In order to improve the performance of transient stability predictions based on convolutional neural networks(CNNs) and demonstrate multi-angle auxiliary decision-making information, such as transient stability classification and margin, etc., this paper proposes a multi-task model for transient stability prediction using the multi-level feature fusion based spatial pyramid pooling convolutional network. Firstly, the short-time disturbed trajectories of each generator can be obtained by phasor measurement units (PMUs), and a three-dimension information matrix may be constructed using these trajectories. Then, the spatial pyramid pooling layer and multi-level pooling layer can be employed for extracting and fusing multi-scale and multi-level feature based on the CNN. Finally, the MSPP-net is built by hard parameter sharing, so as to achieve transient stability classification, critical generators identification and stability margin prediction. The case studies have been carried on an IEEE 39-bus system, an IEEE 145-bus system and a certain provincial power grid. The results in comparison with those results from the conventional methods show that the proposed methodology is valid, and it shows the applicability when the information of PUMs is incomplete or contains noise.