Abstract: The accurate prediction of the root alarms in the electric power communication network can assist the operation and maintenance personnel to efficiently investigate and quickly locate the high-risk points of the communication network in advance, then avoid regional communication failures and derivative alarms from the root, and reduce network risks and operation and maintenance costs. Aiming at the redundancy of source data and low accuracy of root alarm prediction in the existing research, this paper proposes a prediction model based on APRIORI-Bayesian optimization XGBoost for root alarms of electric power communication network. The APRIORI algorithm is used to optimize the input of the prediction model and mine the association rules among the influencing factors of root alarms. With the aid of the probabilistic method of association rules, the key influence factors are determined to reduce the training data redundancy of the Bayesian optimized XGBoost model, increase the data value density, and then improve the model efficiency and warning prediction accuracy. Then the prediction model is constructed on the basis of the Bayesian optimized XGBoost algorithm with the key factors. Finally, the experimental results show that the proposed algorithm performs well in prediction accuracy, recall and F-value, and achieves the optimal prediction accuracy when the minimum support is 15%, which can provide technical support for efficient maintenance and troubleshooting of power communication network.