Abstract: The widespread use of video conferencing systems meets the real-time communication needs of multiple parties in power scenarios, but there is currently no research on optimizing the quality of user video conferencing experience for the high-quality video conferencing needs of distributed power industry users with multiple concurrent and massive access in the new power system. Therefore, this article uses convolutional networks based on deep learning algorithms for spatial feature analysis of video conferencing quality data. Then, the image feature information extracted by two-dimensional convolution is used for feature learning and data classification through long short-term memory networks. Finally, the system adjusts video stream parameters such as bit rate, frame rate, and resolution end-to-end based on the output of the deep learning model, in order to minimize buffering and latency and improve video quality in real-time. The experimental results show that the model described in this article significantly improves three key indicators of user experience quality, including re buffering rate, average bit rate, and structural similarity index.