A Distributed PV Ultra-short-term Power Deterministic and Probabilistic Forecasting Based on Dynamic Graph Network with Shape-amplitude Criteria

Accurate forecasting of distributed photovoltaic power plays a crucial role in ensuring the safety and stability of an active distribution network. However, most existing research on distributed photovoltaic power forecasting exhibits certain limitations, including: 1) insufficient consideration of the dynamic correlations among power sites; and 2) absence of a training loss function capable of simultaneously aligning the amplitude and shape of forecasting values with the true values. Therefore, a dynamic graph network with a shape-amplitude loss function based distributed photovoltaic ultra-short-term power forecasting method is introduced. Firstly, a data-driven method is used to mine the dynamic correlation and the dynamic graph data can be generated to ensure the effective characterization of the correlation among distributed photovoltaics. Secondly, the dynamic graph network is constructed as the power forecasting model to realize the effective utilization of spatial-temporal correlation features. Then, the shape-amplitude loss function which combines the Dynamic Time Warping and Mean Square Error is used as the criterion of model training to ensure the consistency of the forecasting value and the real in situ shape and amplitude. Meanwhile, the dynamic graph network is combined with quantile regression and the quantile loss function is improved inspired from the idea of shape-amplitude. The forecasting performance of the introduced approach is demonstrated via a distributed photovoltaic power dataset in China.