Abstract: As a typical meteorological disaster, the cold wave weather poses a great challenge to the safe operation of wind power and the new power system with wind power as the main body. Providing accurate wind power prediction will be an effective response. For this reason, this paper presents a combined short-term wind power prediction method under cold weather with small samples. First, the cold wave weather events are defined and the characteristics of wind power output are analyzed. To solve the problem that sample data is scarce and difficult to model in cold weather, TimeGAN algorithm is used to enrich meteorological and power samples. Then, based on XGBoost and Transformer algorithms, the wind power reference and loss prediction models are established to quantify theoretical output and power shortage in cold weather. In addition, a binary Seq2Seq model based on attention mechanism is proposed to predict whether power loss occurs or not, so that the combined prediction is conducted by extracting the time period of wind power loss. Finally, the tests of the cold wave weather events represented by the weather with strong wind, strong rainfall, the combination of strong wind and strong rainfall are carried out. Compared with the traditional prediction model, this method shows good prediction performance under cold wave weather.