The increasing penetration rate of wind power and photovoltaics （PV） in new power systems exacerbates voltage fluctuations in distribution networks, while energy storage （ES） and electric vehicles （EV） play important roles in reducing voltage fluctuations in distribution networks. At the same time, smart meters, smart sensors and improved communication networks are widely deployed, the amount of data available is increasing, and data-driven technology is emerging.This paper proposes a multi-agent deep reinforcement learning （MADRL） based dual time scale active and reactive power coordinated voltage control strategy for distribution networks.Using the double deep Q-network algorithm （DDQN） to solve the optimization problems of capacitor banks （CBs）, on line tap transformers （OLTC）, and ES active and reactive power at a slow time scale. At a fast time scale, the EA-MASAC algorithm with attention mechanism is used to enhance the reactive power of PV, wind turbines （WT）, and static var compensators （SVCs）, as well as the active power of EVs. Finally, the effectiveness of the proposed method is verified on an IEEE-33 node system.