Abstract: Traditional wind mutation monitoring performs threshold judgment on the collected wind direction sequence, which cannot change the shutdown problem of the plant due to large wind direction changes. In response to this situation, an intelligent wind direction time series forecasting method is proposed. The actual operating data of the plant was used as the data set, and the random forest method was applied for feature engineering to solve the problem of too few features. The Sigmoid function was used to classify the sequence and the regression models were constructed based on LightGBM for prediction. Bayesian optimization was applied to tune model parameters and the model performance was optimized. LSTM algorithm was used to establish a residual prediction model for self-correction. The experimental results show that the combined self-correction model improves the prediction accuracy and is feasible.