As the proportion of new energy in power systems gradually increases, new energy power prediction becomes a research focus. However, the power prediction of the new-built wind farm is faced with the problem of historical data insufficiency, and difficulty in feature transfer. A short-term wind power prediction approach based on the deviation compensation TCN-LSTM and step transfer strategies are proposed. First of all, the small amount of data of the target wind farm are divided into two groups according to the correlation with the source wind farm. Then the historical data of the source wind farm is used to train the hybrid model containing the error compensation module. Finally, the model is constructed with the step transfer strategy. The relevant case analysis of this paper exhibits that the prediction accuracy of the compensation step transfer learning model based on TCN-LSTM is increased by 1.23% compared to similar direct prediction models. The effectiveness of the proposed approach is proved by related cases.