DVTransformer: A novel approach for multi-step wind power forecasting using spatio-temporal dynamics and predictive strategies

Short-term wind power forecasting is essential for wind farm management and reliable grid operations. However, the accuracy of turbine-specific forecasting is often compromised by limitations in sequence modeling and misleading information from the surrounding wind turbines. To address these challenges, we proposed a novel DVTransformer (DTW-VARIMA-Transformer) framework for turbine specific forecast integrating Transformer neural networks and a predictive strategy. This approach integrates spatio-temporal dynamics using wind speed from the surrounding turbines to forecast the wind power of the target turbine in a wind farm. Wind turbines were selected using dynamic time warping (DTW) based metrics, which calculate the dynamic distance between wind speed time series, ensuring the reliability of spatial information. Additionally, we incorporated predicted wind speeds of surrounding turbines using vector autoregressive integrated moving average (VARIMA), alongside historical data, to better capture the influence of future wind conditions on the target turbine power output. The DVTransformer performance was assessed against parallel models under various metrics, including mean absolute error (MAE), mean squared error (MSE), and correlation coefficient (R), demonstrating significant improvements in a multi-step ahead forecasting task. The proposed DVTransformer was first validated on two representative turbines to assess turbine-specific forecasting performance. For 3-step-ahead forecasting, DVTransformer achieved an average MSE of 0.085 for turbine T01, corresponding to reductions of 34.92%, 24.77%, 2.20%, and 27.34% compared to the Fast Fourier Transformer (FFTransformer), Informer, Spatio-temporal Long Short-Term Memory (ST-LSTM), and Spatio-temporal Multi-Layer Perceptron (ST-MLP), respectively. Similarly, for turbine T05, the proposed model attained an average MSE of 0.090, achieving MSE reductions of 38.45%, 23.08%, 2.93%, and 25.96% against the same benchmark models, demonstrating robustness across turbines. To further evaluate computational efficiency and scalability, a training-budget sensitivity analysis was conducted by comparing a DVTransformer trained for a single epoch against a fully trained Transformer. The results showed that the DVTransformer achieved comparable prediction accuracy across all turbines while reducing significant computational time. Evident from incorporating reliable spatial information, employing predictive wind conditions and using Transformer to capture long and short-term dependencies within time sequences increased the overall performance of the proposed method.