A Method for Ultra-Short-Term Wind Power Forecasting: Based on Parallel Transformer Encoder Architecture

To address the issue of low wind power prediction accuracy caused by the intermittent nature of wind power forecasting and the inability of the full connection layer in PatchTST and iTransformer to capture temporal correlations within tokens, an ultra-short-term wind power prediction model based on a parallel Transformer encoder is proposed. First, a time series convolution layer is introduced in PatchTST to extract the time series features of endogenous variables. Second, a cross-attention mechanism and a time series convolution layer are introduced in iTransformer to extract the time series features of exogenous variables. Finally, the extracted time series features of endogenous and exogenous variables are dimensionally aligned, and the final prediction results are output through Concatenate and BiLSTM. Through ablation studies and comparative experimental analyses, the proposed model demonstrates high accuracy and robustness in ultra-short-term wind power forecasting.