Deep-learning-based Wind Speed Forecasting Considering Spatial-temporal Correlations with Adjacent Wind Turbines

The accurate prediction of wind speed, which greatly influences the secure and efficient application of wind energy, is still an important issue and a huge challenge. Previous research has largely focused on advanced algorithms, often ignoring the contribution of expanding predictors to predict wind speed. In order to promote the accuracy of forecasting, this study proposes a provisory wind speed forecasting model based on spatial-temporal correlation (SC) theory, in which the target and adjacent wind turbines, as well as the related time-lag characteristics, are examined through Wavelet Coherence Transformation analysis (WCT). Prior to that, the continuous wavelet transforms (CWT) are used to detect the spatial-temporal correlations with adjacent wind turbines. The CWT results show that the adjacent wind turbines which have a strong correlation with the target wind turbine are adopted as important factors of the forecasting model. Moreover, the study focuses on long short term memory (LSTM), a typical deep learning model from the family of deep neural networks, and compares its forecast accuracy to traditional methods with a proven track record of wind speed forecasting. Wind speed series of these model tests are taken from a Buckley City wind farm in Washington State, USA. The results of testing set reveal that (1) the root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) of the proposed model (SC-LSTM) are 0.49 m/s, 0.28 m/s and 2.57%, respectively, which are much lower than those of the conventional Back Propagation (BP) model, Extreme Learning Machines (ELM) model, and Support Vector Machine (SVM) model; (2) the proposed model that considers spatial-temporal correlations with adjacent wind turbines based on the WCT can obtain reliable and excellent prediction results, providing an excellent hybrid model for wind speed forecasts.