Accurate photovoltaic (PV) forecasting is extremely important for power grid dispatching and power security. Generally, the power forecasting for a single PV station relies on its corresponding data, which neglects the valuable information of adjacent stations. Besides, the nonstationarity of PV power has not been considered yet. Against this background, this paper investigates a novel ultra-short-term PV forecasting approach called nsTransraph (nsTransformer graph modeling), which could utilize data from the adjacent stations to enhance the forecasting performance. Specifically, transfer entropy is used to select adjacent stations that provide beneficial information to the target station. Then, an efficient method called Transraph is proposed based on the Transformer and implicit graph modeling, to extract the spatio-temporal relationship of adjacent PV stations simultaneously. Moreover, a cross-attention aggregation mechanism is designed to aggregate the information adaptively within the graph structure. To deal with nonstationarity, a de-stationary module is embedded in the attention layer to focus on the target station and aggregate the data of adjacent PV stations simultaneously. Experimental results illustrate that the proposed nsTransraph method outperforms several state-of-the-art methods in the medium and long-term prediction on both practical datasets.
Wang et al. (Wed,) studied this question.