Aiming at the problem of poor model accuracy caused by poor stability and strong randomness of wind power output. A short-term prediction model of wind power based on quadratic decomposition error compensation was proposed. Firstly, BiLSTM (bidirectional long short-term memory) prediction model is established to predict wind power and output prediction errors. Secondly, an IDBO (improved dung beetle optimizer) algorithm was used to initialize the population by using chaotic mapping, update the position of rolling dung beetles by introducing golden sine strategy, and update the position of thieving dung beetles by adding dynamic adaptive weight coefficient to optimize the parameters of the prediction model. Prevent the network from falling into the local optimal solution, and adaptively search the optimal parameter combination. Then, using the decomposition-reconstruction-decomposition strategy, CEEMDAN (complete ensemble empirical mode decomposition with adaptive noise) was used for the first decomposition. In addition, SE(sample entropy) and K-means are introduced to reconstruct the sequence according to frequency, and the high-frequency error sequence was decomposed into error sequences of different frequency bands by VMD(variational mode decomposition). Improve the prediction efficiency and accuracy of subsequent models. Finally, the input error compensation model of each component was used to predict and the Attention mechanism was introduced to learn the feature relationship of different time steps and give different weight values to enhance the attention to key information. Through the measured data of a wind farm in Xinjiang, the prediction accuracy of the proposed model is proved to be high and has significant advantages.