Short-term wind power prediction is challenging due to the chaotic characteristics of wind speed. Since, for wind power industries, designing an accurate and reliable wind power forecasting model is essential, we deployed a novel composite deep learning-based evolutionary approach for accurate forecasting of the power output in wind-turbine farms, which is developed in three stages. At the beginning stage (pre-processing), the k-means clustering method and an autoencoder are employed to detect and filter noise in the SCADA measurements. In the Next step (decomposition), in order to decompose the SCADA time-series data, we proposed a new hybrid variational mode decomposition (HVMD) method, that consists of VMD and two heuristics: greedy Nelder-Mead search algorithm (GNM) and adaptive randomised local search (ARLS). Both heuristics are applied to tune the hyper-parameters of VMD that results in improving the performance of the forecasting model. In the third phase, based on prior knowledge that the underlying wind patterns are highly non-linear and diverse, we proposed a novel alternating optimisation algorithm that consists of self-adaptive differential evolution (SaDE) algorithm and sine cosine optimisation method as a hyper-parameter optimizer and then combine with a recurrent neural network (RNN) called Long Short-term memory (LSTM). This framework allows us to model the power curve of a wind turbine on a farm. A historical dataset from supervisory control and data acquisition (SCADA) systems were applied as input to estimate the power output from an onshore wind farm in Sweden. Two short time forecasting horizons, including 10 min ahead and 1 h ahead, are considered in our experiments. The achieved prediction results supported the superiority of the proposed hybrid model in terms of accurate forecasting and computational runtime compared with earlier published hybrid models applied in this paper.