A new method to predict short-term load of wind power

Onshore wind power load data at Valencia, Spain (from January 1, 2015 to December 31, 2018) was analyzed in a hybrid model of bidirectional long-term/short-term memory and convolutional neural network (BiLSTM-CNN) with attention mechanism, to predict short-term power load．The attention mechanism was found to significantly improve predictive performance of BiLSTM after weighting input at varied time steps．LSTM-CNN was found more suitable for short-term load forecasting than CNN-LSTM, which could make full use of time series information but did not lose key information at the beginning．The root mean square error (RMSE) and mean absolute percentage error (MAPE) of the BiLSTM-attention-CNN model were 575.35 and 7.02%, respectively．Compared with other models, for MAPE, BiLSTM-attention-CNN was 9.65% lower than the second-best model of CNN-BiLSTM-attention; for RMSE, BiLSTM-attention-CNN was 2.75% lower than the second-best model of CNN-BiLSTM-attention．It is concluded that the present work can be readily applied in short-term forecasting of wind power load．