This paper proposes a deep reinforcement learning method for uncertainty aware scheduling to address the problem of power system scheduling decisions being susceptible to prediction bias, related disturbances, and extreme scenarios under high proportion wind and photovoltaic power integration conditions. Firstly, construct a scheduling environment that includes joint error characterization of wind power, photovoltaic power, and load, and explicitly embed multi-source related deviations into the state space. Secondly, design a Soft Actor Critic (SAC) scheduler that integrates risk sensitive rewards and safety action mapping layers to achieve coordinated optimization between operating costs, wind and solar power curtailment, carbon emissions, insufficient backup, and constraint violations. Based on publicly available time series data and combined with typical days, extreme disturbances, and sample scenarios outside the training set for validation. The results showed that the total operating cost of the proposed method was 52.47 × 104 CNY/day, a decrease of 3.39% compared to the original SAC method and a decrease of 7.75% compared to Model Predictive Control (MPC). And the wind and solar abandonment rate is 3.79%, the constraint violation rate is only 0.21%, and the average single step solving time is 0.045 seconds. At the same time, this method shows better stability and generalization ability under high uncertainty, cross month testing, and extreme weather conditions. Research has shown that this method can provide intelligent decision support with engineering feasibility for online scheduling of new power systems.
