Wind speed forecasting system based on gated recurrent units and convolutional spiking neural networks

Abstract

Deep recurrent neural networks, such as gated recurrent units and long short-term memories, have been widely applied in wind speed forecasting. However, the simulations of the dynamics of the neurons in these models are different from the dynamics of natural neurons, and the useful temporal information is not fully extracted. This results in an unsatisfactory forecasting accuracy for practical wind energy management. In this study, under the hypothesis that a wind speed series can be forecasted using only previous observations (without any other information from the outer environment), a hybrid dual temporal information wind speed forecasting system comprising a third-generation spiking neural network is proposed, aiming to better extract temporal information. A fluctuating feature decomposition strategy is adopted to separate the different modes and adaptively transform the original series into several subseries. Subsequently, the third-generation spiking neural network is integrated with a convolution operation to correct and optimize the forecasting performance of a single recurrent deep learning model. Finally, an effective optimization algorithm is applied to obtain a linear combination of the forecasting outputs of each subseries. Four wind datasets collected from the Liaotung Peninsula in China are used to verify the effectiveness of the designed forecasting system. The experiments indicate that the proposed forecasting system achieves MAPEhengshan=1.43%, MAPExianren=1.40%, MAPEdonggang=1.49%, and MAPEdandong=2.56%, thereby showing excellent forecasting performance.