An accurate prediction of short-term and long-term wind speed is necessary in order to integrate wind energy into large-scale grid power. However, wind speed presents diverse and complex seasonal and stochastic characteristics that impose challenges on wind speed forecasting models. This study proposes a Quaternion Convolutional Neural Network combined with a Bi-directional Long Short-Term Memory recurrent network to forecast wind speed. Quaternion Convolutional Neural Network is used to elicit more effective features from the stochastic sub-signals of wind speed. A new decomposition method is also proposed, comprising variational mode decomposition to decompose the wind speed data into optimal signal components, and an improved arithmetic optimisation algorithm to optimise the parameters of the variational mode decomposition. Furthermore, a fast and effective hyper-parameters tuner is introduced in order to adjust the hyper-parameters and architecture of the proposed hybrid forecasting model. The proposed forecasting model is developed based on data collected from Lesvos and Samothraki Greek islands located in the North Aegean Sea with the forecasting range in one-day ahead (long-term) and achieved considerable accuracy improvements in these case studies compared with the bi-directional long short-term memory model at 13% and 20%, respectively. The experimental outcomes confirm that, first, the proposed hybrid forecasting model considerably outperforms the five existing machine learning and two hybrid models in terms of precision and stability.