Efficient grain storage management is of great significance to the country, and grain storage temperature is one of the key indicators to judge grain storage security. Accurately predicting the temperature of stored grain and making timely and appropriate protective measures can effectively reduce grain storage losses. In order to address the limitations of traditional prediction models, this paper introduces a novel network model integrating a time-domain convolutional network (TCN), a self-attention mechanism, and a bi-directional gated recursive unit (BiGRU). First of all, the local features of grain storage temperature data are obtained by TCN, and the SelfAttention mechanism is incorporated into the network according to the temporal characteristics of grain storage temperature to assign weights to different grain features, highlighting the features that have a greater impact on the prediction of grain storage temperature. In addition, BiGRU network is used to learn the bi-directional dependency of grain condition sequences to obtain more information in the sequences and achieve the prediction of grain storage temperature. The experimental results show that the RMSE of the model is 0.389 5, the MAE is 0.328 1, and the R² is 0.9912. Compared with other models, the proposed method reduces the error and improves the prediction accuracy, thus providing a solid foundation for decision-making in grain storage temperature control.