Abstract: Existing irradiance forecasts are mostly intra-hour models built on coarse weather-type labels, ignoring the intraday weather evolution and the impact of cloud type on irradiance. To address these limitations, a day-ahead direct normal irradiance (DNI) forecasting framework that couples complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), K-means density-based spatial clustering of applications with noise (KDBSCAN) and the Informer model was proposed. Firstly, principal component analysis was employed to compress the dimensions of raw meteorological variables. The derived principal components were then decomposed and reconstructed via CEEMDAN. KDBSCAN was subsequently employed to perform a two-layer clustering, weather-type clustering and intra-weather-type cloud-pattern clustering. Finally, based on the two-layer labels, the day-ahead DNI was forecasted via the Informer network. Results show that by classifying cloud types to consider weather changes throughout the day in the prediction process, the determination coefficient (R²) for predicting irradiance before sunny days reaches 0.98. Compared with prevailing benchmarks, the proposed framework reduces the overall mean absolute error by 32.42 W/m² and root-mean-square error by 31.71 W/m², achieving overall R² of 0.97.