Prediction of wind power ramp events via a self-attention based deep learning approach

Abstract

With the increasing popularity of wind power generation, the importance of accurate wind power ramp events (WPRE) prediction has been further emphasized. At present, WPRE prediction model based on deep learning shows superior performance, which does not need to fully grasp the operation of wind power generation, but ignores the interdependence mechanism of prediction error along the input characteristics of the neural network. This paper develops a self-attention (SAM) and wavelet transform (WT) based hybrid 1DCNN (one-dimensional convolutional neural network) and LSTM (long short term memory networks) for WPRE forecasting. In the proposed model, WT effectively separates the middle and high frequency noise signals in wind power and SAM can redistribute the neural weights in 1DCNN-LSTM, then 1DCNN-LSTM further extracts the space-time information of effective wind power. Finally, the performance of the forecast model proposed in this paper is tested by using the wind power data from the Belgian ELIA website. The experimental results show that, compared with the other six competition models, the proposed model effectively suppresses the influence of randomness, volatility and uncertainty of wind power on the recognition of ramp events. Furthermore, through quantitative analysis, this paper explains the reason of SAM is superior to traditional AM in terms of their application object and weight control range in the proposed model.