to prevent misoperation, this paper proposes a multimodal data fusion-based power grid misoperation diagnosis method. First, video surveillance, sensor data, and operation logs are fused using a bidirectional gated recurrent unit (BiGRU). A multi-interaction attention mechanism is employed to obtain a unified representation of the multimodal data. Then, a Bayesian neural network is used to update the weights of the long short-term memory (LSTM) network. By combining the uncertainty estimation capability of Bayesian methods with the strength of LSTM in handling time-series data, the fused multimodal features are fed into a Bayesian LSTM network to diagnose grid misoperation. Finally, the simulation results show that the proposed method effectively improves the accuracy of misoperation diagnosis and further enhances the safety and stability of grid operations.