With the digitization and intellectualization of power systems, the prediction of distribution transformer overload has become one of the key technologies for realizing intelligent condition-based maintenance. In real-world scenarios, the spatiotemporal factors of distribution transformer overload often exhibit a biased distribution, among which some high-risk and rare (HRR) factors, once occurred, can cause irreversible damage to transformers and should not be ignored. Therefore, this paper proposes a prediction method for distribution transformer overload based on the improved association rules-criticality importance (IAR-CI ) model. Firstly, considering both internal and external factors, multiple data sources are collected to establish a database of distribution transformer operating states, and ICA is used to identify rare high-risk periods and HRR factors that are strongly associated with severe transformer overload. Secondly, based on the criticality importance (CI) metric calculation, a factor weighting method is designed to accurately measure the risk weight of each factor. Finally, the TBFP-Growth algorithm is applied to enhance the operational efficiency of the model. Simulation analysis conducted in a region in southern China demonstrates that the proposed method can improve the prediction performance of severe distribution transformer overload, facilitating the reasonable planning and scientific scheduling of subsequent inspection and testing strategies. This reduces the operation and maintenance costs of power equipment while enhancing the reliability of power supply.