Distributed Photovoltaic Power Prediction Based on GAT-TST

Distributed photovoltaic (PV) power stations are characterized by their dispersed layout, small equipment investment, and relatively poor quality of historical data, along with a general lack of meteorological data collection devices. This makes it challenging to improve the accuracy of power forecasting. To address this, this paper proposes a method for improving prediction accuracy by leveraging the spatiotemporal correlation of distributed PV stations. First, each distributed PV station is defined as a graph node, and the Graph Attention Network (GAT) is used to extract spatial features from the historical power data between stations. Then, taking advantage of the Time Series Transformer(TST) model's ability to efficiently capture temporal features, the dependencies between different time points in the historical PV data are captured, and the final prediction result is output through a fully connected layer. Finally, based on real PV station data, the model is validated for ultra-short-term power forecasting using only historical power data, even in the absence of meteorological data collection devices.