DUSTFN: A dual-path U spatiotemporal fusion network for wind speed prediction

Accurate short-term wind speed prediction is crucial for maintaining wind power systems’ safe, stable, and efficient operation. Therefore, we propose DUSTFN, a fully convolutional network, that fuses multi-source meteorological data, for fine-grid wind speed prediction studies. DUSTFN models the spatial and temporal evolution of the wind and auxiliary variables using a dual-path U-shaped spatiotemporal network and uses multi-source spatiotemporal fusion units to capture the uncertainty of the wind from the a priori evolution of the geopotential and temperature evolution. Finally, wind magnitude and structure are integrated using homoscedasticity uncertainty equilibrium loss to constrain model training. The experiments show that the mean absolute errors of DUSTFN predicted wind speeds are 0.27 m/s at the 1st hour, 0.5 m/s at the 3rd hour, and 1 m/s at the 12th hour; the accuracy of wind direction prediction for eight directions reaches 98.77 % at the 1st hour and 95.85 % at the 3rd hour. The transferability and efficiency experiments demonstrate the outstanding generalization performance and prediction speed of DUSTFN, which can be quickly applied to off-site prediction. Therefore, with its accuracy and efficiency, DUSTFN can be an effective tool for predicting vector wind speeds in large regional wind power centers and can help in ultra-short and short-term deployment planning for wind power.