The time series prediction of Supervisory Control and Data Acquisition (SCADA) systems is a pivotal challenge in the application of artificial intelligence (AI) within the wind turbine operations and maintenance (O&M). Following the great success of transformer-based approaches in neural language processing (NLP) and computer vision (CV), these methods have gained traction for time-series prediction tasks. Transformer architectures, particularly those designed for large time series model, commonly segment the data into patches to serve as input tokens. Nevertheless, the manner in which time series are divided into patches can affect prediction outcomes. The optimal patch size is not universal, varying with the prediction horizon required by different applications, thus raising concerns about the robustness of models that employ a fixed patch size. For comprehensive large time series models intended for diverse problem-solving, this variability in patching design necessitates careful consideration. This study investigates the application of transformers in time series prediction of SCADA data, analysing how different patching designs impact prediction accuracy across varying forecast lengths. The research is grounded in experiments conducted on a SCADA dataset from a real-world Chinese wind farm, and the results shows that encoder-only transformers are not sensitive to the patch size but the optimal patch size is positively correlated to the prediction horizon. This paper providing insights into the design choices that can enhance the performance of transformer-based prediction models.