Day-ahead solar irradiance prediction based on multi-feature perspective clustering

Global horizontal irradiance significantly influences the performance of photovoltaic power generation, and accurate prediction is crucial for enhancing power generation efficiency and optimizing grid dispatch. However, existing single prediction models frequently encounter challenges such as limited adaptability and restricted prediction accuracy when addressing complex and variable weather conditions. To tackle this challenge, this paper proposes a multi-feature perspective clustering method that integrates time-domain and frequency-domain features. By performing similar day clustering, historical data are segmented into sub-datasets representing three weather conditions, which are then used to construct targeted prediction sub-models, thereby improving the model's adaptability and training performance. Additionally, a probabilistic reconstruction model based on Bayesian theorem is developed. This model combines the features of fuzzy C-means clustering and variational autoencoders to optimize the weighting strategy, further enhancing prediction accuracy. Experimental results on multi-regional datasets demonstrate that the proposed method significantly outperforms traditional approaches across multiple evaluation metrics, highlighting its superior performance and robust cross-regional adaptability in GHI day-ahead prediction.