Decomposed solar radiation models are commonly used to separate direct and diffuse irradiance from global irradiance. However, most of these models are designed to process hourly data, which may not be sufficient to capture the rapid changes in solar irradiance that occur within a shorter timescale. To address this issue, we examined the performance of existing decomposition models at different temporal resolutions ranging from 1 min to 1 h. We found that the errors in the decomposition models increased as the temporal resolution decreased. Specifically, as the timescale was reduced from hourly to every minute, the relative root-mean-square error(rRMSE) increased by more than 5%. These findings highlight the need to develop accurate models that can process sub-hourly data. Accordingly, we propose the use of deep learning models to estimate the direct irradiance using sub-hourly data. The proposed models significantly reduced the rRMSE by more than 7% compared to the existing models on a 1-min time scale. The results indicate that deep-learning models can provide accurate estimates of direct irradiance, even at sub-hourly temporal resolutions.