Deep Reinforcement Learning Strategy for Electric Vehicle Charging Considering Wind Power Fluctuation

Electric vehicles (EVs) can inhibit the wind power fluctuations in the generalized form of energy storage. However, optimizing the charging process of EVs under wind power fluctuations is difficult because of the uncertainties of wind power output and user demands. A charging control strategy based on deep reinforcement learning (DRL) was proposed in this study to address the influence brought by uncertain environmental factors to the control. This strategy mined the deep relation between perceiving the uncertainties of environmental factors and learning charging laws by virtue of the perceptual and learning abilities of DRL. An immediate reward mechanism that acts upon the environment was constructed from the angle of neural network fitting function. The EV charging control model was expressed as a Markov decision process (MDP) that contain the state, action, and transfer functions and reward and discount factors through temporal discretization. Next, the single-step updating and experience replay mode were combined to construct the DRL algorithm, followed by the comparative convergence experiment with the reinforcement learning (RL) algorithm that expressed the reward function in mathematical form. In the end, the agent obtained through training was used for the verification of the calculated example. Results show that the constructed RL algorithm is converged by 8,500 episodes earlier. The charging control strategy based on DRL meets the charging requirements when the proportion of optimization objectives is 0.5 and 0.9, and users are allowed to change the allowed charging time temporarily. This study demonstrates that the charging control strategy based DRL can optimize the EVs charging process under many uncertain factors.