부산대학교 도서관

In order to effectively analyze the nonlinear and non-stationary wind power and improve the prediction accuracy, an improved ultra-short-term wind power prediction model is proposed. Firstly, the time series of wind power historical data is used to forecast, and CEEMDAN algorithm is used to divide the original wind power time series into multiple modal components, which effectively reduces the complexity between the series. Secondly, aiming at the existence of many hyperparameters in BiLSTM, Brown-Levy movement is introduced by using the idea of marine predator algorithm (MPA) for reference, and the sea squirt swarm algorithm is improved by combining greedy rules and elite reverse learning strategy. The wind power prediction accuracy is the best by optimizing the relevant super parameters of BiLSTM. Finally, the evaluation indexes of the improved model are compared with those of BP, PSO-BiLSTM, SSA-BiLSTM and ISSA-BiLSTM models, and the RMSE and MAPE of the improved model are 2.89% and 0.089, respectively, which shows that the error of the improved model is smaller and the prediction effect is better than other models. The results show that the proposed ultra-short-term prediction model of wind power has better performance.