Log anomaly detection is one of the key technologies to monitor the operation of distribution master station system and identify abnormal behavior. Existing log anomaly detection methods based on deep learning rely on a large amount of in-domain training data, and the scarcity of training data will lead to a significant decline in performance. Aiming at the above problems, based on the contextual reasoning characteristics of large language models, an adaptive hint strategy is designed and a training-free anomaly detection scheme for distribution master logs is implemented. Firstly, a demonstration example filtering algorithm is designed to dynamically select several high-quality demonstration examples from a small number of labeled local logs for different online logs. Then, combined with the task description and human experience knowledge, a text hint is automatically constructed to guide the large language model to complete the anomaly detection task of distribution master station logs. The experimental results on the general data set and the self-built distribution master station data set show that the proposed scheme has better performance than the existing methods, showing higher flexibility and generalization.