Multivariate Wind Speed Forecasting based on Improved TDformer

The wind speed exhibits by diverse periodicity and trend, directly inputting the wind speed sequence into the model to forecast significantly impacts the prediction accuracy. By using decomposition algorithms to decompose the wind speed sequence, the decomposed sequence serves as additional variables. Modeling with these decomposed components, along with factors such as temperature, humidity, wind direction, and dew point temperature, can enhance prediction accuracy. This paper focuses on wind speed as the research target, collecting influencing factors such as temperature, humidity, wind direction, and dew point temperature. VMD (variational mode decomposition) is utilized to decompose the wind speed into multiple components. These decomposed components, along with the collected influencing factors, are jointly used as features inputted into the TDformer model for wind speed prediction. Additionally, TSMixer is employed to enhance the trend prediction module of TDformer. Through ablation experiments and comparative experiments, the results indicate: (1) VMD leads to a decrease of over 40% in MAE and MSE. (2) Utilizing TSMixer to improve TDformer results in a decrease of 3%-20% in MAE and MSE. Moreover, the improved model exhibits over 8% improvement in MAE and MSE metrics compared to other time series forecasting models on this dataset.