Effective rehabilitation methods are essential for the recovery of lower limb dysfunction caused by stroke. Nowadays, robotic exoskeletons have shown great potentials in rehabilitation. Nevertheless, traditional rigid exoskeletons are usually heavy and need a lot of work to help the patients to put them on. Moreover, it also requires extra compliance control to guarantee the safety. In contrast, soft exoskeletons are easy and comfortable to wear and have intrinsic compliance, but their complex nonlinear human-robot interaction dynamics would pose significant challenges for control. In this work, based on the pneumatic actuators inspired by origami, we design a rehabilitation exoskeleton for knee that is easy and comfortable to wear. To guarantee the control performance and enable a nice human-robot interaction, we first use Deep Koopman Network to model the human-robot interaction dynamics. In particular, by viewing the electromyography (EMG) signals and the duty cycle of the PWM wave that controls the pneumatic robot's valves and pump as the inputs, the linear Koopman model accurately captures the complex human-robot interaction dynamics. Next, based on the obtained Koopman model, we further use Model Predictive Control (MPC) to control the soft robot and help the user to do rehabilitation training in real-time. The goal of the rehabilitation training is to track a given reference signal shown on the screen. Experiments show that by integrating the EMG signals into the Koopman model, we have improved the model accuracy to great extent. In addition, a personalized Koopman model trained from the individual's own data performs better than the non-personalized model. Consequently, our control framework outperforms the traditional PID control in both passive and active training modes. Hence the proposed method provides a new control framework for soft rehabilitation robots.