An Ultra-short Term Wind Power Prediction Method Based on Spatio-temporal Hybrid Attention Mechanism

In response to the challenges posed by poor forecasting accuracy of wind power and inadequate meteorological sensors in certain wind farms, a wind power prediction model algorithm independent of meteorological data is proposed. Firstly, the sliding window based method is utilized to split the data and construct the time sequence data. The one-dimensional convolution network and long short-term memory network (LSTM) are subsequently employed to extract spatial and time dimension features from time series data with long term dependence. Finally, a spatio-temporal hybrid attention mechanism is employed to fuse features and predict power. Through the case analysis of three-year data from a wind farm in eastern China, the practicability of the above prediction model is proved and the results are more accurate than those of CNN-LSTM model, CNN model and LSTM model. Both validity and practicability of the proposed prediction method are proved, providing substantial support for reliability analysis of power prediction in actual scenarios.