Short-term wind power prediction based on ICEEMDAN decomposition and BiTCN-BiGRU-multi-head self-attention model

In order to address the security threats posed by the volatility and stochasticity of large-scale distributed wind power, this paper proposes an attention-based hybrid deep learning approach for more efficient and accurate wind power sequence prediction. Firstly, the Pearson correlation coefficient (PCC) is used to identify the main meteorological variables as input sequences. Secondly, the intrinsic complete ensemble empirical mode decomposition with adaptive noise is used to decompose the sequence of wind power. Then, the hidden information such as wind speed, wind direction, and wind magnitude are extracted by bidirectional temporal convolutional networks (BiTCN), and the acquired information is inputted into bidirectional gated recurrent units (BiGRU) optimized by a multi-head self-attention mechanism for prediction. Finally, the predicted values of each component are summed to obtain the final prediction results. By comparing with the other 12 models, the results show that the two-scale integrated model of BiTCN and BiGRU can obtain better prediction accuracy. Compared with other benchmark models, the RMSE of this paper's model is reduced by more than 9.4%, indicating that this paper's model can fit the wind power data better and achieve better prediction results.