[Objective] With the continuous growth in energy consumption， the installed capacity of clean energy， particularly wind energy， is steadily increasing. However， the construction of new wind power stations faces significant challenges in accurate wind power prediction owing to the limited data availability and dynamic nature of scene data， which make it difficult to capture time-varying characteristics. [Methods] To address these issue， this study proposes a wind power prediction method for new wind power stations with limited data availability. First， a similarity measurement method combining the Canberra distance and dynamic time warping algorithm is adopted to efficiently identify multiple source-domain wind power stations similar to the new target station. Second， a pre-training model based on a multi-source transfer learning dilated convolutional neural network and bidirectional long short-term memory （dilated convolutional neural network-bidirectional long short-term memory， DCNN-BiLSTM） is established using the wind power data of the selected source domain stations. This approach transfers the experience knowledge of multiple similar source domain stations to the new target wind power station， thereby avoiding over-reliance on single-source domain data. To explore the impact of time-varying scene data on prediction results， a prediction method combining an online adaptive module is proposed. Two self-evolution methods were established considering data-matching update adaptation and weight-update adaptation. The output results of the basic prediction and online models in the online adaptive module were weighted to achieve short-term wind power prediction for the new wind power stations. [Results] The proposed method was verified using data from a wind power station cluster in Northwest China， demonstrating its superior ability to screen source-domain wind power station data and achieve more accurate power predictions. [Conclusions] The method proposed in this article provides a feasible solution to the challenge of scarce power data in under limited information conditions and is expected to broaden the application scope of such methods.