In order to address the challenges of spatiotemporal coupling and edge deployment difficulties in distributed photovoltaic power generation prediction, a multi granularity graph gated long short-term memory network coupled with an output knowledge distillation architecture is proposed. Firstly, utilizing dynamic collaboration mechanisms to unify multi-source heterogeneous observations and grid structure priors. Next, the three-branch graph encoder composed of the original nodes and the multi hop graph attention network extracts spatial context through the Softmax gating algorithm. Then, these spatial features are injected into the time step through conditional bias for joint multi time scale prediction. Finally, a lightweight student network was constructed using knowledge distillation for potential edge deployment. The results indicate that the model achieved an average absolute error of 25.31 MW and an average absolute percentage error of 1.48% within a prediction interval of 1 hour. During the merged time interval, the global average absolute percentage error decreased to 2.64%, which is 42.39% less than the traditional gated cyclic unit model. In addition, the lightweight student model successfully compressed the number of parameters from 319.7K to 44.9K, achieving a compression rate of 85.96%. Although knowledge distillation has brought about a small overall improvement in comprehensive error, its statistical significance is manifested as marginal benefits, and it exposes a certain trade-off of performance degradation in the 24-hour long-term prediction field. This framework achieves deep collaboration between data fusion, grid aware modeling, and lightweight compression, providing engineering reference for potential edge deployment in resource constrained scenarios.
