The demand forecasting plays a crucial role in the predictive physical and virtualized network management in cellular networks, which can effectively reduce both the capital and operational expenditures by fully exploiting the network infrastructure. In this paper, we study the per-cell demand forecasting in cellular networks. The success of demand forecasting relies on the effective modeling of both the spatial and temporal aspects of the per-cell demand time series. However, the main challenge of the spatial relevancy modeling in the per-cell demand forecasting is the irregular spatial distribution of cells in a network, where applying grid-based models (e.g., convolutional neural networks) would lead to degradation of spatial granularity. In this paper, we propose to model the spatial relevancy among cells by a dependency graph based on spatial distances among cells without the loss of spatial granularity. Such spatial distance-based graph modeling is confirmed by the spatiotemporal analysis via semivariogram, which suggests that the relevancy between any two cells declines as their spatial distance increases. Hence, the graph convolutional networks and long short-term memory (LSTM) from deep learning are employed to model the spatial and temporal aspects, respectively. In addition, the deep graph-sequence model, graph convolutional LSTM, is further employed to simultaneously characterize both the spatial and temporal aspects of mobile demand forecasting. Experiments demonstrate that our proposed graph-sequence demand forecasting model could achieve a superior forecasting performance compared with the other two proposed models as well as the traditional auto regression integrated moving average time series model.