Abstract: To address the short-term scheduling challenges of new power systems, a deep learning driven scheduling decision support framework is constructed. The CNN-LSTM attention model and improved PPO algorithm are used to achieve source load prediction and multi-objective optimization. The experiment shows that the MAPE of load forecasting is ≤ 2.1%, the MAPE of new energy output forecasting is ≤ 3.5%, the dispatch cost is reduced by 12%, and the new energy consumption rate in conventional/extreme scenarios reaches 95% and 91.2% respectively. The method has better efficiency and robustness, and can provide technical support for short-term dispatch of power systems.