Abstract: A new hybrid model incorporating error correction was proposed to effectively address the challenges posed by the randomness of wind power in forecasting. The raw wind power data and prediction errors were first decomposed through complete ensemble empirical mode decomposition with adaptive noice (CEEMDAN), t-test was used to measure the characteristics of each subsequence obtained after decomposition. A bidirectional long short-term memory network (BiLSTM) enhanced by SelfAttention mechanism was applied to predict sequences with high-frequency characteristics, and the autoregressive integrated moving average (ARIMA) model was used to process sequences with low-frequency characteristics. Finally, the initial wind power prediction results were added to the corresponding error prediction values to obtain the wind power prediction values based on error correction. Power samples from different wind fields were selected and compared with other 12 models to validate the performance of the proposed model CEEMDAN-ARIMA-BiLSTM-SelfAttention. Results show that the proposed model significantly improves the performance of wind power forecasting by at least 70.19%, 68.67%, 25.41% and 7.59% in terms of root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and coefficient of determination (R²) respectively under the WF1 autumn scenario compared to the single model. Therefore, the error correction and decomposition strategy effectively enhances forecasting accuracy. Compared to the baseline model, the proposed model demonstrates the most reliable forecasting method in multiple scenarios.