HiGN-ARec: A Hierarchical Graph Network with Adaptive Reconciliation for PV Power Forecasting in Spatial Hierarchy

Prediction of photovoltaic (PV) power is extremely important in power grid which is organized in a hierarchical framework. In this paper, we propose a novel end-to-end deep network, named hierarchical graph network with adaptive reconciliation (HiGN-ARec), to predict the PV power for all levels. Specifically, HiGN-ARec consists of two parts: the base forecasts and the reconciliation. For the base forecasts, we adopt the state-of-the-art spatio-temporal (ST) modules to exploit the intralevel information and propose interaction modules to capture the valuable inter-level information. The cooperation of both modules in the base forecasts contributes to the significant enhancement of the accuracy of base forecasts. For the reconciliation, we propose a reconciliation matrix P which is used to adjust the base forecasts and aggregation matrix S to impose the consistency constraint on the network. The reconciliation matrix P is learnable and determined with other parameters in the network via training. It incorporates knowledge about hierarchy structure and learned parameters that adjust base forecasts effectively. We use synthetic solar PV power plant data from the National Renewable Energy Laboratory (NREL) to verify the effectiveness of the proposed method. Experimental results demonstrate the superiority of our proposed approach over the competing methods.