In mineral processing industry, accurately predicting concentrate grade can help engineers adjust process parameters in advance and improve flotation performance. However, the prediction accuracy of concentrate grade has been restricted by limited sample sizes, high-dimensional data and complex temporal correlations in actual mineral processing. A time-series data generation model called LS-WGAN is proposed according to the problem of prediction for small sample data, which combines wasserstein generative adversarial network (Wasserstein GAN) and long short-term memory (LSTM) neural network. LSTM network is mainly used to capture the time correlation in mineral processing data, then Wasserstein GAN is used to generate samples similar to the original data distribution for data augmentation. In order to improve concentrate grade prediction accuracy, a mineral processing prediction model called C-LSTM is established, the prediction accuracy of the proposed method is verified by experiments based on real froth flotation process data.