Due to the defects of traditional genetic algorithm(GA) such as premature and poor local search ability, a simulated annealing algorithm(SA) is introduced to modify GA for better optimizing result. Afterwards, the modified genetic simulated annealing algorithm(GSA) is implemented to search for the optimal weight and threshold of BP neural network, which improves the prediction accuracy of BP neural network by overcoming its drawbacks of local minimum and slow convergence. Thus, GSA-BP neural network model is established for the back analysis of surrounding rock mechanical parameters. Finally, the model is applied to an engineering case, Wudongde Power Station, to regress surrounding rock mechanical parameters of the underground powerhouse on the right side with in-situ measured displacement data. By applying the regressive mechanical parameters to numerical model, displacements of surrounding rock are computed; and the computed displacements agree well with the measured ones; which indicates the GSA-BP neural network model is feasible for back analysis of surrounding rock mechanical parameters in real-world engineering cases.