Flexible HVDC transmission, as a new generation of DC transmission technology, is an important way to build a smart grid. During the operation of the core equipment in the valve hall of the converter station, due to the safety distance, maintenance personnel are unable to approach for detection. Its detection relies on indoor monitoring systems and inspection robots, etc. The risk of equipment failure generated cannot be dealt with in a timely manner. Therefore, risk prediction of the detection data generated in the flexible HVDC valve hall is particularly important. A risk prediction model for multi-modal data generated by flexible HVDC valve halls is proposed. This model is based on the generative adversarial network, uses non abnormal data for training, learns the potential distribution of data, reconstructs the data, and judges the degree of abnormality of data according to the error generated by reconstruction. Through the actual collected experimental data, it has been proven that the constructed model can effectively learn the potential distribution of the data, and can effectively identify abnormal data. Compared to traditional methods, this model is more accurate.