Abstract: When a high proportion of intermittent distributed generators（DGs） and electric vehicles are connected to the distribution network, it is easy to cause frequent, fast and dramatic fluctuations in power and voltage. This paper combines both the data-driven and physical modeling approaches to propose a coordinated optimal strategy for dual-timescale active and reactive power in distribution networks. For the short-timescale（minute or second） power fluctuations, a second-order cone programming（SOCP） model and a quadratic programming model are constructed for balanced and unbalanced distribution networks, respectively, with static var compensator（SVC） and reactive power of DG as decision variables and network loss minimization as the objective function, taking into account physical constraints. For the long-timescale（hour scale） optimization, the Markov decision process（MDP） is constructed with the tap ratio of on-load transformer changers（OLTCs）, the tap position of switchable capacitors reactors（SCRs）, and the charging/discharging power of energy storage systems（ESSs） as the actions, network loss as the cost, taking into account the crossing penalty of bus voltage. To overcome dimension curses in continuous-discrete action space, this paper uses a relaxation-prediction-correction based deep deterministic policy gradient（DDPG） reinforcement learning algorithm. The effectiveness of the proposed method is verified by IEEE 33-bus and IEEE 123-bus distribution systems.