The high cost and difficulty of data collection result in discrete intermittency and limited available data for most datasets, posing challenges to modeling and prediction needs. Thus, it is crucial to explore methodologies suitable for discrete small sample data. Concurrently, unsupervised learning methods are employed for small sample prediction to circumvent the issue of low prediction accuracy resulting from inadequate a priori knowledge. Current data modeling and prediction methods rely heavily on neural network deep learning models, which are mainly designed for continuous processes and show low accuracy for intermittent processes. This paper proposes a data expansion method based on Time Generative Adversarial Networks (TimeGAN) and a prediction method based on the combination of Generative Adversarial Networks (GAN) and Deep Forest (DF). Firstly, the TimeGAN model is proposed to learn small sample data for data expansion. Then, the GAN-DF model is constructed so that the model is used for data prediction of unsupervised discrete processes. On this basis, we introduce the DFGAN and LSTM-DFGAN models, which embed DF and LSTM-DF models as generators within the GAN framework to enhance prediction accuracy for small sample data. Simulation validation using contaminant data from a food processing process demonstrates the efficacy of the TimeGAN method in data expansion. The results demonstrate that the TimeGAN method is capable of effectively expanding the data, while the LSTM-DFGAN model exhibits the most optimal combined prediction effect.