With the expansion of power systems, in order to satisfy the real-time requirements of large-scale power system calculation, a mass expression evaluation parallel algorithm based on 'expression forest' was proposed. A batch of expressions can be converted to an optimal parallel expression forest by a generation algorithm. Then, based on GPU-CPU framework, the proposed parallel algorithm can realize expression evaluation. This algorithm has universal applicability in varieties of computational fields. In the case analysis, this algorithm was used to realize parallel formation of large-scale Jacobian matrices in the Newton method for power flow calculation. Numerical results show that, the proposed algorithm has high efficiency with small memory usage, which could be generalized to other power system calculations and applied to engineering practice.