Ultra-short term wind power prediction applying a novel model named SATCN-LSTM

Abstract

Accurate and reliable wind power forecasting has become very important to power system scheduling and safely stable operating. In this paper, a novel self-attention temporal convolutional network (SATCN) is combined with long-short term memory (LSTM) to forecast wind power for guaranteeing the continuous electricity supply. In the proposed SATCN-LSTM model, the structure of SATCN with a self-attention mechanism is conducted to pay more attention to features that contribute more to the output. The strength of SATCN is performed through extracting temporal feature of meteorological data and correlation characteristics between variables. LSTM is used after SATCN to further build the connection between features and outputs for predicting future ultra-short time wind power. The effectiveness and advancement of the proposed method is tested by using meteorological data and wind power data from two different wind farms in the U.S. The experimental results reveal that the SATCN-LSTM model is more accurate comparing to other methods. Taking California's fourth quarter wind power forecast results as an example, the proposed method has carried out a reduction of 17.56%, 10.99%,11.34% and 3.68% on the root mean square error compared with LSTM, TCN, CNN-LSTM, TCN-LSTM.