Near-infrared (NIR) spectroscopy has been widely used to predict the gasoline properties that are difficult to measure online during gasoline blending. NIR models should be prepared in advance to apply this technique successfully. Obtaining a high-accuracy NIR model in practice is hard because abundant labelled samples are difficult to acquire. A new modelling method on the basis of Wasserstein generative adversarial network is proposed in this study to overcome this weakness. Abundant artificial labelled samples are generated firstly using the proposed method, and sample selection is performed to select the appropriate artificial samples. Real and selected artificial samples from the selection results are combined to train the NIR model that could be established efficiently when labelled samples are scarce. An actual dataset obtained during gasoline blending is provided to validate the effectiveness of the proposed method, and several traditional methods are adopted for comparison.