Abstract: In allusion to the defect in the small sample exact prediction brought by unreasonable allocation distribution transformer capacity as well as frequent heavy overload, a new early warning method for heavy overloaded distribution transformers was proposed. Firstly, an extended sample pool was formed to meet the learning requirement of large data samples. Secondly, by means of collecting load data of distribution transformers, social development and statistical data and meteorological data, the input characteristic variables that might impact heavy overload were rough selected, then aggregating them to form well-chosen feature data samples. And then, a deep belief network learning model for heavy overload early warning was constructed to analyze the development trend of heavy overloaded distribution transformers, and by means of early warning of short- and medium-term prediction and selecting annual load curve the K-means cluster analysis was performed to form heavy overload early warning list to implement the pre-judgment of the potential dangers of heavy overloaded distribution transformers. The proposed method could cope with the insufficient training sample data caused by the short operation time of the sampling system, the risk prevention of heavy overloaded distribution transformer as well as the optimization and adjustment of distribution transformers’ capacity could be realized. The early warning performance and the effectiveness of the proposed method are verified by calculation example.