Accurate short-term solar irradiance forecasting is essential for reliable solar power integration and grid stability, yet remains difficult due to rapid, cloud-driven fluctuations. Traditional irradiance sensors offer real-time measurements but lack predictive capability. Sky cameras, on the other hand, provide visual information that supports intra-hour forecasting. This study implements a convolutional neural network (CNN)-based model to forecast solar irradiance up to 60 min ahead using sky imagery. Mesa, Arizona, was selected as the study site due to its flat terrain and arid desert climate, making it ideal for image-based irradiance forecasting. A fisheye sky camera captured high-resolution (1056 x 1056 pixel) images, which were downscaled to 128 x 128 pixels and synchronized with pyranometer data collected at the same location. With this simple configuration and a standard CNN model, we achieved RMSE values between 66-118 W/m2, depending on sky condition. Compared to the persistence model, our approach reduced the 15 min forecast error by over 22.4 %. The model showed the best performance under mixed sky conditions, effectively capturing complex cloud dynamics. These results demonstrate the promise of combining sky imagery and deep learning techniques to enhance short-term solar forecasting and support more efficient grid management.