Power system real-time dispatch (RTD) problem optimizes generator outputs considering operational constraints and high-dimensional system information such as nodal loads and network topology. Conventional deep reinforcement learning methods aim to find optimal dispatch decisions through random exploration and lead to low computational efficiency. In this paper

we decompose RTD into three subproblems

namely unit commitment

power adjustment and voltage adjustment. A reinforcement agent is constructed using an expert knowledge-based twin delayed deep deterministic policy gradient algorithm (EK-TD3) to give power adjustment decisions

while unit commitment and voltage adjustment decisions are based on capacity-rules and volt-var sensitivity rules

respectively. Regularization terms based on expert knowledge are introduced to guide exploration of the agent

including generation-load mismatch

midpoint-maintaining rule of balance generator

renewable energy consumption rate and reserve capacity availability. Weight of different regularization terms can be adjusted periodically to reflect priority of dispatch objectives. The proposed method is tested on the State Grid 126-bus system (SG-126). Results prove the proposed method has better convergence performance compared with the conventional deep reinforcement learning method

enabling its application in large-scale power systems.