This paper proposes to improve Automated Essay Scoring (AES) models by effectively encoding rhetorical information based on rhetorical structure theory (RST). Each essay is first parsed into an RST tree and then incorporated into the main AES model with two kinds of encoding methods, involving RST sequence encoding and feature vector encoding. Detailed experimentation on the Automated Student Assessment Prize (ASAP) dataset and its upgraded dataset (ASAP++) demonstrates that our best AES model has a significant gain not only in performance but also in coherence measurement capability compared with other strong AES baselines. Moreover, the auxiliary experimental results show the proposed RST encoding method is universal and can be easily applied in other classic neural network-based AES models.