Long-term solar irradiance prediction can be more effective to plan and manage solar power systems. Existing methods have demonstrated the effectiveness of decomposing time series to enhance solar irradiance prediction models. However, these methods still have limitations in extracting potential sequence information from complex data sources. This makes the long-term prediction of irradiance still uncertain. In this paper, an improved complete ensemble empirical mode decomposition adaptive noise (ICEEMDAN) combined with TimesNet (ICEEMDAN-TimesNet) method is proposed. The ICEEMDAN method decomposes the subsequence features of the original irradiance, and the multidimensional spatial mapping of the data in the TimesNet model ensures that the model can efficiently extract historical information for the week-ahead hourly prediction of irradiance. In order to explore the performance of ICEEMDAN-TimesNet in the solar irradiance prediction task, several multi-step prediction models were established for comparison. Overall, the ICEEMDAN-TimesNet framework can report satisfactory testing results. The prediction curve is fitted to the trend of irradiance change which can play an excellent guiding role in scheduling electricity generation.