Forecasting the crowd: An effective and efficient neural network for citywide crowd information prediction at a fine spatio-temporal scale

•A new method to predict citywide crowd information at fine spatio-temporal scales.•Evaluating the proposed method with five baselines on two different datasets.•Our method has advantages in both predictive accuracy and training time costs.•It can better predict sudden massive crowds gathering with low training time costs.•It can make better predictions for regions with high and irregular crowd changes.