Traditional transformer fault diagnosis method has poor classification effect in the condition of small example and low recognition efficiency of mass monitoring data. Aiming at the problems above, this paper proposes a parallel variable predictive model based on the Spark computing framework. Firstly, HDFS is used as memory storage system, RowMatrix is stored as a distributed matrix storage structure, and broadcast variable and adjustment of the partition number are used to optimize the degree of parallelism. Secondly, the optimal model type and model parameters are obtained by training four mathematical models to diagnose the transformer fault. The experimental results show that the proposed algorithm has higher accuracy than support vector machine, better computing efficiency than stand-alone environment, and good adaptability to high dimensional feature vector. This work is supported by National Natural Science Foundation of China (No. 5167702).