Wind speed forecasting plays an important role in estimating the power produced from wind farms. However, because of the non-linear and non-stationary characteristics of the wind speed time series, it is difficult to model and predict such series precisely by traditional wind speed forecasting models. In this paper, a novel hybrid modelling method is proposed, in which time series decomposition, feature selection, and basic forecasting model are combined in a synchronous optimisation framework. In this method, the above-mentioned modelling factors, which affect model performance, could make a concerted effort to improve the model. Specifically, variational mode decomposition, the Gram-Schmidt orthogonal, and extreme learning machine, are optimized synchronously by gravitational search algorithm in the proposed hybrid short-term wind speed forecasting model. First, variational mode decomposition is employed to decompose the original wind speed time series into a set of modes and into one bias series. Subsequently, the Gram-Schmidt orthogonal is used to select the important features. Next, the set of modes are forecasted using the ELM. Finally, the key parameters of the models in three stages are optimized synchronously by gravitational search algorithm. Seven data sets from the Sotavento Galicia wind farm and two wind farms in China have been adopted to evaluate the proposed method. The results show that the proposed method achieves significantly better performance than the traditional signal forecasting models both on one-step and multi-step wind speed forecasting with at least 40% average performance promotion over all the seven competitors.