Efficient Training of Graph Convolutional Networks for Dynamic Phenomena