In order to solve the problems of low transmission efficiency, poor real-time performance and poor processing effect of distribution network fault data, this paper proposes a technology for efficient transmission and intelligent processing of distribution network fault data integrated with GPT. By constructing a lightweight GPT model and embedding it into the edge computing node, the adaptive quantization coding technology is used to compress the features of the fault recording data. Combined with the priority dynamic sharding mechanism, high-frequency fault features and low-frequency steady-state data are transmitted in layers. In the data processing stage, the multi-head self-attention architecture based on GPT models the spatiotemporal correlation of fault time series data, and fine-tunes the parameters of the pre-trained model through transfer learning to adapt it to the distribution characteristics of the distribution network fault features. At the same time, the adversarial generative network is introduced to enhance the generalization ability of small sample fault modes. The dynamic sharding transmission protocol has a maximum bandwidth utilization rate of 99.5%, and the average delay is only 15.27 ms, which provides strong support for the safe and stable operation of smart grids.