Abstract: Timely forecast of stored grain condition is a necessary means to ensure the safety of stored grain. At present, traditional forecasting methods mostly predict the stored grain situation from one side, and cannot achieve accurate comprehensive assessment of the risk of stored grain situation. However, deep learning methods have bottleneck problems such as a large number of required training samples, high training difficulty and long training time. In view of this situation, by using the feature extraction and fusion method based on broad learning and the training method based on incremental learning, and combined with the multi-modal characteristics of grain situation data, a grain situation risk prediction model was proposed on the basis of the existing framework of the broad learning system. The results showed that, compared with the existing deep learning model, the training difficulty and time-consuming of the model were greatly reduced without reducing the accuracy of prediction. The predictive model proposed in the paper may be an effective alternative to deep learning models.