Abstract: In view of the intermittent, random and fluctuating characteristics of wind power output and the time-consuming drawback of its previous combined forecasting models, a combined wind power prediction model is developed based on a combination of the parallel technology, ensemble empirical mode decomposition (EEMD), and bidirectional long short-term memory (BiLSTM) neural networks. First, the initial sequence of wind power output is decomposed into a series of intrinsic mode functions using EEMD. Then, a parallel BiLSTM neural networks sub-model array for the intrinsic mode functions is constructed with the help of the multi-process message passing interface, and a Bayesian optimization algorithm is used to calibrate the hyperparameters of each sub-model. Finally, wind power predictions are obtained by synthesizing the parallel sub-model prediction sequences. Example verification shows that this prediction model has certain advantages over the five models compared in terms of one-step prediction, multi-step prediction, and execution efficiency. The results are useful for the formulation of power grid generation plans and the economic operation of power systems.

Key words: ensemble empirical mode decomposition, bidirectional long short-term memory, parallel computing, Bayesian optimization, combination model