Solar irradiance prediction using K-MTransformer model based on data input from K-means++ clustering

The global warming and environmental pollution issues are accelerating the development of renewable energy. As a key form of renewable energy, photovoltaic power generation relies on the stability of solar irradiance for its efficiency, making the accurate prediction of solar irradiance crucial. Current mainstream single-irradiance prediction models face problems such as difficulty in capturing periodic features, low feature utilization, and insufficient generalization ability. Therefore, this paper proposes the K-MTransformer method for solar irradiance prediction to improve accuracy under complex conditions. The method first uses the K-means++ algorithm to cluster time-series meteorological data and label each time series. It then employs a Deep Feature Analysis module to extract temporal characteristics, capturing both periodic patterns and sudden dynamics. Finally, by integrating a MoE model gating unit, it adaptively selects expert submodels for prediction based on cluster labels and features extracted from the time series data, achieving precise irradiation forecasting. The model's effectiveness is validated using three public datasets. The experimental results show that the K-MTransformer model achieved RMSEs of 69.3681 W/m2, 49.4675 W/m2, and 31.1840 W/m2 across the three datasets, all significantly lower than comparison models, highlighting its superior performance in irradiance prediction.