Electric power and transportation networks become increasingly coupled through electric vehicles (EV) charging station (EVCS) as the penetration of EVs continues to grow. In this paper, we propose a holistic framework to enhance the operation of coordinated electric power distribution network (PDN) and urban transportation network (UTN) via EV charging services. Under this framework, a bi-level model is formulated to optimally determine EVCS charging service fees (CSF) for guiding EV charging behaviors and minimizing the total social cost. At the upper level, PDN with wind power generation is formulated as a second-order cone problem (SOCP) where CSF is determined. Given the settings calculated at the upper level, the lower level problem is described as a traffic assignment problem (TAP) which is subject to the user equilibrium (UE) principle and captures the individual rationality of single EV owners in UTN. The uncertainties in wind power output and origin-destination (O-D) traffic demands are considered in the proposed model and a deep reinforcement learning (DRL)-based solution framework is developed to decouple and approximately solve the stochastic bi-level problem. Both gradient-based and gradient-free training algorithms are implemented in this paper and the respective results are compared. The case studies on a 5-node system, 24-node Sioux-Falls system and real-world Xi'an city in China are conducted to verify the effectiveness of the proposed model, which demonstrates the enhanced operation of coordinated PDN and UTN networks by reducing the traffic congestion and improving the integration of renewable energy.