A hybrid wind speed prediction model using improved CEEMDAN and Autoformer model with auto-correlation mechanism

This study addresses the critical need for precise and reliable wind speed predictions in the context of global environmental challenges and the increasing demand for sustainable energy. To overcome the challenges posed by the unpredictability of seasonal and stochastic winds, a novel and hybrid methodology is proposed in this study. The proposed hybrid methodology consisting improved version of the data denoising algorithm complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN), and Autoformer (AF) architecture with an Auto-Correlation (ACE) mechanism for the wind speed prediction (WSP). ICEEMDAN solves the problems in CEEMDAN: mode-mixing, aliasing, and noise. AF model uses a series decomposition block to enables the gradual aggregation of long-term trends from intermediate predictions. ACE mechanism in AF is distinct from self-attention, showing better efficiency and accuracy. The proposed hybrid model is evaluated using wind speed data from Block Island and Gulf Coast wind farms. The performance of current WSP methods is observed to decline with increasing time horizons. Addressing this problem, the proposed hybrid methodology's effectiveness is evaluated using eight separate models and eight hybrid models over six time horizons: 5-min, 10-min, 15-min, 30-min, 1-hour, and 2-hour ahead WSP. Results from the two conducted experiments demonstrate that the proposed methodology demonstrated enhanced performance, leading to a statistically significant improvement across all assessed time horizons.