MIG-EWPFS: An ensemble probabilistic wind speed forecasting system integrating multi-dimensional feature extraction, hybrid quantile regression, and Knee improved multi-objective optimization

Wind speed exhibits strong randomness and intermittency, posing technical challenges to wind energy utilization, especially with a high level of operational uncertainty in wind farms. Therefore, developing precise Probabilistic Wind Speed Forecasting (PWSF) techniques can offer critical insights for decision-making in power systems. Given this, a novel PWSF technique is developed, which fuses an advanced channel attention mechanism to recalibrate the important differences among features. However, the limitations of single models prevent effective mining of the interactions among wind speed and meteorological factors. To break this bottleneck, an adaptive ensemble PWSF architecture based on multi-dimensional feature extraction (MFE) and multi-model fusion is further explored. This framework begins with an enhanced fuzzy rough set that adheres to specific granularity criteria to capture the temporal coupling features among wind speed and meteorological factors. Simultaneously, a library of candidate models is established, upon which multiple high-performing candidates are integrated using an improved Pareto optimization strategy. This technique fully utilizes the advantages of each candidate, effectively mitigating the “No Free Lunch (NFL)” phenomenon. Using wind speed uncertainties in the U.S. and China, the proposed ensemble system demonstrates superior interval sharpness and overall forecasting accuracy, along with significant reliability compared to optimal benchmarks and ensemble methods.