Medium-Term Wind Power Forecasting (MTWPF), with a 7-day forecasting horizon, can provide important support for dispatch plans in the power system and trading strategies in the electricity market. Existing MTWPF methods usually use the Numerical Weather Prediction (NWP) to map the power generated by wind farms. However, the accuracy of the NWP decreases as the forecasting horizon increases, which reduces the performance of the MTWPF. In this paper, we propose a MTWPF model based on multi-head Dynamic Self-Attention mechanism (DSA-MTWPF), which dynamically adjusts the mapping between NWP and wind power according to the change in forecasting horizon. A case study using operational data from one wind farm in China demonstrates that the DSAM-MTWPF outperforms that of the state-of-the-art MTWPF models.