Accurate intra-day forecasting of global horizontal irradiance (GHI) and direct normal irradiance (DNI) is critical for the efficient operation of solar energy systems. While satellite-based data have traditionally been used as the most relevant exogenous spatio-temporal input for such forecasts, numerical weather prediction (NWP) models with increasing resolution now offer promising benefits. This study evaluates the effectiveness of spatio-temporal satellite- and NWP-based inputs (i.e., irradiance and cloud cover forecasts) and their impacts on intra-day GHI and DNI forecasts using deep learning methods from three perspectives: single type of input, input combinations, and ensemble forecasting. Results demonstrate that the combined use of satellite- and NWP-based inputs generally shows better performance than their individual use, achieving skill scores up to 36.35% and 30.87% for 4-hour-ahead GHI and DNI forecasts, respectively. Among the three spatio-temporal inputs, NWP irradiance exerts a greater overall influence, while satellite-derived products contribute more to short-term predictions (less than 2 h ahead), NWP cloud cover is more influential for longer forecast horizons. Based on impacts of satellite- and NWP-based inputs on the forecast, two weighted ensemble strategies are proposed to further enhance the forecasting performance, where skill scores can reach 38.62% for GHI and 34.61% for DNI in 4-hour-ahead forecasts. With the growing capacity addition of solar energy systems, this work highlights the advantages of integrating spatio-temporal satellite-derived products and NWP forecasts in intra-day solar forecasting with deep learning. This offers practical benefits to a wide range of stakeholders in the solar energy community for real-world applications.