A multi-location short-term wind speed prediction model based on spatiotemporal joint learning

Abstract

Using temporal and spatial correlations to predict wind speed is still one of the most challenging and least-researched fields in wind speed prediction. How to make full use of the spatial correlation of wind speed between adjacent wind turbines to improve the accuracy of wind speed prediction is significant. Therefore, to address the challenges of making multi-location, short-term wind speed predictions, we propose a multi-location wind speed prediction model based on spatiotemporal joint learning, which called Mixed-SqueezeNet-BiGRU. By designing the SqueezeNet model with a deformable convolutional network, the spatial correlation of a multi-location time series is learned by using the spatial wind speed matrix; by introducing the Channel Shuffle operation of ShuffleNet, the characteristics of multi-location and multi-channel are output to ensure the orderly flow of information among the channels. To describe the dynamic information of the time series, a time series dynamic network model based on the bi-directional gating cycle unit, BiGRU, is introduced to model the long-term and nonlinear dependence of the time series. Experiments were carried out on three wind farm's wind speed dataset to verify that the Mixed-SqueezeNet-BiGRU short-term wind speed prediction model has a better prediction performance than some advanced methods.