A Multi-Step Wind Speed Prediction Model for Multiple Sites Leveraging Spatio-temporal Correlation

The wind speed prediction with spatio-temporal correlation is a task of great significance and challenges. In this paper, we are concerned with the problem of multi-step wind speed prediction for multiple sites. Based on the nature of the spatio-temporal sequences, a two-stage modeling strategy for spatio-temporal sequence prediction was proposed, i.e., extracting the spatial features firstly and followed by capturing the temporal dependencies among these extracted spatial features. A model with deep architecture for wind speed prediction, termed the deep spatio-temporal network (DSTN), was presented. DSTN is composed of a convolutional neural network (CNN) and a bidirectional recurrent unit (BGRU), which is trained in an end-to-end (E2E) manner and capable of conducting predictions of sequence-to-sequence (S2S). Initially, the spatial features were extracted by the CNN at the bottom of DSTN. Then, the temporal dependencies among these spatial features at contiguous time sections were captured by the BGRU. Finally, the predicted wind speed was generated by the model based on the spatio-temporal correlation. Moreover, three error indices, evaluating the overall average performance and error control ability for the individual sample of the prediction model, were defined. Experiment results on real-world data from the state of California show that the proposed DSTN is capable of predicting wind speed with spatio-temporal correlation effectively, and it outperforms the prior arts. © 2019 Chin. Soc. for Elec. Eng.