Medium-term wind speed ensemble forecasting using ICEEMDAN and dual attention mechanism

Wind power output has strong randomness, volatility, and intermittency. To maintain the safety and stability of the large power grid under the new power system, high-precision medium-term wind power forecasting is urgently needed. This paper fully leverages the temporal dynamics of the wind power dataset and proposes a medium-term wind power ensemble forecasting model that integrates transfer entropy, improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) decomposition, dual attention mechanism, and multiple recurrent neural networks. Firstly, we compare the transfer entropy of wind power and meteorological factors to determine the direction of information flow and select the set of characteristic variables. Next, utilizing the ICEEMDAN signal decomposition algorithm, the wind power sequence is segmented into various intrinsic mode functions, and attention-based LSTM, GRU, and BiLSTM models are established. After aggregation and reconstruction, three sets of predictions are obtained. Finally, the attention mechanism is combined to dynamically weight the three models to achieve ensemble predictions. Actual examples show that compared with several benchmark models, the proposed model notably enhances the predictive accuracy of medium-term wind power forecasting.