Many existing solar power forecasting studies still focus on a single, fixed forecasting resolution, with limited investigation into how model performance changes when temporal data are aggregated to coarser time scales and formulated as single-step predictions, or how related Recurrent Neural Network (RNN) architectures compare across multiple short-term forecasting intervals. Therefore, this study is conducted to evaluate the performance of four advanced RNN architectures including GRU, BiGRU, LSTM, and BiLSTM, for short-term solar power forecasting using a single-step forecasting formulation across multiple temporally aggregated time resolutions: 5, 15, 30, and 60 minutes. The results demonstrated that all RNN-based models achieved reasonable accuracy in short-term prediction, with BiGRU consistently outperforming the others at most intervals. Notably, for the 5-minute resolution, BiGRU achieved an R² of approximately 0.94 and a NMAPE of just 3.41%, confirming its adaptability and precision. Its performance remained strong for 15- and 30-minute forecasts; although a gradual decline was observed as the temporal resolution became coarser. At the 1-hour mark, BiLSTM outperformed BiGRU slightly, achieving R² = 0.85 and NMAPE = 7.01%, indicating its better ability to model long-term dependencies in aggregated solar power data. These findings not only extend the practical applications of bidirectional RNNs in renewable energy forecasting but also provide a solid foundation for developing more intelligent and efficient power scheduling systems.