Abstract: The disadvantage of the Simulated Annealing(SA) algorithm and BP algorithms was analyzed in this paper, design an optimization network algorithm, organic combine SA and Powell, insteaded of gradient falling algorithm of BP network to train network weight. it can get high accuracy and fast convergence speed. Today using multiple seismic information to predict thin interbeded reservoir parameter(thickness of grit,porosity)is an important research subject in the petroleum exploration. Thin interbeded reservoir have the features of reservoir thickness thin and change violently in horizontal, Traditional BP network can not get contentment result to predict thin interbeded reservoir parameters. Using the new algorithm in this paper can improve predict accuracy and convergence speed . The prediction results of an example proves this algorithm feasible and practicablity.