Because back-propagation(BP) neural networks can not avert partial minimum and genetic algorithms(GA) is of property of the good global searching, a novel approach, GA-BP neural networks is presented to analyze and forecast deformation data. The novel approach optimizes initial weights of BP neural networks by GA, then implements the learning of networks by BP. Experimental results show that this approach remains stable number of learning and stable final value of weights, and is better than standard BP neural networks in accuracy, speed and robustness for the forecast of deformation data.