The state prediction of key components in manufacturing processes plays an important role in intelligent manufacturing, as it could improve the production quality, efficiency and reduce costs. Data-driven methods could learn well-performed prediction models from large volume of data. However, in complex manufacturing systems, the lack of prior knowledge limits the performance of prediction models, where the manufacturing environments changes continuously. In order to address this issue, this paper proposed a zero-shot prediction method for complex manufacturing systems based on causal inference. A deep convolutional neural network and a causal representation model are jointly optimized to extract invariant causal signal features, which can be generalized to non-stationary manufacturing environments without any new data. The experiment of tool wear prediction under non-stationary working conditions is carried out as a research example. The proposed method is verified with the open dataset on tool wear prediction, and experimental results show that the prediction accuracy could be obviously improved over existing methods.