In this paper, we propose a conservation-informed end-to-end learning framework to achieve accurate human motion prediction based on conservation laws. Existing methods, which rely purely on data-driven approaches, often produce physically implausible results due to their neglect of the physical principles of human movement. This shortcoming often leads to prediction results that are not applicable to real-world scenarios. To address these challenges, our proposed Conservation-informed Neural Network (CiNN) treats the human body as an independent spatiotemporal dynamical system. By solving the conservation equations that describe the rules of human motion, CiNN calculates the next frame of human movement. The physical constraints provided by the conservation laws effectively mitigate the inherent cumulative error problem in human motion prediction. Extensive experiments demonstrate that our proposed conservation framework generates results that are more in line with physical laws and shows significant performance improvement in long-term prediction.