Abstract: In order to describe the complex nonlinear relationship between earthquake magnitude and its sensitive factor, the genetic algorithm ( GA ) combined with general regression neural network ( GRNN ), and the principal component analysis ( PCA ) were used to reduce the dimension of earthquake sensitive factor. Then normalizing the extracted principal components which were used as input vectors of general regression neural network, and earthquake magnitude was used as output vector, 20 typical cases of earthquake magnitude were used for training data, then using GA to optimize the best smooth factor, finally the prediction model for earthquake magnitude based on PCA-GA-GRNN was established, and it was used to predict 8 prediction samples. The result shows that the minimum error, maximum error and average error of PCA-GA-GRNN model were 51.5630%、4.8780% and 2.647% respectively. Compared with GSM-GRNN model and GRNN model, the average error of PCA-GA-GRNN model is reduced by 5.6667% and 5.0264%, respectively. Therefore, PCA-GA-GRNN model has higher prediction accuracy.