Predicting the grain losses and waste rate (LWR) is critical for agricultural planning and grain policy development. Capturing the stage interaction and generating robust features are the main challenges in grain LWR prediction. In this paper, we propose MTGA, a Multi-Task GRU Autoencoder approach to: 1) obtain the robust feature representation for prediction task and 2) explore the time-ordered interaction among different stages of grain chain. Specifically, we propose multiple GRU encoder-decoder pairs to co-reconstruct the features in a common space for robust feature learning. Then, an attention mechanism is proposed to better fuse the reconstructed features from the GRU encoder-decoder pairs. Furthermore, we utilize the multi-task for reconstructed loss and grain LWR prediction. We introduce the reconstructed loss task as an auxiliary task to help us better represent the robust features. Besides, we propose the LWR prediction as main task to learn the parameters for prediction task. We collected the data with questionnaires, interviews or data from grain management institutes for experiments. The evaluation results show that grain LWR prediction by our approach achieves the best results compared to several state-of-the-art prediction model. Moreover, our method gains overall performance decline of 12.5%-18.3% on MAE and RMSE metrics.