There is uncertainty in power transformer condition evaluation, and so an evaluation based on cloud similarity and evidence fusion is proposed. First, considering the randomness and fuzziness of each indicator at the state-level boundary of the power transformer, a cloud model is used to establish the basic framework of condition evaluation. Second, considering the rigor and fuzziness of state-level classification, an improved cloud entropy optimization algorithm is used to determine the cloud entropy of the cloud model. Then, given the uncertainty of the index data itself, a forward cloud generator and cloud synthesis algorithm are used to generate the identification and scale clouds for each test item, and the similarity between both is computed using fuzzy closeness as the basic probability assignment of the evidence sources. Finally, conflicting evidence correction methods that consider the credibility and uncertainty of the evidence are used to correct the evidence sources and blend different evidence to determine the final state of the power transformer. It is verified by examples that, compared with the traditional method, the method can effectively deal with the uncertainty in the condition evaluation process, and the evaluation results are in accord with the actual situation of the power transformer. This has a certain reference value for condition evaluation.