Electric vehicle routing with stochastic customers using deep reinforcement learning and graph convolutional networks

In this paper, we study the electric vehicle routing problem with stochastic customer requests. We propose a novel graph convolutional network-based deep reinforcement learning framework for energy-optimal electric vehicle routing. Our proposed routing method uses Graph Convolutional Networks (GCN) to encode the global and local routing network information which helps the routing system understand the routing network structure and energy consumption information efficiently. The proposed routing model incorporates uncertainty in energy consumption using a probabilistic energy consumption estimation model. We present a heuristic trip completion reward to ensure successful trip completion of the routing agent without extra computation. We have trained and validated our proposed routing method on two different routing network datasets of different sizes. Our experimental results show that our deep reinforcement learning-based routing method outperforms other existing routing methods on large network instances and our method is able to reduce the training time significantly compared to the existing routing methods.