Genetic algorithm is a non-derived random optimization method based on the regulations of nature selection and evolution. Generally, when genetic algorithm is applied to wavelength selection, only how to improve the prediction accuracy of a regression model is concerned. But the robustness of a regression model, that is, the anti-interference ability towards external measuring conditions variance (such as the ambient temperature, place and so on), is usually ignored. Therefore, when the measuring conditions of the predicted samples change, the regression model would predict the measured samples with high prediction errors. In this paper, genetic algorithm combined with experimental design method was studied to increase the robustness of the multivariate calibration model. In our experiments, the training set was divided into the calibration set and the monitor set to establish the regression model. The spectra of the calibration set samples were measured under the ordinary measuring conditions. The measuring conditions when obtaining the monitor set sample spectra could be arranged according to experimental design method. Kennard/Stone algorithm was used to select monitor set samples from training set. The calibration model could be built with the calibration set samples and optimized with the monitor set samples measured under the designed measuring conditions. And then the validation set samples, which were independent of the training set ones, were employed to evaluate the prediction ability of the regression model. In order to obtain a regression model with high prediction accuracy and robustness, the spectral information caused by the changes of measuring conditions need to be considered and those wavelengths which were easily interfered by external measuring factors need to be rejected when the calibration model was trained. In this paper, the modified wavelength selection method of genetic algorithm was applied to the temperature experiments of the glucose aqueous solution samples. Results revealed that not only fewer wavelengths or principal components were needed to build the calibration model but also the robustness and prediction accuracy of the calibration model were greatly improved. This modified method not only makes the regression model insensitive to external measuring conditions, but also could be applied to the calibration transfer between different instruments of the same type.