In the context of achieving carbon neutrality and carbon peaking, new energy source is a key approach. The development momentum of new energy in China is rapid, and more and more new energy is connected to the power system, but due to the unstable characteristics of wind power generation systems (such as fluctuating output, intermittent operation, and random changes), it has become one of the important factors affecting the stability of power supply and power quality. This paper discusses the challenges of short-term wind power forecasting in the context of increasing wind power penetration and the need for privacy protection. Vertical Federated Learning (VFL) integrates spatial-temporal correlation of wind data from multiple wind farms while ensuring data privacy. However, VFL model has some defects. In this paper, an improved VFL model combining temporal convolutional network (TCN) encoder data encryption and self-attention mechanism real-time screening of cooperative stations is proposed. A case study shows that the model is effective compared with the traditional method.