Ultra short-term photovoltaic (PV) power generation forecasting is great significance for the reliable operation of power grid. However, existing forecasting method generally focus on mining the temporal characteristic of PV power time series, without considering the interactions between multiple PV sites. Graph neural network has shown excellent performance in modeling spatial correlation of non-Euclidean structured data, but still has some drawbacks. Specifically, it predefines graph structures based on prior information, which all are static. While the spatial dependence between PV sites changes over time due to uncontrollable meteorological factors. To address this problem, this paper designs a dynamic spatial-temporal graph neural network (DSTGNN) for ultra short-term PV power generation forecasting. it consists of a dynamic graph learning module, a stacked spatial-temporal convolution module and an output module. The stacked spatial-temporal convolution module uses temporal convolution module to model temporal dependence, while utilizing graph convolution module over a dynamic graph to model dynamic spatial dependence. The dynamic graph is constructed based on attention mechanism and real-time input. Experiment results validate that the proposed method achieves better prediction performance than existing methods.