Prediction of Day-Ahead Low-Power Output Events in Regional Wind Farms Using Feature Spectrums With Dilated Causal Convolution and Squeeze-Excitation ShuffleNet Lightweight Deep Learning

The accurate prediction of low-power output events in regional wind farms is important for the grid dispatch of the power system. However, conventional wind power prediction methods mainly focus on improving the overall accuracy, so wind low-power output events are rarely discussed separately. An innovative prediction method for day-ahead low-power output events in regional wind farms using feature spectrums with dilated causal convolution (DCC) and squeeze-excitation (SE)-improved ShuffleNet is proposed in this paper. First, the time-series regional features are transformed into spectrum images, in which three possible feature permutations are introduced and discussed after feature creation and selection. Second, the DCC-SE-ShuffleNet lightweight deep learning neural network is proposed as the prediction model for low-power output events; the input consists of the previously obtained spectrum images and the output is the categorized sequence to represent whether the low-power output events happened. Finally, case studies are presented at multiple thresholds of low-power output events to verify the proposed method. Based on the proposed low-power output events prediction method, the recall rate, accuracy rate, critical success index, and F1-score can be improved by 17.52%, 13.97%, 20.78%, and 15.99%, respectively, compared with conventional regression-based approaches, thus demonstrating the effectiveness and superiority of the proposed method.