A graph attention network framework for generalized-horizon multi-plant solar power generation forecasting using heterogeneous data

Abstract

Accurate forecasting of solar power output from multiple photovoltaic plants simultaneously for different time horizons is crucial for their large-scale integration into the electric grid. Forecasting strategies for PV output power significantly vary depending on the forecasting horizons, quality, variety, resolution of data, and fields of application. While the researchers addressed many of these particular cases to achieve high forecasting accuracy, the literature lacks sufficient discussion on integrating strategies for various forecasting scenarios into a general framework. This article proposes such a framework facilitating PV power forecasting with variable time horizons and imparting data of various types and granularity by introducing a single adjustable module into the framework. Moreover, by proposing a geographic distance-based graph construction, ensuring minimal vertex connectivity and adjustable sparsity, the technique captures the spatiotemporal correlation among the PV plant through a graph attention network for accurate forecasting. The proposed technique is highly scalable archiving excellent forecasting performance (an average absolute error of $<5\text{\%}$ of the plant capacity for 5 min to 3-day forecasting horizon) up to thousands of PVs. The results indicate that the proposed method outperforms state-of-the-art approaches, such as Long Short-Term Memory networks, in terms of both accuracy and scalability. Additionally, this paper analyzes the suitability of features for forecasting in different scenarios and performance sensitivity to various model parameters.