Publication

Abstract : To address the significant long-term charging costs associated with Plug-in Electric Vehicles (PEVs), this paper proposes a demand response (DR) strategy that considers various factors such as traffic conditions, fluctuating energy prices, energy consumption patterns, and the stochastic nature of driving behavior. Therefore, employing a Dyna-Q reinforcement learning approach in EV charging enables rapid learning and adaptation to other unknown parameters, such as energy prices and driving patterns Deep reinforcement learning techniques are utilized to address the problem, which is a Markov Decision Process (MDP) with uncertain transition probabilities. Unlike existing methods that either delay convergence or rely on user data, the Dyna-Q algorithm integrates both model-free and model-based learning, updating the model with each real encounter. A deep neural network-based value approximation method determines the longterm expected reward of state-action pairs because the state space is too big for a table-look-up method to work. Historical values are averaged to forecast future prices. Without the initial PEV owner data, simulations of three different user behaviors show that a Dyna Q network (DQN) method that does not use any models often leads to battery drain during trips. The results indicate that the recommended approach improves upon existing PEV charging techniques by more quickly and efficiently determining the optimal charging strategy.