Short-term wind power prediction based on two-layer decomposition and BiTCN-BiLSTM-attention model

Abstract

In order to solve the security threat brought by the volatility and randomness of large-scale distributed wind power, this paper proposed a wind power prediction model which integrates two-layer decomposition and deep learning, effectively realizing the accurate prediction of wind power series with non-stationary characteristics. Initially, pearson correlation coefficient (PCC) is employed to identify primary meteorological variables as input series. Second, the wind power series are smoothed by implementing complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), and then all subseries are decomposed and obtained by utilizing empirical wavelet transform (EWT) for the components with the highest complexity. Subsequently, hidden information related to wind speed, wind direction, and wind power series are extracted through the bidirectional temporal convolutional network (BiTCN), and the obtained information is fed into a bidirectional long short-term memory network (BiLSTM) optimized by attention mechanism for prediction. Finally, the predicted values of all components are summed to derive the final prediction results. In addition, the significant advantages of the prediction model in this paper are verified by five comparison experiments. The mean absolute error (MAE) and root mean square error (RMSE) of the model's one-step prediction in the January dataset are 2.1647 and 2.8456, respectively.