The accuracy of photovoltaic (PV) power prediction significantly impacts power grid dispatching and economic benefits, as PV generation output is heavily influenced by various meteorological factors. However, existing prediction methods struggle to effectively process complex nonlinear relationships among multidimensional meteorological features, resulting in unstable performance especially in small-sample similar-day scenarios. To address these issues, this study proposes a feature-enhanced Transformer deep learning framework integrating t-distributed stochastic neighbor embedding (t-SNE) and variational Bayesian Gaussian mixture model (VBGMM) weather clustering with intelligent optimization for ultra-short-term PV power forecasting. First, multidimensional meteorological features are preliminarily screened using Pearson correlation coefficients, followed by secondary dimensionality reduction with t-SNE. Then, VBGMM adaptive clustering is employed to identify key weather patterns. Next, a feature-enhanced Transformer (FET) model is designed to extract temporal features, while an improved osprey optimization algorithm (IGOOA) optimizes the key parameters of FET. The framework is validated using different real-world datasets, demonstrating significant advantages in both single-step and recursive multi-step prediction tasks, making it suitable for PV power prediction scenarios with multidimensional features and small sample sizes.
扫码关注我们
求助内容:
应助结果提醒方式:
