To provide highly accurate wind power forecasting data for system dispatch centers to correct intraday generation schedule deviations, a minute-scale ultra-short-term encoder-decoder forecasting model integrating bidirectional gated recurrent units (BiGRU) and multi-head attention mechanism is proposed. First, improved dynamic time warping algorithm is used to quantify the similarity and cluster wind turbines. Next, improved complete ensemble empirical mode decomposition with adaptive noise is adopted to decompose historical power series. Pearson correlation coefficients are used to select highly correlated subsequences. Finally, an encoder-decoder network is built to predict each subsequence and aggregate to obtain total wind farm power. Validation shows the proposed method performs better in 1-4 hours ahead forecasting.