Conducting limit-exceeding and abnormal data mining for power quality (PQ) monitored data, and giving the corresponding early warning can help find PQ problems more early. However, the research on setting the threshold of abnormal data is less, and the existing threshold judgment algorithm mainly concentrated in single assignment. Combinatorial weighting method for early warning threshold of PQ is first put forward in this paper. Clustering analysis method based on distance and improved Otsu are used to be objective weighting method. At the same time, expert scoring system is applied to be subjective weighting method. Finally, the weight.is determined by combinatorial weighting method. The example results show that the above method can set PQ early warning threshold accurately, and it can be used to in the early warning system for PQ abnormal data effectively.