Abstract
Wind power forecasting (WPF) models play an increasingly important role in integrating wind power into electricity systems. Portability of such models allows for quantified calibration features which can be taken from one farm and applied to another without compromising forecast accuracy. This paper investigates the portability of WPF methods by exploring the influence of model hyperparameter configurations on forecasting performance. The performance of two hybrid WPF methods are evaluated and compared, Variational Mode Decomposition & Feed Forward Neural Network (VMD-FFNN) and Ensemble Empirical Mode Decomposition & Feed Forward Neural Network (EEMD-FFNN). Supervisory Control and Data Acquisition (SCADA) data from wind farms in Ireland and the UK are utilised. The robustness and portability of the WPF methods when applied to different datasets are examined. The models demonstrated good forecasting accuracy, with the VMD-FFNN model achieving 3.42% NMAE error for the Irish site. For portability, the forecasting performance is found to be sensitive to two of the four model hyperparameters examined. A low number of modes used in signal decomposition, beyond a threshold of ~4 modes, is adequate for accurate prediction, although calibration is still required depending on the wind farm. Additionally, the number and variety of datasets improved model robustness.