Attention-SAC: a DRL approach for optimal operation of new-type power systems

Under the dual-carbon target and increasing renewable energy penetration, new-type power systems face significant challenges due to high-dimensional state variables, strong source-grid-load-storage coupling, renewable energy uncertainty, and conflicting dispatch objectives. This study is motivated by the limitation of existing optimization and reinforcement learning methods in simultaneously achieving economic efficiency, renewable energy utilization, and operational security under complex nonlinear operating conditions. It is hypothesized that incorporating a feature-wise attention mechanism into the Soft Actor-Critic framework can enhance the extraction of critical dispatch information and improve adaptive scheduling performance. Therefore, an attention-enhanced Soft Actor-Critic method (Attention-SAC) is proposed for optimal operation of new-type power systems. A comprehensive dispatch model integrating thermal power, hydropower, wind power, photovoltaic generation, energy storage, local load, and external power demand is established, where operational constraints and a multi-objective reward function are considered to coordinate economic, environmental, and security objectives. Furthermore, a lightweight feature-wise self-attention encoder and a balance-guaranteed action projection mechanism are introduced to improve state representation and ensure physically feasible dispatch decisions. Comparative simulations with proximal policy optimization (PPO), conventional soft actor-critic (SAC), twin delayed deep deterministic policy gradient (TD3), advantage actor-critic (A2C), and deep deterministic policy gradient (DDPG) demonstrate that Attention-SAC achieves faster convergence, improved operational stability, reduced operating cost, and enhanced renewable energy utilization. The results verify the effectiveness of the proposed framework and indicate its potential for intelligent low-carbon dispatch of complex power systems.