Abstract: The existing researches on the photovoltaic (PV) power output prediction have insufficient consideration for the influence of the complex spatio-temporal correlations and the black box nature of the deep learning hinders the interpretability of the prediction. To improve the accuracy of the PV power prediction at the multiple spatio-temporal scales and enhance the model's interpretability, a PV power prediction model that combines the spatio-temporal attention with the deep neural networks is proposed, as well as an analysis method for its interpretability. Firstly, a multi-dimensional attention mechanism is established, considering the temporal, spatial, and feature dimensions. This mechanism is combined with the deep neural networks and the quantile regression models to construct a PV interval prediction model. The model optimization is guided by the attention factors. Furthermore, a framework for explaining the prediction process and results of the deep learning models is proposed. The prediction mechanism is explained based on the structure of the model using the gradient method of neuron conductance. Additionally, the attention weights are used to identify the core spatio-temporal features that influence the power prediction. To validate the reliability of the interpretability results, the global marginal contributions of features are quantified using the Shapley additive principle. The prediction basis of the model is also combined with the sample-tracing models. Finally, the proposed model is validated using the data from the distributed PV power plants in a province in China. The results show that the proposed model has a higher prediction accuracy compared to the traditional prediction model.