The invention relates to a genetic algorithm and Kalman filtering based RBFN (Radial Basis Function Networks) combined training method which comprises the following five steps: I, setting a random initialized population according to parameters to code central values: v11, v12,..., v1m, v21,..., and vcm; II, calculating adaptive values of individuals in the population and storing the optimal adaptive value, wherein the target of training RBFN is to minimize an output error E and a fitness function is set as follows: Fit(fi)=1/E; III, if set evolutionary algebra is realized or a current optimal individual satisfies the condition, returning network parameters and skipping to the step IV, otherwise, skipping to the step II after selection, crossing and genetic variation operation; IV, correcting the network parameters in a self-adaptive manner by a Kalman filtering algorithm, wherein the network parameter value optimized by a genetic algorithm is taken as the network initial parameter of the Kalman filtering algorithm; and V, ending a program when the maximum iterative time limitation is realized or the current network error meets the requirement, otherwise, skipping to the step 4 to operate continuously.