This paper proposes walk pattern modeling for powered exoskeleton based on complex-valued neural network and reports its validity through experiments. We have been developing a powered exoskeleton to support workers at a nuclear power plant in time of hazard. The objective of the powered exoskeleton is to support a worker wearing a heavy radiation protection suit. We believe that conventional reactive power assist control based on EMG sensors is not feasible because they fail to measure the worker muscle activity robustly in the radiation protection suit as the worker has a lot of sweat in the high temperature and the humidity in the suit. Therefore, we have developed feed-forward control to assist the worker's motion based on the recognition of the worker's motion. Our previous studies use a simple k-nearest neighbor algorithm to model the motion of the powered exoskeleton, however, the algorithm is not for online learning and the estimated trajectory is not as smooth as we expected. This paper proposes a new modeling of walk motion of the powered exoskeleton based on a complex-valued neural network. The complex-valued neural network generally has good properties on learning speed and stability. This paper shows its validity for the modeling of powered exoskeleton's walk through experiments.