A novel interpretable wind speed forecasting based on the multivariate variational mode decomposition and temporal fusion transformer

Accurate and efficient wind speed forecasting is essential for the stable operation of wind farm and power grids. However, the high volatility of wind speed, coupled with its correlation with local meteorological factors, makes accurate wind speed forecasting a significant challenge. To achieve precise wind speed forecasting and model interpretability, this study proposes a short-term interpretable wind speed forecasting model based on the joint decomposition of multi meteorological feature data, combined with the Temporal Fusion Transformer (TFT) and the Crested porcupine optimizer (CPO) algorithm. Initially, wind speed data and various meteorological features are input into the Multi-variant Variational Mode Decomposition (MVMD) algorithm for decomposition, resulting in multiple Intrinsic Mode Functions (IMFs). The CPO algorithm will concurrently be utilized to intelligently optimize the hyperparameters of the MVMD. Mutual Information (MI) will then be employed to select the IMFs derived from MVMD decomposition that exhibit a higher correlation with wind speed. These IMFs, along with various meteorological features, will collectively form the input data for the TFT model. Subsequently, the TFT model will be used to achieve high-precision wind speed predictions and generate interpretable results. Finally, the CPO algorithm is used to finely tune the hyper parameters of the TFT, yielding the optimal hyper parameter combination. Experimental results demonstrate that compared with other common forecasting methods, the proposed CPO-MVMD-MI-CPO-TFT model offers higher forecasting accuracy. Additionally, its interpretable results can provide robust data support for decisions related to wind farm site selection and wind turbine scheduling.