Deep Learning Ensemble Based New Approach for Very Short-Term Wind Power Forecasting

Abstract

This paper presents a new prediction approach based on deep learning ensemble for very short-term (10-minuteahead) wind power forecasting for a look-ahead period of 1h, 3h, and 6h. The proposed deep learning ensemble approach combines several individual and hybrid deep learning models, such as Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), Hybrid Deep Neural Network (HDNN), with the formation of four different ensembles, in particular HDNN+CNN,HDNN+LSTM, CNN+LSTM, and HDNN+CNN+LSTM. The proposed approach considers the historical data of wind speed as major input through ensemble averaging in order to produce the final wind power prediction. The major advantage of the proposed ensemble learning is that they make the best use of predictions from multiple deep learning models and their capability to effectively “cancel out” the individual errors, which in turn help enhance the final prediction accuracy. The simulation on actual data, acquired from the real wind farm in Texas, demonstrates the effectiveness of the presented approach to produce a higher degree of very short-term wind power forecast accuracy for multiple seasons of the year in comparison to other soft computing as well as to individual deep learning models.